Decoding Cancer Dynamics: A Comprehensive Guide to RNA Velocity in Single-Cell Analysis

Abstract

This article provides researchers, scientists, and drug development professionals with a comprehensive overview of RNA velocity and its transformative application in single-cell cancer research. We explore the foundational principles of RNA velocity, which leverages spliced and unspliced mRNA to infer transcriptional dynamics and predict future cell states from static snapshots. The review systematically compares cutting-edge computational tools—including scVelo, Dynamo, TSvelo, spVelo, and VeloVGI—highlighting their unique strengths in modeling complex tumor microenvironments, multi-lineage differentiation, and therapy response. We address critical challenges such as batch effects, data sparsity, and model selection while offering practical troubleshooting guidance. Through validation frameworks and biological applications in identifying cells of origin, tracking tumor evolution, and characterizing therapeutic resistance, this guide establishes RNA velocity as an indispensable methodology for uncovering cancer mechanisms and informing novel therapeutic strategies.

The RNA Velocity Paradigm: From Splicing Kinetics to Cancer Fate Prediction

Single-cell RNA sequencing (scRNA-seq) has revolutionized biology by enabling high-throughput quantification of gene expression at the individual cell level [1]. However, a significant limitation persists: standard scRNA-seq provides only static cellular snapshots, obscuring dynamic temporal processes like differentiation, reprogramming, and disease progression [1]. RNA velocity, introduced in 2018, offers a groundbreaking solution to this problem by leveraging the inherent kinetics of RNA transcription [1]. The method exploits the relationship between unspliced pre-mRNA (nascent) and spliced mRNA (mature) to infer instantaneous gene expression change rates, effectively predicting future transcriptional states over hour-long timescales [1] [2]. This approach has become indispensable for navigating the complex, dynamic cellular world, transforming our understanding of temporal biological processes from static observations to predictive, dynamic insights that illuminate cellular fate decisions and disease mechanisms, particularly in cancer research [1].

Core Concepts and Theoretical Foundations

The Biochemical Basis of RNA Velocity

The fundamental premise of RNA velocity lies in the transcriptional dynamics of mRNA synthesis. For each gene, the process involves transcription (producing unspliced pre-mRNA), splicing (converting unspliced to spliced mRNA), and degradation of spliced mRNA [2]. The key insight is that unspliced mRNA serves as a leading indicator of spliced mRNA abundance, providing a window into the cell's future transcriptional state [2]. The rate of change of spliced mRNA ((ds/dt)) is defined as RNA velocity, which represents the direction and speed of movement for individual cells in gene expression space [3] [4].

Mathematical Models of Transcriptional Dynamics

The original RNA velocity concept relies on a steady-state model based on ordinary differential equations (ODE) [4]. This model assumes a constant transcriptional state and infers velocity from deviations from the steady-state ratio of unspliced to spliced mRNAs [3]. Second-generation methods like scVelo introduced a dynamical model that recovers the full transcriptional dynamics, using an expectation-maximization (EM) algorithm to iteratively update ODE rate parameters and cell-specific latent time [4]. More recent approaches like TIVelo bypass explicit ODE assumptions by determining velocity direction at the cluster level based on trajectory inference, better capturing complex transcriptional patterns [5].

Table 1: Key RNA Velocity Estimation Methods and Their Characteristics

Method	Underlying Model	Key Features	Applications
Velocyto [1]	Steady-state ODE	Robust regression for degradation rates	Basic velocity estimation
scVelo [3]	Dynamical ODE	EM algorithm for parameters and latent time	Complex systems, multiple trajectories
TIVelo [5]	Trajectory inference	Cluster-level direction inference	Systems violating ODE assumptions
VeloVGI [6]	Variational graph autoencoder	Batch effect correction via graph networks	Multi-batch, multi-condition datasets
TopicVelo [7]	Stochastic model with topic modeling	Disentangles multiple concurrent processes	Complex systems with branching points

Experimental Protocol: scVelo RNA Velocity Analysis

Data Preprocessing Requirements

The computational protocol begins with loading single-cell data containing both spliced and unspliced counts. The AnnData object format serves as the standard container, storing the data matrix (adata.X), observations (adata.obs), variables (adata.var), unstructured annotations (adata.uns), and layers for spliced and unspliced counts [3].

Essential preprocessing steps include:

• Gene selection by detection (minimum shared counts of 20) and high variability (dispersion)
• Normalization of every cell by its total size
• Logarithmization of the data matrix (X)
• Computation of moments among nearest neighbors in PCA space (30 PCs, 30 neighbors)

These steps are implemented in scVelo as follows [3]:

Velocity Estimation and Visualization

After preprocessing, RNA velocity is estimated using transcriptional dynamics of splicing kinetics. The standard approach uses stochastic mode (default), though deterministic (mode='deterministic') and dynamical (mode='dynamical') modes are available [3].

Key steps in velocity estimation:

• Velocity computation via scv.tl.velocity(adata)
• Velocity graph calculation using scv.tl.velocity_graph(adata)
• Projection onto low-dimensional embeddings (UMAP, t-SNE)
• Visualization as streamlines, gridlines, or single-cell vectors

The transition probabilities between cells are computed using cosine correlation between potential cell-to-cell transitions and the velocity vector, stored in a velocity graph matrix of dimension (n{obs} \times n{obs}) [3].

Interpretation and Validation

Critical interpretation of velocity results requires examining individual gene dynamics through phase portraits, which plot spliced against unspliced counts for each gene [3]. The black line in phase portraits represents the estimated 'steady-state' ratio, with RNA velocity determined as the residual from this line [3]. Positive velocity indicates gene up-regulation (higher unspliced mRNA than expected), while negative velocity indicates down-regulation [3].

Diagram 1: RNA velocity analysis workflow.

The Scientist's Toolkit: Essential Research Reagents and Computational Tools

Table 2: Essential Research Reagent Solutions for RNA Velocity Analysis

Tool/Reagent	Function	Application Context
scVelo [3]	Python-based velocity estimation	Dynamical modeling of transcriptional dynamics
Velocyto [1]	Initial RNA velocity implementation	Steady-state model applications
CellRank [1]	Fate probability estimation	Identifying initial, intermediate, terminal states
TIVelo [5]	Cluster-level trajectory inference	Systems with complex transcriptional patterns
VeloVGI [6]	Batch effect correction	Multi-batch, multi-condition datasets
TopicVelo [7]	Process-disentanglement via topic modeling	Complex systems with concurrent processes
Pancreas Dataset [3]	Endocrine development benchmark	Protocol validation and method testing

Advanced Methodologies and Integration with Multimodal Data

Addressing Current Limitations

Despite its promise, RNA velocity has important limitations. A significant challenge is its reliance on smoothing via the k-nearest-neighbors (k-NN) graph, which can result in considerable estimation errors when the graph fails to accurately represent the true data structure [4]. RNA velocity performs poorly at estimating speed except in very low noise settings, and users are advised against over-interpreting expression dynamics, particularly in terms of speed [4]. A novel quality measure has been introduced to identify when RNA velocity should not be used [4].

Integration with Multimodal Data

Recent advances integrate RNA velocity with other data modalities. MultiVelo combines chromatin accessibility data, protaccel incorporates protein abundances, Dynamo uses new/total labeled RNA-seq, PhyloVelo leverages phylogenetic trees, and TFvelo incorporates transcription factor information [5]. These integrations help address fundamental limitations and provide more comprehensive views of cellular dynamics.

Diagram 2: Evolution beyond basic ODE models.

Applications in Cancer Dynamics Research

In cancer research, RNA velocity provides unique insights into tumor evolution, drug resistance development, and metastatic processes. The method reveals novel disease mechanisms by analyzing immune cell differentiation and state transitions in complex tumor microenvironments [1]. For cancer systems, methods like TopicVelo are particularly valuable as they can disentangle multiple concurrent processes such as proliferation, stress response, and differentiation, which often occur simultaneously in tumor cells [7]. The ability to predict cellular fate decisions without prior knowledge makes RNA velocity particularly powerful for studying rare cell populations and transition states that drive cancer progression and therapeutic resistance.

RNA velocity has evolved significantly from its initial implementation, with current methods moving beyond simple ODE assumptions to incorporate cluster-level inference, deep learning, and multimodal data integration. While limitations remain, particularly regarding speed estimation and sensitivity to preprocessing choices, the method continues to provide unprecedented insights into cellular dynamics. For cancer researchers, the growing toolkit of velocity methods offers powerful approaches to unravel tumor heterogeneity, plasticity, and progression mechanisms. Future directions will likely focus on improving model robustness, integrating additional biological layers, and developing more rigorous validation frameworks to establish RNA velocity as a quantitative rather than qualitative tool in single-cell cancer dynamics research.

RNA velocity analysis has emerged as a transformative computational method for predicting cellular dynamics from single-cell RNA sequencing (scRNA-seq) data. By leveraging the intrinsic kinetics of RNA splicing, this approach allows researchers to infer the direction and speed of cellular state transitions, making it particularly valuable for studying cancer progression, tumor heterogeneity, and treatment response [1] [8]. The core principle rests on distinguishing between unspliced (nascent, pre-mRNA) and spliced (mature, mRNA) transcripts within individual cells, then using their ratio to predict future transcriptional states [8] [9]. In cancer research, this provides a powerful "window" into dynamic processes such as drug resistance emergence, metastatic evolution, and stem cell lineage commitment, moving beyond static snapshots to model the temporal dynamics that define tumor behavior [10].

Theoretical Foundations of Splicing Kinetics

The Splicing Process and Kinetic Modeling

The journey from gene to functional protein begins with transcription, where RNA polymerase II produces pre-messenger RNA (pre-mRNA) containing both exonic and intronic regions. This nascent RNA is classified as unspliced (u). Through the complex process of splicing, performed by the spliceosome, introns are removed and exons are joined together to form spliced (s) mature mRNA [10] [11]. This mature mRNA is then exported to the cytoplasm for translation.

The kinetic relationship between these two molecular species is typically modeled using a two-step process described by ordinary differential equations (ODEs):

Where α(t) represents the transcription rate, β denotes the splicing rate constant, and γ is the degradation rate constant for the mature mRNA [12] [8]. The key observable quantity, RNA velocity, is defined as the time derivative of spliced mRNA abundance (ds/dt). A positive velocity indicates future upregulation of the gene, while a negative velocity predicts downregulation [8] [5].

Computational Paradigms in RNA Velocity

As the field has evolved, three distinct computational paradigms have emerged for inferring transcriptional kinetics from unspliced and spliced mRNA data:

Table 1: Computational Paradigms in RNA Velocity Analysis

Category	Underlying Principle	Representative Methods	Strengths	Limitations
Steady-State Methods	Assumes constant splicing rates and transcriptional equilibrium; uses least-squares regression on steady-state subpopulations	Velocyto, scVelo (stochastic model)	Simple, fast, and interpretable; effective for clear differentiation processes	Assumptions often violated in heterogeneous populations; inaccurate for complex kinetics [8]
Trajectory Methods	Estimates kinetic parameters to construct phase portrait trajectories aligning cells with corresponding cell times	scVelo (dynamical model), UniTVelo, DeepCycle	Captures more complex dynamics; assigns latent cell time	May struggle with highly discontinuous processes [8] [9]
State Extrapolation Methods	Leverages expected future cell states to guide estimation of cell-level RNA velocity vectors	VeloVAE, Pyro-Velocity, LatentVelo	Flexible modeling of complex patterns; incorporates uncertainty	Higher computational demand; more complex interpretation [8]

Experimental and Computational Workflows

Wet-Lab Protocol: From Cells to Sequencing Data

Sample Preparation and RNA Isolation

• Starting Material: Begin with high-quality single-cell suspensions from cancer tissue or cell lines. Preserve cells immediately in RNA stabilization reagent to prevent degradation.
• RNA Extraction: Use PicoPure RNA isolation kit or equivalent. Work with RNase-free reagents and surfaces to prevent RNA degradation.
• Quality Control: Assess RNA purity using NanoDrop (aim for 260/280 ratio ~2.0). Evaluate RNA integrity using Agilent TapeStation (RIN >7.0 required). Minimum 500 ng total RNA needed for standard protocols [13] [14].

Library Preparation and Sequencing

• Poly-A Selection: Isolate mRNA using NEBNext Poly(A) mRNA magnetic isolation kits or equivalent.
• cDNA Synthesis and Library Prep: Use NEBNext Ultra DNA Library Prep Kit for Illumina. For low-input samples (≤500 pg RNA), employ SMART-Seq v4 Ultra Low Input RNA kit.
• rRNA Depletion: Remove ribosomal RNA using QIAseq FastSelect (~14 minutes, >95% efficiency).
• Sequencing: Run on Illumina NextSeq 500 or similar platform with 75-cycle single-end high-output sequencing kit. Target ≥20,000 reads per cell for velocity analysis [13] [14] [9].

Computational Protocol: From Raw Data to Velocity Estimates

The computational workflow for RNA velocity estimation follows a structured pipeline that transforms raw sequencing data into interpretable velocity vectors:

Figure 1: Computational Workflow for RNA Velocity Analysis

Data Preprocessing Steps

• Quality Control: Process raw FASTQ files with FastQC to assess Phred quality scores (>Q30), adapter contamination, and GC content.
• Read Trimming: Use Trimmomatic or cutadapt with quality threshold of 10 to remove low-quality bases and adapters.
• Quantification: Employ alignment-free tools like Kallisto or Salmon for rapid quantification of unspliced and spliced counts, or alignment-based tools like STAR or HISAT2 for spliced alignment [13] [8].
• Filtering: Apply filterByExpr from edgeR to retain genes with sufficient counts (≥10 counts in enough samples) [13].

Velocity Estimation and Visualization

• Gene Selection: Identify genes with sufficient expression and clear kinetic patterns for velocity analysis.
• Model Fitting: Apply appropriate velocity method (see Table 1) to estimate kinetic parameters (transcription, splicing, degradation rates).
• Projection: Embed high-dimensional velocity vectors into 2D space using UMAP, t-SNE, or PCA.
• Visualization: Generate velocity streamplots using scVelo or similar packages to visualize predicted cell state transitions [8].

Advanced Methodologies in RNA Velocity

Integrative Methods for Complex Biology

Recent advancements have extended RNA velocity beyond basic splicing kinetics to incorporate additional biological layers particularly relevant to cancer research:

TSvelo models the cascade of gene regulation, transcription, and splicing using neural Ordinary Differential Equations (ODEs). It incorporates transcriptional regulation by modeling transcription factor-target relationships using databases like ChEA and ENCODE, providing more accurate dynamics for complex processes like cancer lineage specification [12].

Multi-omic Integration approaches include:

• MultiVelo: Integrates scATAC-seq data with splicing information to connect chromatin dynamics with transcriptional outcomes.
• protaccel: Incorporates protein abundance data to bridge the gap between mRNA and protein dynamics.
• TFvelo: Explicitly models transcription factor dynamics to uncover regulatory drivers in cancer progression [12] [8].

Deep-Learning Approaches such as DeepCycle use autoencoder neural networks to fit circular patterns in unspliced-spliced space, particularly effective for cycling processes like cell division in cancer cells [9].

The Scientist's Toolkit: Essential Research Reagents and Computational Tools

Table 2: Essential Resources for RNA Velocity Experiments

Category	Item	Specific Examples	Function/Purpose
Wet-Lab Reagents	RNA Stabilization Reagents	Liquid nitrogen, dry-ice ethanol baths, RNAlater	Preserve RNA integrity immediately post-collection
	RNA Isolation Kits	PicoPure RNA Isolation Kit, QIAseq UPXome RNA Library Kit	Extract high-quality RNA from limited samples
	Library Prep Kits	NEBNext Ultra DNA Library Prep Kit, SMART-Seq v4 Ultra Low Input RNA Kit	Prepare sequencing libraries from extracted RNA
	rRNA Depletion Kits	QIAseq FastSelect	Remove abundant ribosomal RNA to improve detection of mRNA
Computational Tools	Quantification Tools	Kallisto, Salmon, STAR, HISAT2	Quantify unspliced and spliced mRNA abundances
	Velocity Methods	Velocyto, scVelo, TSvelo, DeepCycle	Estimate RNA velocity from count matrices
	Visualization Packages	scVelo, Scanpy	Project and visualize velocity vectors
Reference Databases	TF-Target Databases	ChEA, ENCODE	Curated transcription factor-target relationships for regulatory models [12]

Cancer-Specific Applications and Considerations

Splicing Dysregulation in Cancer Biology

Cancer cells frequently exhibit widespread splicing alterations that can be exploited through RNA velocity analysis. Key mechanisms include:

• Recurrent Mutations: Driver mutations in splicing factors like SF3B1, SRSF2, U2AF1, and ZRSR2 occur in various hematologic and solid malignancies, creating distinct splicing landscapes that RNA velocity can track [10].
• Splicing Factor Dysregulation: Even without mutations, cancer cells show altered expression of splicing factors that rewire splicing patterns to support tumorigenesis, including effects on apoptosis, DNA repair, and metabolism pathways [10].
• Noncoding RNA Interactions: Cancer-associated noncoding RNAs like snaR-A interact with core splicing machinery (e.g., U2 snRNP protein SF3B2) and disrupt mRNA processing, creating molecular fingerprints detectable through velocity analysis [11].

Protocol for Investigating Cancer Cell Transitions

Objective: Identify transitional states and directionality in tumor cell populations using RNA velocity.

Step-by-Step Procedure:

• Sample Collection: Process tumor and matched normal tissue simultaneously to minimize batch effects. Include technical replicates.
• Cell Sorting: Use fluorescence-activated cell sorting (FACS) with appropriate markers to enrich for viable single cells. Sort directly into RNA stabilization buffer.
• Library Preparation: Employ low-input protocols (e.g., SMARTer Stranded Total RNA-Seq Kit v2 - Pico Input) to handle limited clinical samples.
• Sequencing: Aim for higher sequencing depth (>50,000 reads/cell) to adequately capture both unspliced and spliced counts for velocity estimation.
• Velocity Analysis:
  • Apply TSvelo for complex cancer lineages to model regulatory cascades
  • Use DeepCycle for cycling tumor subpopulations
  • Implement TIVelo when ODE assumptions may be violated in heterogeneous tumors
• Validation: Integrate with spatial transcriptomics where possible to validate predicted transitions against tumor morphology [15].

Troubleshooting Tips:

• If velocity vectors appear random or contradictory, check RNA quality (RIN >7) and increase sequencing depth.
• For inconsistent phase portraits, apply more flexible models like TSvelo or neural ODE approaches that don't assume constant kinetic rates.
• When working with sparse clinical samples, employ imputation methods designed for velocity analysis [12] [9].

RNA velocity analysis based on splicing kinetics provides a powerful framework for modeling cancer dynamics from static single-cell RNA-seq data. The continuous refinement of computational methods—from simple steady-state models to sophisticated integrative approaches—has significantly enhanced our ability to predict tumor progression, therapeutic resistance, and metastatic pathways. As the field advances, key challenges remain in improving model accuracy for highly heterogeneous cancer samples, integrating multi-omic data layers, and validating predicted dynamics through perturbation experiments and spatial mapping. For cancer researchers and drug development professionals, these methods offer increasingly refined tools to identify critical transitional states in tumor evolution, potentially revealing novel therapeutic targets for interrupting progressive cancer pathways.

Ordinary Differential Equations (ODEs) serve as the cornerstone for modeling the dynamic processes of gene expression at single-cell resolution. In single-cell cancer dynamics research, ODE-based models power RNA velocity analysis, a transformative methodology that predicts cellular trajectories from snapshot transcriptional data. These models mathematically represent the unobserved temporal dynamics of transcription, splicing, and degradation, enabling researchers to forecast cell states and fate decisions critical to understanding tumor progression, heterogeneity, and drug response mechanisms. This foundation provides the mechanistic framework needed to move beyond static observations toward predictive models of cancer biology.

Core Mathematical Principles

The application of ODEs in transcriptional modeling is built upon a framework that describes the kinetics of RNA metabolism. These equations formalize the cascade of molecular events from gene activation to mature mRNA degradation.

The Fundamental Dynamical System

The foundational model for RNA velocity describes the time evolution of unspliced (pre-mRNA) and spliced (mature mRNA) transcript abundances for each gene [16]. The system is defined by a coupled pair of ordinary differential equations:

$$ \begin{aligned} \frac{dug}{dt} &= \alphag(t) - \betag ug \ \frac{dsg}{dt} &= \betag ug - \gammag s_g \end{aligned} $$

where:

• $u_g(t)$ represents the unspliced mRNA concentration for gene $g$ at time $t$
• $s_g(t)$ represents the spliced mRNA concentration for gene $g$ at time $t$
• $\alpha_g(t)$ denotes the transcription rate function
• $\beta_g$ is the splicing rate constant
• $\gamma_g$ is the degradation rate constant for spliced mRNA

This formulation captures the essential biological processes where unspliced transcripts are produced at rate $\alphag(t)$, converted to spliced transcripts at rate $\betag$, and degraded at rate $\gamma_g$.

Modeling Transcriptional Regulation

The transcription rate $\alphag(t)$ can be modeled with varying complexity depending on the biological context. In basic models, it is often treated as a constant or a switching function between active and inactive states. More sophisticated approaches model $\alphag(t)$ as a function of transcription factor (TF) activities that regulate gene $g$ [12]:

$$ \alphag(t) = f\left(\sum{TF \in TFs(g)} w{TF,g} \cdot x{TF}(t)\right) $$

where $TFs(g)$ represents the set of transcription factors regulating gene $g$, $w{TF,g}$ are regulatory weights, and $x{TF}(t)$ are the TF expression levels. This formulation enables the integration of regulatory network information into the dynamical model.

Current ODE-Based Methodologies in RNA Velocity

Computational methods implementing ODE-based RNA velocity have evolved from simple steady-state assumptions to complex generative models. The table below summarizes key methodologies and their mathematical foundations:

Table 1: ODE-Based Methodologies for RNA Velocity Analysis

Method	Mathematical Approach	Key Parameters	Cancer Research Applications
Steady-State Model [16]	Constant transcription rate assumption: $\alphag(t) = \alphag$	$\betag$, $\gammag$ estimated from extreme quantiles	Limited to systems at transcriptional equilibrium
EM Model (scVelo) [16]	Expectation-Maximization for parameter inference; gene-specific latent time	$\alphag$, $\betag$, $\gammag$, $tc$	Identifying differentiation trajectories in cancer stem cells
veloVI [17]	Deep generative modeling; Bayesian inference with variational autoencoder	Posterior distributions over all parameters	Quantifying uncertainty in trajectory inference
TSvelo [12]	Unified ODE incorporating regulation, transcription, and splicing: $ \alphag(t) = \sum{TF} w{TF,g} \cdot x{TF}(t) $	$w{TF,g}$, $\betag$, $\gamma_g$, unified latent time	Multi-lineage tumor progression analysis
Cell-MNN [18]	Locally linear ODE in latent space: $\dot{z} = A(z,t)z$	Matrix $A(z,t)$ defining local dynamics	Scalable analysis of large cancer datasets
SCODE [19]	Linear ODE framework: $dx = Axdt$	Regulatory network matrix $A$	Efficient gene regulatory network inference

Methodological Advancements

Recent methodological advances have addressed key limitations in earlier ODE-based approaches. veloVI introduces a deep generative modeling framework that provides uncertainty quantification through posterior distributions over velocities, enabling researchers to assess confidence in predicted trajectories [17]. TSvelo integrates transcriptional regulation with splicing kinetics through a comprehensive ODE framework that simultaneously models all selected genes, allowing for the inference of a unified latent time across the transcriptome [12]. Cell-MNN employs a locally linearized ODE in a latent space to efficiently capture complex dynamics while maintaining interpretability through explicit gene interaction terms [18].

Experimental Protocols

Standard RNA Velocity Workflow

The following protocol outlines a standard workflow for RNA velocity analysis using ODE-based methods, with specific notes for cancer applications:

Table 2: Key Research Reagents and Computational Tools

Category	Specific Tools/Reagents	Function in Analysis
Data Generation	10x Genomics Single-Cell RNA-seq	Generation of spliced/unspliced count matrices
Preprocessing	Scanpy, Scanny	Quality control, normalization, and filtering
Velocity Estimation	scVelo, Velocyto, veloVI	ODE parameter inference and velocity calculation
Visualization	matplotlib, scVelo plotting	Stream plots and embedding visualization
Validation	FUCCI cell cycle indicators, metabolic labeling	Orthogonal validation of directionality

Step 1: Data Acquisition and Preprocessing

• Generate single-cell RNA sequencing data using protocols that preserve strand information (e.g., 10x Genomics)
• Quantify spliced and unspliced counts using tools like Velocyto or kallisto-bustools
• Filter genes based on minimum expression thresholds (typically ≥20 counts for both spliced and unspliced) [16]
• Normalize cell sizes and apply log1p transformation to reduce outlier effects
• Select highly variable genes (typically 2,000-3,000 genes) for downstream analysis

Step 2: Data Smoothing and Moment Calculation

• Perform principal component analysis (PCA) on the spliced count matrix
• Construct a k-nearest neighbor (k-NN) graph based on PCA coordinates
• Calculate first-order moments (neighborhood averages) for both spliced and unspliced counts using the k-NN graph: $Ms(c) = \frac{1}{|N(c)|} \sum{c' \in N(c)} s(c')$ where $N(c)$ is the neighborhood of cell $c$ [16]
• This smoothing step is critical for reducing technical noise and obtaining accurate velocity estimates

Step 3: ODE Parameter Estimation The specific implementation varies by method:

For EM Model (scVelo):

• Initialize parameters and latent variables
• E-step: Infer latent times $tc$ and transcriptional states $zc$ for each cell
• M-step: Estimate kinetic parameters $\alphag$, $\betag$, $\gamma_g$ that maximize likelihood
• Iterate until convergence of the likelihood function [16]

For veloVI:

• Train variational autoencoder to infer posterior distributions over parameters
• Encoder networks process unspliced and spliced abundances to output posterior parameters
• Likelihood of observed data computed as function of latent time and kinetic parameters
• Optimize using gradient-based methods [17]

Step 4: Velocity Calculation and Visualization

• Compute high-dimensional velocity vectors: $vg = \betag ug - \gammag s_g$ for each gene
• Project velocities to low-dimensional embeddings (e.g., UMAP, t-SNE) using transition probability method
• Visualize as streamlines or vector fields to interpret cellular dynamics
• Construct velocity graph to model transitions between cell states

Specialized Protocol: Multi-Lineage Analysis with TSvelo

For complex cancer datasets with multiple lineages, TSvelo provides a specialized protocol:

Step 1: Gene Selection and Regulatory Network Integration

• Select velocity genes based on expression and variability
• Integrate TF-target information from databases (ChEA, ENCODE)
• Formulate the comprehensive ODE model:

$\frac{dug}{dt} = \alphag(t) - \betag ug$, $\frac{dsg}{dt} = \betag ug - \gammag sg$, with $\alphag(t) = \sum{TF \in TFs(g)} w{TF,g} \cdot x_{TF}(t)$ [12]

Step 2: Unified Latent Time Inference

• Initialize pseudotime using diffusion-based methods
• Employ Expectation-Maximization to iteratively optimize:
  • ODE parameters ($w{TF,g}$, $\betag$, $\gamma_g$) in M-step
  • Unified latent time across all cells in E-step
• Use Neural ODE for numerical solution when analytical solutions are intractable

Step 3: Model Validation

• Calculate velocity consistency: Agreement between velocities of neighboring cells
• Assess in-cluster coherence: Directional consistency within annotated cell types
• Evaluate cross-boundary correctness: Agreement with known differentiation pathways

Signaling Pathways and Workflows

The following diagrams illustrate key signaling pathways and computational workflows in ODE-based transcriptional modeling.

Transcriptional Kinetics Signaling Pathway

Diagram 1: Transcriptional kinetics pathway. This diagram illustrates the core signaling pathway of transcriptional kinetics, showing the conversion from DNA to unspliced pre-mRNA, splicing to mature mRNA, and eventual degradation. Transcription factors (TFs) regulate the transcription rate, forming the basis for ODE modeling of RNA velocity.

RNA Velocity Computational Workflow

Diagram 2: RNA velocity computational workflow. This workflow diagram outlines the key steps in RNA velocity analysis, from raw scRNA-seq data preprocessing to final visualization, highlighting the central role of ODE parameter estimation with various methodological approaches.

Applications in Cancer Dynamics Research

ODE-based RNA velocity models have enabled significant advances in understanding cancer dynamics:

Tumor Heterogeneity and Evolution

RNA velocity analysis reveals lineage relationships and temporal ordering within tumors, mapping progression from cancer stem cells to differentiated states. In single-cell studies of leukemia, ODE models have reconstructed differentiation blocks and identified regulatory programs that maintain stem-like populations [1]. The parameter estimates from these models (α, β, γ) provide quantitative insights into transcriptional dysregulation across subpopulations.

Drug Response Mechanisms

By applying RNA velocity to time-course scRNA-seq data from treated tumor cells, researchers can track early transcriptional shifts that predict eventual drug response or resistance. The dynamical information captures transitional states that are missed in static analyses, potentially revealing novel therapeutic targets to prevent resistance emergence.

Metastatic Progression

ODE models help decipher the regulatory programs driving epithelial-mesenchymal transition (EMT) and metastatic seeding. In pancreatic cancer studies, velocity analysis has revealed bidirectional plasticity in EMT programs and identified key transcription factors regulating these transitions [12].

Limitations and Future Directions

Despite their transformative potential, ODE-based transcriptional models face several challenges:

Technical Limitations

Current RNA velocity methods depend critically on the accuracy of k-nearest neighbor graphs for data smoothing, and errors in graph construction propagate to velocity estimates [4]. The assumption of constant kinetic rates may be violated in complex biological systems, particularly during rapid state transitions in cancer. Additionally, velocity estimates for speed (magnitude) are less reliable than direction, except in very low-noise settings.

Validation Challenges

Direct experimental validation of RNA velocity estimates remains difficult, with studies showing poor correspondence between splicing-based velocities and those derived from metabolic labeling for some genes [4]. Developing robust validation frameworks is essential for advancing these methods in cancer research contexts.

Emerging Solutions

Newer approaches address these limitations through:

• Uncertainty quantification (veloVI) providing confidence estimates for predictions [17]
• Time-dependent transcription rates modeling complex regulatory dynamics [12]
• Multi-omic integration combining splicing data with epigenetic information
• Foundation models (GET) that learn generalizable transcriptional principles across cell types [20]

Table 3: Comparison of Key ODE Model Parameters Across Methods

Parameter	Steady-State	EM Model	veloVI	TSvelo	Biological Interpretation
Transcription Rate (α)	Constant	Constant or switching	Time-dependent	TF-regulated	Gene activation strength
Splicing Rate (β)	Global constant	Gene-specific	Gene-specific	Gene-specific	pre-mRNA processing efficiency
Degradation Rate (γ)	Gene-specific	Gene-specific	Gene-specific	Gene-specific	mRNA stability and turnover
Latent Time	Not inferred	Gene-specific	Cell-specific	Unified global	Cellular progression along trajectory
Uncertainty Estimation	None	Implicit	Explicit posterior	Limited	Confidence in predictions

ODE-based transcriptional modeling represents a powerful framework for unraveling dynamic cancer processes from single-cell data. The mathematical foundation provided by coupled differential equations enables researchers to move beyond static snapshots to predictive models of tumor evolution, treatment response, and metastatic progression. As methods continue to advance through deeper integration of regulatory networks, improved uncertainty quantification, and multi-omic data integration, these approaches will increasingly enable truly predictive cancer biology, with potential applications in personalized treatment forecasting and therapeutic target discovery. The ongoing development of more sophisticated ODE frameworks promises to further enhance our ability to model and ultimately control oncogenic processes at single-cell resolution.

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity in cancer, yet it provides only static snapshots of transcriptional states. RNA velocity, by modeling the temporal dynamics of gene expression from spliced and unspliced mRNA ratios, overcomes this limitation by predicting future cellular states and uncovering the directionality of cellular transitions. This application note details how RNA velocity is transforming cancer biology by enabling the tracking of cell plasticity, tumor origin, and evolutionary dynamics. We provide structured protocols for implementing velocity analyses in cancer research, supported by quantitative data comparisons and visual workflows designed for researchers and drug development professionals.

Cancer is not a static condition but a dynamic system characterized by continuous evolution and cellular plasticity. Traditional scRNA-seq analyses identify distinct cell populations within tumors but cannot capture the ongoing transitions that underlie critical processes such as therapeutic resistance, metastatic progression, and cellular reprogramming [21]. RNA velocity analysis addresses this gap by inferring the instantaneous rate of change of gene expression, effectively predicting cellular futures on a timescale of hours from standard single-cell snapshots [22].

The core premise of RNA velocity lies in distinguishing between nascent (unspliced) and mature (spliced) messenger RNA transcripts. The relative abundance of these RNA species reveals the current transcriptional trajectory of each gene within a cell. A positive RNA velocity indicates active gene induction, while a negative velocity signifies repression [8]. When aggregated across the transcriptome, these vectors form a high-dimensional velocity field that predicts the future state of individual cells, illuminating developmental trajectories and state transitions directly from static samples [1] [22].

Key Applications in Cancer Biology

Mapping Cellular Plasticity and Fate Decisions

Cell plasticity—the ability of cells to alter their phenotypes without genetic change—is a fundamental driver of tumor adaptability, therapy resistance, and metastatic potential [21]. RNA velocity directly probes this plasticity by revealing transient cellular states and directional fate decisions.

• Identifying Metastatic and Drug-Tolerant Trajectories: RNA velocity can reconstruct the trajectories that epithelial cells follow as they undergo epithelial-to-mesenchymal transition (EMT), a key plasticity program in metastasis. Similarly, it can identify the early transcriptional shifts that precede the emergence of a drug-tolerant persister state, offering a window into adaptive resistance mechanisms before they become fixed. In studying Alzheimer's disease, which involves diverse cellular perturbations, researchers found that genes with differential RNA velocity were qualitatively distinct from those with differential expression alone. These dynamically altered genes were specifically associated with synaptic organization and cell development processes [23]. This paradigm can be directly applied to cancer to distinguish the active drivers of plasticity from passively altered genes.

• Uncovering Branching Lineage Trees: During development and in cancer stem cell hierarchies, cells face binary fate decisions. RNA velocity has proven powerful in resolving these branching points. In the developing mouse hippocampus, for instance, RNA velocity revealed a complex manifold with multiple branches, accurately showing directional flow towards distinct neuronal and glial fates [22]. In cancer, this capability can delineate the branching choices between self-renewal and differentiation within a tumor, pinpointing the transcriptional regulators that govern cell fate.

Tracing Cellular Origins and Evolutionary Histories

Understanding a tumor's cellular origin and evolutionary history is crucial for deciphering its biology and clinical behavior. RNA velocity provides a causal lens through which to view these relationships.

• Reconstructing Tumor Phylogenies: By ordering cells based on their transcriptional dynamics rather than mere similarity, RNA velocity can be used to infer pseudo-temporal lineages that trace the evolutionary history of a tumor population. This helps resolve the sequence of molecular events leading to aggressive subclones.
• Revealing Cell-of-Origin Phenotypes: Some cancers may arise from distinct cell types of origin, which can impart specific clinical properties. RNA velocity analysis can identify the residual transcriptional momentum that reflects a cell's origin, even after significant transformation. This is achieved by analyzing the phase portraits of key marker genes, where deviations from steady-state relationships indicate active induction or repression, serving as a trace of a cell's recent past [22].

Quantifying Transcriptional Kinetics and Heterogeneity

The speed and heterogeneity of transcriptional processes are key to tumor evolution. Newer RNA velocity models move beyond directionality to quantify the kinetic parameters of gene regulation.

• Inferring Rate Parameters: Advanced methods like scVelo and dynamo infer gene-specific rates of transcription, splicing, and degradation [1] [8]. In a cancer context, comparing these rates between treatment-naive and resistant cells can reveal how tumors rewire their regulatory infrastructure to survive therapy. For example, the widespread dysregulation of RNA velocity observed in Alzheimer's disease suggests that the underlying kinetic rates of transcription are perturbed [23]. A similar analysis in cancer could pinpoint which gene regulatory networks have altered kinetics in response to therapeutic pressure.

Table 1: RNA Velocity Method Categories and Their Applicability to Cancer Research

Category	Key Methods	Underlying Principle	Strengths in Cancer Research	Limitations
Steady-State Methods	Velocyto, scVelo (stochastic)	Assumes constant splicing rate and identifies cells at transcriptional equilibrium [8].	Simple, fast, interpretable; good for clear differentiation trajectories.	Assumptions often violated in highly heterogeneous tumors; inaccurate for complex kinetics.
Trajectory Methods	scVelo (dynamical), UniTVelo, dynamo	Estimates kinetic parameters to align cells along a latent time trajectory [8].	Infers full transcriptional dynamics; can assign latent time and identify key driver genes.	Computationally intensive; model complexity may require deep sequencing data.
State Extrapolation Methods	VeloVAE, LatentVelo, Pyro-Velocity	Leverages expected future states to optimize high-dimensional velocity vectors [8].	Flexible; can incorporate multimodal data and correct for batch effects.	"Black-box" nature can reduce biological interpretability.

Experimental and Computational Protocols

Wet-Lab Protocol: Generating scRNA-seq Data for RNA Velocity

Principle: Successful velocity analysis hinges on high-quality sequencing data that robustly captures both unspliced and spliced mRNA molecules. Most common scRNA-seq protocols (10x Genomics, SMART-seq2, inDrop) are suitable, as a significant portion (15-25%) of their reads typically originate from intronic sequences, representing unspliced pre-mRNA [22].

Procedure:

• Cell Isolation & Lysis: Isolate single cells from fresh tumor dissociates or cryopreserved samples using standard procedures. Lyse cells to release RNA, ensuring RNase inhibition.
• Library Preparation: Proceed with your chosen scRNA-seq protocol. Critical: Use oligo-dT primers for reverse transcription. This enriches for polyadenylated RNA but still captures a substantial amount of intron-containing, unspliced pre-mRNA due to secondary priming events within introns [22].
• Sequencing: Sequence the libraries. To ensure sufficient coverage for velocity analysis, aim for a sequencing depth that is 1.5 to 2 times deeper than standard gene expression studies. High-depth datasets (e.g., ~30,000 UMIs/cell) have been shown to clearly reveal cyclic patterns for dynamic genes [9].

Validation: For method validation, compare velocity estimates from single-nucleus RNA-seq (snRNA-seq) with matched single-cell RNA-seq (scRNA-seq) data. A strong correlation (e.g., 0.94-0.99) between the velocity estimates confirms the precision of the assay [23].

Computational Protocol: A Standard RNA Velocity Workflow

Principle: The computational workflow transforms raw sequencing data (BAM files) into interpretable velocity vectors and trajectories through a series of standardized steps [8].

Procedure:

• Data Quantification: Use tools like Velocyto [22] or kallisto | bustools [8] to quantify unspliced and spliced mRNA counts from aligned BAM files. This generates two count matrices (unspliced, spliced) for the same set of genes and cells.
• Preprocessing & Normalization: Filter cells and genes for quality. Normalize the matrices for library size. Many workflows then log-transform the data and perform principal component analysis (PCA) on the spliced expression matrix.
• Velocity Estimation: Choose a model from Table 1 and estimate RNA velocity.
  • Example using scVelo (Stochastic Model):

• Visualization & Interpretation: Project the velocity vectors onto a low-dimensional embedding (e.g., UMAP, t-SNE).

The following diagram summarizes the key steps and decision points in this standard workflow.

Advanced Integrations and The Scientist's Toolkit

Integrating Spatial Transcriptomics with spVelo

The emergence of spatial transcriptomics technologies allows for the mapping of gene expression within the tissue context. The spVelo framework integrates spatial information to significantly enhance RNA velocity inference in complex tissues like tumors [24].

Principle: spVelo combines a Variational Autoencoder (VAE) for gene expression data with a Graph Attention Network (GAT) that incorporates spatial location proximity. This joint modeling approach leverages the spatial neighborhood of cells to inform the velocity estimation, leading to more accurate and biologically plausible trajectory inferences, especially in multi-batch datasets [24].

Application: In a tumor microenvironment, spVelo can reveal how velocity patterns are spatially organized—for instance, showing a directional flow of differentiating cells from the tumor core to the invasive front, or uncovering localized zones of immune cell activation.

The Scientist's Toolkit: Essential Research Reagents and Tools

Table 2: Key Reagents and Computational Tools for RNA Velocity Analysis

Item Name	Type	Function/Biological Role	Example Use Case
Oligo-dT Primers	Wet-Lab Reagent	Enriches for polyadenylated RNA, capturing both spliced mRNA and unspliced pre-mRNA via intronic priming [22].	Fundamental for library prep in most scRNA-seq protocols to ensure intronic read capture.
10x Chromium Platform	Wet-Lab Platform	High-throughput droplet-based scRNA-seq system. Generates data with sufficient intronic reads for robust velocity analysis [22].	Profiling thousands of cells from a tumor biopsy to characterize cellular heterogeneity and dynamics.
Velocyto	Computational Tool	The pioneering command-line tool for quantifying spliced/unspliced matrices from BAM files [22] [8].	The first step in any standard RNA velocity pipeline to generate the required input data.
scVelo	Computational Tool	A widely-used Python package that generalizes the velocity framework with dynamical and stochastic models [8].	The primary tool for velocity inference, latent time calculation, and identifying key driver genes.
CellRank	Computational Tool	A toolkit that leverages RNA velocity to compute robust transition probabilities and fate likelihoods [1].	Modeling probabilistic fate decisions in branching lineages, such as stem cell differentiation in cancer.
spVelo	Computational Tool	A framework for RNA velocity inference that integrates spatial transcriptomics data [24].	Analyzing spatially-resolved tumor samples to understand the geography of cell state transitions.

RNA velocity has moved from a novel computational concept to an indispensable methodology in single-cell biology. For cancer research, it provides the critical dimension of time, enabling the prediction of cellular futures, the mapping of plasticity routes, and the quantification of transcriptional kinetics directly from static snapshots. As methods continue to advance—integrating spatial information, handling multi-batch designs, and employing more flexible deep-learning models—RNA velocity is poised to deepen our understanding of cancer origins, evolution, and adaptive resistance, ultimately guiding the development of more effective therapeutic strategies.

The advent of single-cell RNA sequencing (scRNA-seq) fundamentally transformed biological research by enabling unprecedented resolution in the examination of cellular heterogeneity. However, a significant limitation remained: standard scRNA-seq provides only static cellular snapshots, obscuring the very dynamic processes that unfold temporally, such as differentiation, reprogramming, and disease progression [1]. The introduction of RNA velocity in 2018 offered a groundbreaking solution to this problem. By leveraging the inherent kinetic information in the ratio of unspliced pre-mRNA to spliced mRNA, RNA velocity models infer instantaneous gene expression change rates and effectively predict future transcriptional states [1] [8]. This concept has rapidly evolved from a foundational model to a suite of sophisticated tools, each refining the original approach to provide more accurate, uncertain-aware, and broadly applicable insights into cellular dynamics. This evolution is particularly critical for cancer research, where understanding the temporal dynamics of tumor evolution, drug resistance, and metastatic transitions can illuminate novel therapeutic vulnerabilities.

The Foundational Paradigm and Its Limitations

The original RNA velocity framework, introduced by La Manno et al. and implemented as Velocyto, rests on a elegant biophysical model [8] [25]. It describes the transcription, splicing, and degradation of mRNA using a system of ordinary differential equations (ODEs). The core idea is that for a given gene, a steady-state ratio of unspliced to spliced mRNA exists. A cell with an abundance of unspliced mRNA above this steady state is predicted to be up-regulating that gene (positive velocity), whereas an abundance below suggests down-regulation (negative velocity) [3]. The combination of velocities across all genes in a cell defines a vector in high-dimensional expression space, predicting the cell's future state [3] [8].

Despite its revolutionary impact, the steady-state assumption proved to be a significant limitation. The model assumes constant transcription, splicing, and degradation rates across all cells, an assumption often violated in complex, heterogeneous systems like tumors, which involve multi-stage and multi-lineage transitions [8] [26]. Furthermore, initial methods lacked any notion of uncertainty quantification, making it difficult to assess the robustness of predictions [17]. These limitations spurred the development of a second generation of computational tools designed to overcome these challenges and expand the applicability of the RNA velocity concept.

The Evolution of RNA Velocity Methods

The landscape of RNA velocity tools has diversified significantly, moving beyond the steady-state model to incorporate more flexible and powerful computational frameworks. The following table summarizes the key evolutionary milestones and the distinct classes of methods that have emerged.

Table 1: Evolution of RNA Velocity Methodologies

Method Class	Representative Tools	Core Innovation	Key Advantages	Primary Limitations
Steady-State Methods [8]	Velocyto [25], scVelo (deterministic/stochastic) [3]	Leverages steady-state ratio of unspliced/spliced mRNA	Simple, fast, and highly interpretable	Fails in non-steady-state or complex kinetic regimes
Trajectory Methods [8]	scVelo (dynamical) [3] [8], UniTVelo [4], Dynamo [8]	Infers full transcriptional dynamics and latent cell time	Relaxes steady-state assumption; infers time	Computationally intensive; sensitive to noise
Deep Learning & Generative Models	cellDancer [26], veloVI [17], VeloVAE [26]	Uses neural networks to infer cell-specific kinetics	Cell-specific parameters; uncertainty quantification	"Black box" nature can reduce interpretability
Multi-Modal & Spatial Extensions	MultiVelo [8], KSRV [27], GraphVelo [28]	Integrates ATAC-seq data or spatial coordinates	Enables velocity inference in spatial context	Relies on accurate data integration
Regulatory-Informed Models	TSvelo [12]	Incorporates TF-regulatory information into ODE model	More biologically accurate dynamics	Requires prior knowledge of TF-target relations

Key Advancements in Methodological Paradigms

• From Global to Local Kinetics: A major leap was the move from global, gene-specific kinetics to cell-specific inferences. Tools like cellDancer employ a "relay velocity model," using deep neural networks to infer transcription, splicing, and degradation rates (( \alpha, \beta, \gamma )) for each cell individually based on its neighbors [26]. This is crucial in cancer, where subpopulations of cells within a tumor may exhibit drastically different transcriptional kinetics.

• Quantifying Uncertainty: The introduction of deep generative models like veloVI provided, for the first time, a robust framework for quantifying uncertainty in velocity estimates. By learning a posterior distribution of RNA velocity, veloVI allows researchers to identify cell states where directionality is estimated with high uncertainty, adding a critical layer of confidence to downstream analyses [17].

• Integrating Regulation and Splicing: Newer frameworks like TSvelo integrate transcriptional regulation directly into the kinetic model. By modeling the transcription rate ( \alpha_g(t) ) as a function of the expression of transcription factors (TFs) that regulate a target gene ( g ), TSvelo provides a more holistic and accurate model of the gene expression cascade [12].

• Towards Multi-Modal and Spatial Velocity: The field is rapidly expanding beyond pure transcriptomics. GraphVelo provides a graph-based framework to project and refine velocity vectors, enabling the inference of multi-modal velocities (e.g., for chromatin accessibility or protein abundance) and ensuring these vectors are consistent with the low-dimensional manifold of the data [28]. Simultaneously, methods like KSRV integrate scRNA-seq with spatial transcriptomics data to infer spatial RNA velocity, allowing researchers to model differentiation trajectories within the anatomical context of a tissue or tumor [27].

Experimental Protocol: A Standard RNA Velocity Workflow

The following protocol describes a standard workflow for RNA velocity analysis using a tool like scVelo, which can be adapted for other methods. This workflow is essential for researchers aiming to apply these techniques to their own single-cell data, such as investigating cancer dynamics.

Preprocessing and Data Loading

• Data Quantification: Begin by quantifying spliced and unspliced transcripts from raw sequencing data (BAM files) using a tool like velocyto.py [25]. This generates a count matrix (often in a loom file) where each cell has counts for both spliced and unspliced molecules for every gene.
• Data Loading and Merging: Load the data into an AnnData object, the standard data structure for single-cell analysis in Python. If you have an existing AnnData object from a standard scRNA-seq analysis, you can merge the velocity data into it [3].

• Filtering and Normalization: Filter genes based on a minimum count threshold and select highly variable genes. Normalize the data for total molecular count per cell and apply a logarithmic transformation to the spliced counts [3].

• Moments Calculation: Compute the first and second-order moments (means and uncentered variances) among nearest neighbors in PCA space. This step performs k-nearest neighbor (k-NN) based smoothing, which is critical for reducing noise and obtaining robust velocity estimates [3] [4].

Velocity Estimation and Interpretation

• Velocity Inference: Estimate RNA velocity vectors. This can be done using the stochastic (mode='stochastic'), deterministic (mode='deterministic'), or more computationally intensive dynamical model (mode='dynamical'), which requires recovering the full gene dynamics first [3].
  The computed velocities are stored in adata.layers [3].
• Velocity Graph Construction: Compute a cell-to-cell transition probability matrix based on the cosine correlation between the velocity vector of a cell and its potential transitions to neighboring cells. This graph forms the basis for all downstream visualizations and analyses [3].

• Visualization: Project the velocities onto a low-dimensional embedding (e.g., UMAP or t-SNE) to visualize the vector field. This is typically done using streamlines, grid arrows, or embedding arrows [3].

• Critical Interpretation with Phase Portraits: Do not base biological conclusions solely on the projected vector field. It is essential to examine the dynamics of individual genes through phase portraits, which plot spliced against unspliced counts for a single gene. This validates how the inferred directionality is supported by specific genes [3] [4].

The following diagram illustrates the key computational and analytical steps of this workflow.

Figure 1: Standard RNA Velocity Analysis Workflow.

Table 2: Key Research Reagent Solutions for RNA Velocity Analysis

Item / Resource	Function / Description	Example Tools / Implementation
Spliced/Unspliced Quantifier	Parses BAM alignment files to distinguish and count spliced vs. unspliced transcripts for each gene.	velocyto.py [25]
Analysis Framework	Provides the core computational environment for data manipulation, model fitting, and visualization.	scVelo (Python) [3], velocyto.R (R) [25]
Kinetic Model	The mathematical engine that fits transcriptional parameters and infers velocity vectors from counts.	Steady-state, Dynamical (scVelo) [3], Deep Learning (cellDancer, veloVI) [26] [17]
Visualization Engine	Projects high-dimensional velocity vectors onto 2D/3D embeddings for intuitive interpretation.	Stream, grid, and embedding plots in scVelo [3]
Reference Datasets	Well-curated public datasets with known dynamics used for method benchmarking and validation.	Pancreas endocrinogenesis [3], Dentate gyrus neurogenesis [8]
Gene Regulatory Database	Source of prior knowledge on TF-target interactions for regulatory-informed models.	ChEA, ENCODE [12]

Application in Cancer Dynamics Research

The application of RNA velocity in oncology provides a powerful lens through which to view the dynamic processes that static analyses miss. Its utility spans several critical areas of cancer biology. RNA velocity can reconstruct the lineage relationships and cellular plasticity within a tumor. For instance, it can help trace the progression from a cancer stem cell state to more differentiated tumor cells, or identify transitional states that exhibit markers of multiple lineages, a phenomenon common in aggressive cancers [1] [8]. Furthermore, by predicting the future state of individual cells, velocity analysis can help identify and characterize pre-resistant cell states that exist within a treatment-naïve tumor. This allows for the study of the transcriptional programs that are activated en route to full-blown drug resistance, potentially revealing targets for combination therapies to block this transition [1]. The integration of RNA velocity with spatial transcriptomics via tools like KSRV or SIRV is particularly powerful for studying the tumor microenvironment [27]. It can model the influx and differentiation of immune cells into the tumor, or the dynamic crosstalk between cancer-associated fibroblasts and tumor cells at the leading edge of a carcinoma, providing spatial context to cellular dynamics.

The field of RNA velocity is advancing at a rapid pace. Key future directions include the tighter integration of additional molecular layers, such as chromatin accessibility (ATAC-seq) and protein expression, to build a more unified and causal model of cellular dynamics [28] [12]. Furthermore, the development of best practices for model selection and uncertainty interpretation will be crucial for robust application in translational settings like cancer research [8] [4] [17].

In conclusion, the evolution of RNA velocity from the foundational Velocyto to the current generation of sophisticated tools has transformed it from a novel concept into an indispensable component of the single-cell omics toolkit. By moving from static snapshots to predictive, dynamic insights, it allows researchers and drug developers to not only see the cellular states present in a tumor but to computationally simulate its trajectory. This paradigm shift holds the promise of uncovering the molecular drivers of cancer progression and therapy failure, ultimately guiding the development of more effective and preemptive cancer therapeutics.

Advanced Tools and Workflows: Applying RNA Velocity to Decipher Cancer Complexity

RNA velocity analysis has emerged as a powerful computational approach for modeling cellular dynamics from single-cell RNA sequencing (scRNA-seq) data. By leveraging the ratio of nascent (unspliced) to mature (spliced) mRNAs, this method enables the recovery of directed dynamic information and the prediction of future cellular states, providing insights into developmental trajectories, lineage commitments, and state transitions that are fundamental to understanding cancer biology [29] [1]. The original concept has evolved into a diverse toolbox of computational methods, each with distinct strengths, modeling assumptions, and applications. For cancer researchers, these tools offer unprecedented opportunities to dissect tumor heterogeneity, identify rare cell populations, map drug resistance pathways, and characterize the cellular hierarchies that drive cancer progression and metastasis.

This article provides a comprehensive overview of the current RNA velocity toolbox, focusing on four prominent tools—scVelo, Dynamo, UniTVelo, and emerging methods—while framing their application within single-cell cancer dynamics research. We present structured comparisons, detailed protocols, and visualization resources to equip cancer researchers and drug development professionals with practical guidance for implementing these cutting-edge analytical techniques.

The Core RNA Velocity Toolbox

Table 1: Core RNA Velocity Methods and Their Applications in Cancer Research

Method	Key Innovation	Modeling Approach	Strengths for Cancer Research	Limitations
scVelo [29]	Generalized dynamical modeling	Expectation-Maximization framework; models transcription, splicing, and degradation kinetics	Identifies putative driver genes and regimes of regulatory change; estimates latent time	Assumes constant kinetic rates; may struggle with complex cancer lineages
Dynamo [30] [31]	Transcriptomic vector field reconstruction	Incorporates metabolic labeling data; differential geometry analysis	Predicts optimal reprogramming paths and in silico perturbation outcomes; maps regulatory networks	Complex implementation; computationally intensive
UniTVelo [32]	Temporally unified latent time	Top-down strategy with radial basis function (RBF) for spliced RNA dynamics	Unifies latent time across transcriptome; handles multi-lineage datasets common in cancer	May oversimplify complex gene-specific dynamics
cellDancer [26]	Cell-specific kinetics via relay velocity model	Deep neural network estimating cell-dependent rates	Infers cell-specific kinetic rates; robust in heterogeneous cancer populations	Computationally demanding for very large datasets
TSvelo [12]	Cascade modeling of regulation, transcription, splicing	Neural Ordinary Differential Equations (ODEs) incorporating TF regulation	Models regulatory interactions; interpretable parameters for mechanistic insights	Requires prior TF-target knowledge which may be incomplete in cancer contexts
cell2fate [33]	RNA velocity module decomposition	Fully Bayesian model with linearization of ODEs	Captures weak dynamical signals in rare cell populations (e.g., cancer stem cells)	Complex model specification; newer method with less extensive validation

Kinetic Models and Rate Estimation

Table 2: Comparison of Kinetic Modeling Approaches Across Methods

Method	Transcription Rate (α)	Splicing Rate (β)	Degradation Rate (γ)	Regulatory Integration
scVelo	Gene-specific, constant or dynamic	Gene-specific, constant	Gene-specific, constant	Not directly incorporated
Dynamo	Estimated from metabolic labeling	Estimated from metabolic labeling	Estimated from metabolic labeling	Through RNA Jacobian analysis
UniTVelo	Derived from spliced RNA function	Derived from spliced RNA function	Derived from spliced RNA function	Not directly incorporated
cellDancer	Cell-specific and gene-specific	Cell-specific and gene-specific	Cell-specific and gene-specific	Not directly incorporated
TSvelo	Modeled as function of TF expression	Gene-specific, constant	Gene-specific, constant	Explicitly models TF regulation
cell2fate	Decomposed into modular components	Gene-specific, constant	Gene-specific, constant	Through transcription rate modules

Methodological Protocols for Cancer Research

Core Computational Workflow

The following diagram illustrates the generalized analytical workflow for RNA velocity analysis in cancer studies, integrating aspects from multiple methods:

Protocol 1: scVelo Analysis for Cancer Cell States

Objective: Identify transitional cancer cell states and predict differentiation trajectories in tumor ecosystems using scVelo's dynamical model.

Materials and Reagents:

• Computing Environment: Python (3.8 or higher) with scVelo, Scanpy, and CellRank packages installed
• Input Data: CellRanger output directories (BAM files and feature-barcode matrices)
• Reference Annotations: Species-appropriate GTF annotation file (e.g., GENCODE vM25 for mouse, Gencode v38 for human)
• Repeat Masking: Repeat region GTF file (e.g., mm10_rmsk.gtf for mouse) [34]

Procedure:

• Data Preprocessing:
  • Process BAM files using Velocyto command line tool to generate spliced and unspliced count matrices:

  • Merge loom files from multiple samples, ensuring barcode consistency across datasets [34].
  • Filter genes and cells based on quality metrics (minimum counts, mitochondrial percentage).

• Velocity Estimation:

  • Preprocess data using scv.pp.filter_and_normalize() and scv.pp.moments().
  • Compute RNA velocity using the dynamical model:

• Visualization and Interpretation:

  • Generate velocity stream plots embedded in UMAP space:

  • Identify putative driver genes using scv.tl.rank_velocity_genes().
  • Estimate latent time with scv.tl.latent_time() to reconstruct temporal sequences.

Cancer Research Application: This protocol can reveal transitional states in tumor development, such as epithelial-to-mesenchymal transition (EMT) intermediates or drug-tolerant persister cells, by identifying regions with consistent velocity flows toward specific phenotypic states.

Protocol 2: Dynamo for In Silico Perturbation in Cancer

Objective: Predict cancer cell fate diversions and identify key regulatory factors using Dynamo's vector field reconstruction and perturbation capabilities.

Materials and Reagents:

• Computing Environment: Python with dynamo-release package installed
• Input Data: Time-resolved metabolic labeling scRNA-seq data (e.g., scNT-seq, scEU-seq) or conventional spliced/unspliced matrices
• Additional Tools: Dynamo notebooks and tutorials from official GitHub repository [30]

Procedure:

• Data Preprocessing and Vector Field Reconstruction:
  • Load and preprocess data using dyn.pp.recipe_monocle().
  • Calculate RNA velocity using dyn.tl.dynamics().
  • Reconstruct continuous vector field with dyn.vf.VectorField().

• Differential Geometry Analysis:

  • Identify fixed points (stable cell states) using dyn.vf.fixed_points().
  • Calculate acceleration and curvature to pinpoint fate decision points.
  • Compute RNA Jacobian to extract cell-state dependent regulatory networks.

• In Silico Perturbation:

  • Predict cell-fate diversion after genetic perturbations:

  • Identify optimal reprogramming paths using least action path method.

Cancer Research Application: This approach enables virtual screening of therapeutic targets by simulating the effects of oncogene knockdown or tumor suppressor reactivation on cell fate decisions, potentially identifying intervention points to divert cells from malignant trajectories.

Protocol 3: UniTVelo for Complex Cancer Lineages

Objective: Resolve complex branching lineages in heterogeneous tumor ecosystems using UniTVelo's unified latent time.

Materials and Reagents:

• Computing Environment: Python with UniTVelo package installed
• Input Data: Spliced and unspliced matrices from complex multi-lineage datasets
• Visualization Tools: Standard scRNA-seq visualization packages (Scanpy, UMAP)

Procedure:

• Data Preprocessing:
  • Follow standard scRNA-seq preprocessing including normalization and highly variable gene selection.
  • Select velocity genes using scv.pp.filter_genes() or method-specific selection.

• Unified Time Inference:

  • Initialize UniTVelo with appropriate mode for cancer dataset:

  • Allow the algorithm to iteratively optimize latent time and kinetic parameters using Expectation-Maximization.

• Multi-Lineage Analysis:

  • Visualize velocity streams on embedding to validate directionality.
  • Compare with known cancer lineage markers to confirm biological relevance.
  • Identify genes with inconsistent dynamics across lineages as potential lineage-specific regulators.

Cancer Research Application: Particularly valuable for dissecting intra-tumor heterogeneity, where multiple subclonal populations coexist with distinct differentiation trajectories, such as in glioblastoma or advanced carcinomas with mixed lineage states.

The Scientist's Toolkit

Table 3: Essential Computational Tools and Resources for RNA Velocity in Cancer Research

Tool/Resource	Function	Application in Cancer Research	Availability
Velocyto.py [34]	Spliced/unspliced matrix generation	Preprocessing of cancer scRNA-seq data for all downstream velocity methods	Python command line tool
Scanpy	scRNA-seq data analysis ecosystem	General data manipulation, visualization, and integration with velocity results	Python package
CellRank [29]	Cell fate mapping using RNA velocity	Identifying terminal states and transition probabilities in cancer cell populations	Python package
TF-target Databases (ChEA, ENCODE) [12]	Regulatory relationship information	Informing models like TSvelo that incorporate transcriptional regulation	Public databases
Metabolic Labeling Data (scSLAM-seq, scNT-seq) [31]	Direct measurement of transcriptional kinetics	Improving velocity accuracy in Dynamo for cancer dynamical processes	Experimental design
Benchmarking Framework [35]	Method comparison and evaluation	Selecting appropriate velocity tools for specific cancer research questions	GitHub repository

Method Selection Guide for Cancer Applications

Decision Framework

The following diagram outlines a strategic approach for selecting the most appropriate RNA velocity method based on specific cancer research questions and data characteristics:

Practical Considerations for Cancer Studies

When applying RNA velocity methods to cancer datasets, several unique challenges emerge:

• High Heterogeneity: Tumors contain diverse cell states with potentially different kinetic regimes. Methods like cellDancer that estimate cell-specific rates may outperform approaches assuming uniform kinetics [26].
• Aneuploidy and Copy Number Variations: Altered gene copy numbers in cancer cells can affect splicing and degradation kinetics. Consider normalizing for CNV effects when possible.
• Complex Lineage Relationships: Cancer progression often involves non-tree-like lineage relationships with convergence, reversal, and cyclic states. Methods with strong multi-lineage capabilities like UniTVelo are advantageous [32].
• Rare Cell Populations: Cancer stem cells or drug-resistant precursors may be rare but critical. cell2fate has demonstrated particular strength in capturing weak dynamical signals from rare populations [33].
• Validation Strategies: Always correlate velocity predictions with known cancer markers, spatial localization (when available), and functional validation where possible.

Future Directions in Cancer Research

The RNA velocity field continues to evolve with several emerging trends particularly relevant for cancer studies:

• Integration with Spatial Transcriptomics: Combining velocity with spatial data can reveal how tumor microenvironment influences cellular trajectories and state transitions.
• Multi-omic Extensions: Velocity concepts are expanding to incorporate epigenetic data (e.g., ATAC-seq) and protein abundance, providing more comprehensive views of cancer cell regulation.
• Therapeutic Application: As models improve, in silico perturbation predictions may help prioritize combination therapies and identify resistance mechanisms before clinical testing.
• Clinical Translation: With standardization and benchmarking, velocity analysis could potentially contribute to diagnostic and prognostic assessments based on differentiation states in tumor ecosystems.

For cancer researchers embarking on RNA velocity analyses, beginning with scVelo provides a solid foundation, while gradually incorporating more specialized tools like Dynamo or cell2fate based on specific research questions and data availability. The protocols and comparisons provided here offer a starting point for leveraging these powerful methods to unravel the dynamic processes driving cancer progression and treatment response.

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to explore cellular heterogeneity, yet it provides only static snapshots of gene expression, obscuring dynamic temporal processes such as differentiation and disease progression [1] [8]. RNA velocity has emerged as a powerful computational concept that addresses this limitation by leveraging the ratio of unspliced (immature) to spliced (mature) messenger RNA to infer the instantaneous rate of gene expression change and predict future cellular states [1] [8]. This approach models transcriptional dynamics using systems of ordinary differential equations (ODEs) based on mRNA splicing kinetics [8].

However, conventional RNA velocity models face significant challenges. They often treat genes independently, failing to incorporate the fundamental biological reality of gene regulatory networks, and struggle with the high noise and short time scales of splicing dynamics [12]. These limitations are particularly problematic in cancer research, where understanding the dynamic regulatory programs driving tumor progression, metastasis, and drug resistance is crucial for therapeutic development.

To address these challenges, researchers have developed next-generation velocity tools that integrate regulatory information. TSvelo (comprehensive RNA velocity by modeling the cascade of gene regulation, transcription and splicing) and TFvelo (gene regulation inspired RNA velocity estimation) represent significant advances that explicitly incorporate transcriptional regulation into RNA velocity frameworks [12] [36] [37]. By modeling the influence of transcription factors (TFs) on target gene expression, these methods provide more accurate reconstructions of cellular dynamics in complex systems, including cancer ecosystems where regulatory programs are frequently rewired.

Methodological Frameworks

TFvelo: Integrating Regulatory Information into Velocity Estimation

TFvelo introduces a fundamental expansion of the RNA velocity concept by incorporating gene regulatory information. Traditional velocity models rely primarily on the phase delay between unspliced and spliced mRNA, which may not provide sufficient signal for robust dynamic modeling across all genes [37]. TFvelo addresses this limitation by modeling the time derivative of RNA abundance while accounting for gene regulatory relationships, enabling more accurate phase portrait fitting for individual genes [36].

The methodological foundation of TFvelo is built upon a generalized Expectation-Maximization (EM) algorithm that iteratively optimizes parameters and latent variables [36] [37]. The key innovation is the incorporation of regulatory information from established TF-target databases such as ChEA and ENCODE to identify potential regulators for each target gene [36]. This allows TFvelo to model transcriptional rates as being influenced by TF expression levels, creating a more biologically grounded framework.

Table 1: Key Hyperparameters in TFvelo Implementation

Hyperparameter	Default Setting	Function
init_weight_method	Correlation	Method to initialize regulatory weights
WX_method	lsq_linear	Method to optimize regulatory weights
n_neighbors	User-defined	Number of neighbors for graph construction
WX_thres	User-defined	Maximum absolute value for regulatory weights
TF_databases	ENCODE & ChEA	Databases for candidate TF selection
max_n_TF	User-defined	Maximum TFs used for modeling each gene
max_iter	User-defined	Maximum iterations in generalized EM algorithm
n_time_points	User-defined	Number of time points in time assignment

The TFvelo workflow begins with data preprocessing similar to standard velocity tools, followed by initialization of regulatory weights using correlation methods. The core algorithm then alternates between assigning latent time to each cell and optimizing the regulatory weights and parameters in the dynamical function [37]. This iterative process continues until convergence, resulting in a model that accurately captures gene dynamics while accounting for regulatory influences.

Figure 1: TFvelo Workflow Integrating Regulatory Information. The diagram illustrates the key steps in TFvelo analysis, highlighting the integration of TF-target databases and the iterative EM algorithm for parameter optimization.

TSvelo: A Comprehensive Mathematics Framework for Regulatory Cascade Modeling

TSvelo represents a more comprehensive framework that models the complete cascade of gene regulation, transcription, and splicing using interpretable neural Ordinary Differential Equations (ODEs) [12]. Unlike approaches that treat genes independently, TSvelo integrates all selected genes into a single unified ODE model, enabling the inference of a global latent time shared across genes within each cell.

The mathematical foundation of TSvelo models the dynamics of unspliced (ug) and spliced (sg) RNA abundance for each gene g using the system:

dug/dt = αg(t) - βg ug

dsg/dt = βg ug - γg s_g

where αg(t), βg, and γg represent the transcription, splicing, and degradation rates, respectively [12]. The key innovation in TSvelo is modeling the gene- and cell-specific transcriptional rate αg(t) as being influenced by the expression of transcription factors that regulate the target gene:

αg(t) = Σ(TF ∈ TFs(g)) W(TF→g) · xTF(t) + b_g

where TFs(g) refers to transcription factors regulating gene g, W(TF→g) represents the regulatory weight, xTF(t) is the expression of the TF, and b_g is a gene-specific bias term [12].

Table 2: Comparative Framework of TFvelo and TSvelo

Feature	TFvelo	TSvelo
Core Approach	Gene regulation-inspired RNA velocity	Comprehensive cascade modeling
Regulatory Integration	TF-target databases (ENCODE/ChEA)	TF-target databases with neural ODEs
Dynamical Modeling	Generalized EM algorithm	Neural ODEs with EM framework
Time Inference	Gene-specific latent time	Unified latent time shared across genes
Splicing Information	Uses unspliced-spliced dynamics	Models full transcription-unspliced-spliced 3D dynamics
Multi-lineage Capability	Not explicitly mentioned	Explicitly designed for multi-lineage datasets
Implementation Base	Built on scVelo framework	Independent implementation

TSvelo employs a neural ODE framework to solve the dynamical system, which is optimized through an EM algorithm that iteratively refines both the latent time and ODE parameters [12]. This approach allows TSvelo to capture complex 3D dynamics in the transcription-unspliced-spliced space, providing a more comprehensive view of gene regulation compared to traditional 2D phase portraits.

Figure 2: TSvelo Comprehensive Framework for Regulatory Cascade Modeling. The diagram illustrates TSvelo's integrated approach combining regulatory information, transcription, and splicing within a unified ODE model optimized through neural ODEs and EM algorithms.

Performance Benchmarks and Biological Applications

Experimental Validation and Performance Metrics

Both TFvelo and TSvelo have undergone rigorous validation using established scRNA-seq datasets to demonstrate their advantages over previous RNA velocity methods. In evaluations using the pancreas development dataset, TSvelo demonstrated superior performance in capturing cell differentiation from ductal to endocrine cells [12]. The method achieved the highest median velocity consistency, in-cluster coherence, and cross-boundary direction correctness, indicating that the high-dimensional velocity vectors learned by TSvelo are more coherent within neighboring cells and better align with ground truth annotations [12].

TSvelo's ability to accurately model 3D gene dynamics provides particular advantages for genes with complex expression patterns. For example, when analyzing the MAML3 gene, TSvelo could distinguish cell types that overlap in traditional 2D phase portraits by incorporating transcriptional representation from multiple TFs [12]. Similarly, for ANXA4, which exhibits a non-monotonic expression pattern (initial decrease followed by increase), TSvelo successfully captured this dynamic where previous methods struggled [12].

TFvelo has been validated on both synthetic data and multiple real scRNA-seq datasets, demonstrating accurate phase portrait fitting and improved pseudotime inference [37]. On synthetic data with known ground truth dynamics, TFvelo achieved high Spearman correlation between inferred and true velocities, as well as between inferred and true regulatory weights [37].

Application to Cancer Biology

The integration of regulatory networks in TFvelo and TSvelo provides particular value for cancer research, where understanding the dynamic regulatory programs driving tumor progression is essential. While the search results do not explicitly detail cancer-specific applications, the methodologies are directly applicable to studying:

• Tumor cell plasticity and state transitions: By modeling the regulatory drivers of cell identity, both methods can help identify TFs responsible for transitions between differentiation states in cancer cells.

• Therapy resistance mechanisms: The dynamics of resistance development often involve regulatory rewiring, which can be captured through regulatory-informed velocity analysis.

• Metastatic progression: The regulatory programs enabling invasion and metastasis can be inferred from primary tumor data using these approaches.

• Tumor heterogeneity: By capturing continuous transitions between cell states, these methods can reveal the regulatory architecture underlying intratumoral diversity.

Experimental Protocols

Protocol 1: Implementing TFvelo Analysis for Regulatory Dynamics

Purpose: To analyze RNA velocity incorporating transcription factor regulation using TFvelo.

Materials and Reagents:

• Single-cell RNA sequencing data (unspliced/spliced counts)
• Computational environment with Python 3.8+
• TFvelo package (https://github.com/xiaoyeye/TFvelo)
• TF-target databases (ENCODE and ChEA)

Procedure:

• Data Preprocessing
  • Load unspliced and spliced count matrices from scRNA-seq data
  • Perform quality control filtering to remove low-quality cells and genes
  • Normalize library sizes and log-transform the data
  • Identify highly variable genes for downstream analysis

• TFvelo Configuration

  • Initialize TFvelo with default hyperparameters:

• Model Training

  • Run the generalized EM algorithm to optimize parameters:
  • The E-step assigns latent time to each cell
  • The M-step optimizes regulatory weights and dynamical parameters
  • Iterate until convergence (typically 50-100 iterations)

• Result Interpretation

  • Visualize velocity stream plots on UMAP embeddings
  • Examine phase portraits for key genes of interest
  • Extract regulatory weights to identify key TF-target relationships
  • Analyze pseudotime ordering for trajectory inference

Troubleshooting Tips:

• If phase portrait fitting is poor, adjust the WX_thres parameter to constrain regulatory weights
• For large datasets, increase n_jobs for parallel processing
• If convergence is slow, increase max_iter or adjust learning rates

Protocol 2: Applying TSvelo for Multi-lineage Dynamics

Purpose: To model complex transcriptional dynamics across multiple lineages using TSvelo.

Materials and Reagents:

• scRNA-seq data with unspliced/spliced counts
• Python environment with PyTorch
• TSvelo implementation (available from original publication)
• High-performance computing resources (for neural ODE optimization)

Procedure:

• Data Preparation
  • Process raw scRNA-seq data to obtain unspliced and spliced matrices
  • Select velocity genes based on expression and dynamics
  • Preprocess data using TSvelo's initialization functions

• Model Configuration

  • Set up the neural ODE architecture for dynamical modeling
  • Configure EM algorithm parameters for joint optimization
  • Specify the number of modules and regulatory constraints

• Model Training

  • Initialize unified latent time for all cells
  • Alternate between neural ODE optimization and time assignment
  • Monitor convergence through loss function stabilization
  • Validate model fit through phase portrait examination

• Downstream Analysis

  • Extract velocity vectors for trajectory inference
  • Identify key regulatory relationships from weight matrices
  • Visualize 3D dynamics for critical genes
  • Perform multi-lineage analysis for complex differentiation processes

Validation Steps:

• Compare velocity consistency with ground truth annotations
• Assess cross-boundary direction correctness between cell types
• Verify that inferred pseudotime matches known developmental ordering

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for Regulatory-Informed Velocity Analysis

Resource	Type	Function in Analysis	Source
ENCODE TF-Target Database	Database	Provides curated transcription factor-target interactions for regulatory network construction	ENCODE Consortium
ChEA Database	Database	Supplies experimentally validated TF-target relationships from chromatin enrichment data	ChEA Project
scRNA-seq Data with Unspliced/Spliced Counts	Data	Raw input data containing both immature and mature mRNA counts for velocity calculation	Cell Ranger, kallisto, BUStools
TFvelo Python Package	Software	Implements regulatory-inspired RNA velocity estimation	GitHub: xiaoyeye/TFvelo
TSvelo Framework	Software	Provides comprehensive cascade modeling of regulation, transcription, and splicing	Contact original authors
Pancreas Development Dataset	Reference Data	Benchmark dataset for validating velocity methods in endocrine differentiation	Original Velocyto paper
Gastrulation Erythroid Dataset	Reference Data	Complex dataset for testing multi-rate kinetics and complex dynamics	Original citation

Discussion and Future Perspectives

The development of TFvelo and TSvelo represents a significant advancement in RNA velocity analysis by moving beyond pure splicing kinetics to incorporate gene regulatory information. These approaches address fundamental limitations in conventional velocity methods, particularly their inability to model the regulatory drivers of transcriptional changes. This is especially valuable in cancer research, where transcriptional regulation is frequently disrupted.

The search results suggest several promising directions for future development. First, there is growing interest in spatial RNA velocity, as evidenced by methods like spVelo, which integrates spatial information with velocity estimation [38]. Combining regulatory information with spatial context could provide unprecedented insights into how cell-cell communication influences regulatory dynamics in tumor microenvironments.

Second, multi-omic integration represents another frontier. While TFvelo and TSvelo primarily leverage transcriptomic data, incorporating epigenetic information such as chromatin accessibility could further improve regulatory network inference. Methods like MultiVelo have begun exploring this direction by integrating scATAC-seq data [8].

Third, uncertainty quantification remains challenging in velocity estimation. Bayesian approaches such as those implemented in VeloCycle and cell2fate offer promising frameworks for assessing confidence in velocity predictions [33] [39]. Future iterations of regulatory-informed velocity methods could incorporate similar statistical rigor.

For cancer researchers, these methodological advances translate to improved ability to model dynamic processes such as therapy resistance emergence, metastatic progression, and tumor cell plasticity. By capturing the regulatory drivers of these transitions, TFvelo and TSvelo provide a more mechanistic understanding of cancer dynamics that could ultimately inform therapeutic strategies targeting these regulatory programs.

TFvelo and TSvelo represent a paradigm shift in RNA velocity analysis by integrating gene regulatory information into dynamical models of transcriptional dynamics. While TFvelo focuses on incorporating TF regulation to improve phase portrait fitting, TSvelo offers a more comprehensive framework modeling the complete cascade from regulatory inputs through splicing kinetics. Both methods demonstrate superior performance compared to previous approaches in benchmarking experiments, particularly for genes with complex dynamics and in multi-lineage settings.

For researchers studying cancer dynamics, these methods provide powerful tools for uncovering the regulatory programs driving tumor progression, heterogeneity, and therapeutic resistance. The experimental protocols outlined in this article offer practical guidance for implementing these analyses, while the toolkit provides essential resources for getting started. As the field advances, we anticipate that regulatory-informed velocity methods will become increasingly central to unraveling the dynamic regulatory architecture of cancer ecosystems.

The tumor microenvironment (TME) is a highly structured ecosystem containing cancer cells surrounded by diverse non-malignant cell types, collectively embedded in an altered, extracellular matrix [40]. Through intricate spatial interactions between multiple components, the TME plays a pivotal role in shaping tumor progression, metastasis, and responses to therapy [40] [41]. While single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity, it provides only static snapshots and loses critical spatial context [38] [41]. RNA velocity has emerged as a powerful concept that overcomes this limitation by leveraging unspliced and spliced mRNA counts to model transcriptional dynamics and predict future cellular states [38] [1]. However, conventional RNA velocity methods face significant challenges in complex TMEs due to batch effects and inability to incorporate spatial information [38] [42]. This creates an pressing need for advanced computational frameworks capable of multi-batch integration and spatial modeling to accurately resolve cellular dynamics in cancer ecosystems.

Computational Frameworks for RNA Velocity Analysis

spVelo (spatial Velocity inference) represents a significant advancement as the first framework designed specifically for RNA velocity inference in multi-batch spatial transcriptomics data [38] [43]. The method integrates a Variational Autoencoder (VAE) for modeling gene expression with a Graph Attention Network (GAT) for incorporating spatial location information [38]. A key innovation is the addition of a Maximum Mean Discrepancy (MMD) penalty between latent spaces of different batches, enabling effective integration of multiple spatial datasets while preserving batch-specific biological signals [38] [43]. This architecture allows spVelo to jointly model spatial location and gene expression data, then estimate kinetic rates and latent time through variational posterior inference [38].

VeloVGI addresses the batch effect challenge in scRNA-seq data through a different approach, combining optimal transport (OT) with mutual nearest neighbors (MNN) to connect similar cells across different batches [42]. The method employs a variational graph autoencoder (VGAE) structure that leverages graph networks to understand relationships between cells, resulting in more accurate predictions of cellular development trajectories [42]. Unlike spVelo, VeloVGI focuses specifically on scRNA-seq data without explicit spatial component integration, making it suitable for larger-scale but spatially-unresolved studies of tumor heterogeneity.

Table 1: Comparative Overview of spVelo and VeloVGI Frameworks

Feature	spVelo	VeloVGI
Primary Data Type	Spatial transcriptomics	Single-cell RNA-seq
Batch Correction	Maximum Mean Discrepancy (MMD) penalty	Optimal Transport + Mutual Nearest Neighbors
Core Architecture	VAE + Graph Attention Network	Variational Graph Autoencoder
Spatial Integration	Explicit via Graph Attention Network	Not applicable
Velocity Estimation	Bayesian deep generative framework	Graph-based network analysis
Uncertainty Quantification	Yes, via posterior distributions	Limited
Key Applications	Complex trajectory patterns, temporal CCC	Lineage inference in batch data

Performance Benchmarking in Tumor Environments

Comprehensive evaluations demonstrate that both methods significantly outperform previous approaches in key metrics relevant to TME analysis. spVelo has been benchmarked using spatial data simulated from mouse pancreas data and real oral squamous cell carcinoma (OSCC) data [38]. When evaluated based on velocity confidence score (measuring reliability of inferred velocities), transition score (assessing probability of true cell-to-cell transition), and direction score (evaluating consistency with known cell type transitions), spVelo consistently achieved the highest average scores across all datasets [38]. Notably, spVelo excelled in direction score, which is particularly important for evaluating velocity's performance in trajectory inference within complex TMEs [38].

Ablation tests conducted by the spVelo developers revealed that integration of spatial information during model training significantly improves performance of velocity and trajectory inference [38]. This highlights the critical importance of spatial context for understanding cellular dynamics in structured environments like tumors, where cell-cell communication and positional relationships drive functional behaviors.

Table 2: Performance Metrics in Tumor-Relevant Contexts

Performance Metric	spVelo Results	VeloVGI Results	Traditional Methods
Velocity Consistency	Superior in local neighborhoods	High agreement across batches	Moderate to poor
Batch Integration	Coherent velocity across batches	Accurate cellular connections	Often fails
Trajectory Accuracy	Complex, non-linear patterns	Reveals complex lineage	Typically linear trajectories
Spatial Alignment	High (by design)	Not applicable	Not applicable
Computational Speed	Efficient for spatial data	Fast on large scRNA-seq	Variable

Experimental Protocols for TME Analysis

spVelo Protocol for Spatial Transcriptomics Data

Sample Preparation and Data Acquisition Begin with fresh-frozen or FFPE tumor tissue sections mounted on spatial transcriptomics platforms such as 10X Genomics Visium, NanoString CosMx, or Vizgen MERSCOPE [40]. For optimal results with spVelo, ensure data includes both spliced and unspliced counts from multiple batches or samples to leverage its multi-batch integration capabilities [38]. The protocol requires matching histological images with spatial coordinate information and gene expression matrices for each batch.

Data Preprocessing

• Log-normalize and smooth the spatial transcriptomics data using standard approaches
• Filter uninformative genes based on their contributions to cell development using spVelo's built-in filtering [38]
• Prepare spatial neighborhood graphs using spatial coordinates, defining cell-cell proximity relationships
• Organize data into annotated batches corresponding to different experimental conditions, time points, or patient samples

Model Implementation and Velocity Inference

Downstream Analysis Applications

• Trajectory Pattern Discovery: Visualize complex trajectory patterns using spVelo's non-linear modeling capabilities
• State Driver Identification: Identify biologically significant marker genes using spVelo's driver marker detection
• Gene Regulatory Network Inference: Construct GRNs specific to spatial domains within the TME
• Temporal Cell-Cell Communication: Model communication dynamics across pseudotime using ligand-receptor interaction databases

VeloVGI Protocol for Multi-Batch scRNA-seq Data

Data Collection and Preprocessing Collect scRNA-seq data from multiple batches of tumor samples, ensuring proper quantification of both spliced and unspliced counts. The protocol is optimized for data from 10X Genomics platforms but can adapt to other scRNA-seq technologies [42].

Batch Correction and Graph Construction

• Identify mutual nearest neighbors (MNN) across batches to define cross-batch cell similarities
• Apply optimal transport (OT) to further refine batch alignment and minimize technical variations
• Construct a unified graph representation connecting cells within and across batches based on transcriptional similarity and MNN relationships
• Define the variational graph autoencoder architecture incorporating both gene expression and graph connectivity

Velocity Estimation and Lineage Inference

Research Reagent Solutions for TME Dynamics

Table 3: Essential Research Reagents and Platforms for RNA Velocity Studies

Reagent/Platform	Function	Application in TME
10X Genomics Visium	Spatial barcoding for transcriptome	Captures spatial context of tumor regions
NanoString CosMx	Single-molecule FISH imaging	High-plex spatial profiling of tumor cells
Vizgen MERSCOPE	Multiplexed error-robust FISH	Subcellular localization in tumor tissues
CODEX	Multiplexed protein imaging	Spatial protein expression in TME
Slide-tags	Spatial barcoding for single cells	Multimodal analysis with spatial positions
DBiT-seq	Microfluidics-based spatial barcoding	Simultaneous transcriptome and proteome
Stereo-seq	DNA nanoball arrays	Large-area TME mapping at nanoscale resolution

Signaling Pathways and Biological Workflows

Applications in Cancer Research and Therapeutic Development

The integration of spVelo and VeloVGI into TME analysis pipelines enables several advanced applications with significant translational potential. These methods facilitate the identification of metastatic cell trajectories and therapy-resistant subpopulations by accurately reconstructing cellular dynamics within tumor ecosystems [38] [44]. For instance, spVelo has demonstrated capability in identifying complex trajectory patterns in oral squamous cell carcinoma, revealing potential transitions between cell states that might be missed by conventional methods [38]. Similarly, VeloVGI's accurate lineage inference in batch-corrected data enables tracking of cancer cell evolution across different tumor regions or time points [42].

A particularly promising application lies in immunotherapy response prediction. By modeling temporal cell-cell communication dynamics, spVelo can identify ligand-receptor interactions that evolve along cancer progression trajectories [38]. This provides insights into how immune cells interact with tumor cells over time, potentially revealing mechanisms of immune evasion or therapy resistance. The ability to quantify uncertainty in velocity estimates further allows researchers to distinguish confident predictions from ambiguous transitional states, which is crucial for clinical translation [38] [17].

These methods also enable drug target discovery through the identification of state driver markers – genes that potentially control critical transitions in tumor development [38]. When coupled with spatial context, these drivers can be mapped to specific TME niches, such as the invasive tumor front or immunosuppressive regions, providing spatially-resolved therapeutic targets [40]. The integration of RNA velocity with gene regulatory network inference further illuminates master regulators of cancer cell states, opening opportunities for network-based therapeutic interventions.

spVelo and VeloVGI represent significant advancements in RNA velocity analysis, specifically addressing the critical challenges of spatial context and batch effects in complex tumor microenvironments. Their development marks a transition from generic trajectory inference to specialized frameworks capable of capturing the intricate spatiotemporal dynamics of cancer ecosystems. As spatial multi-omics technologies continue to evolve, with improvements in resolution and multiplexing capacity [40] [41], the integration of these computational methods will become increasingly essential for extracting biologically meaningful insights from complex TME data.

The future development of RNA velocity methods will likely focus on multi-omic integration, combining not just transcriptomics but also spatial epigenomic, proteomic, and metabolomic data within unified dynamical models. Additionally, clinical translation of these approaches requires further refinement of uncertainty quantification and interpretability features to support diagnostic and therapeutic decision-making. As these tools become more accessible and user-friendly, they will empower cancer researchers to move beyond static cellular cataloging toward truly dynamic understanding of tumor progression, treatment resistance, and metastatic evolution – ultimately enabling more effective therapeutic strategies targeting the vulnerable points in cancer ecosystem dynamics.

Small cell lung cancer (SCLC) is one of the most aggressive malignancies, with a 5-year survival rate below 7% [45]. For decades, the scientific consensus held that SCLC originated from pulmonary neuroendocrine cells (PNECs), which are rare chemosensory cells expressing ASCL1 [45]. However, this paradigm failed to explain the full spectrum of SCLC heterogeneity, particularly the existence of POU2F3-driven tuft-like subsets (SCLC-P) associated with treatment resistance and poor outcomes [45] [46].

This case study details a groundbreaking shift in our understanding of SCLC pathogenesis, demonstrating through integrated genomic approaches that basal stem cells, not neuroendocrine cells, serve as the cell of origin for SCLC and can give rise to its diverse subtypes. This discovery, framed within the context of single-cell RNA sequencing (scRNA-seq) and RNA velocity analysis, reshapes our fundamental understanding of SCLC tumorigenesis and provides new avenues for therapeutic intervention.

Background: SCLC Heterogeneity and Technical Challenges

SCLC comprises distinct molecular subtypes defined by key transcription factors: SCLC-A (ASCL1+), SCLC-N (NEUROD1+), and SCLC-P (POU2F3+) [45]. Human SCLC biopsies frequently demonstrate intratumoral heterogeneity, with co-expression of subtype markers within individual tumors suggesting cellular plasticity between states [45]. Prior to this research, the accepted SCLC cell of origin was the PNEC, yet PNECs with SCLC driver mutations failed to generate SCLC-P in genetically engineered mouse models (GEMMs) [45] [46].

The application of single-cell sequencing technologies has revolutionized cancer research by enabling detailed exploration of cellular heterogeneity, gene regulatory networks, and transcriptional dynamics at unprecedented resolution [47]. Specifically, RNA velocity analysis has emerged as a powerful computational method that models the time derivative of gene expression by linking unspliced (immature) and spliced (mature) mRNA levels through Ordinary Differential Equations (ODEs) to infer cellular trajectory and fate decisions [12].

Experimental Findings: Basal Cells as the Origin of SCLC

Key Evidence from Genetic Mouse Models

Researchers utilized multiple genetically engineered mouse models (GEMMs) of SCLC to investigate the cell of origin question. The study employed Rb1fl/flTrp53fl/flMycT58A (RPM) mice, which harbor key SCLC genetic alterations - RB1 and TP53 loss with MYC activation [45]. To initiate tumors specifically from basal cells, the team combined naphthalene injury (which ablates club cells and expands basal cells) with basal-specific Ad–KRT5 (K5)–Cre [45].

Table 1: Tumor Formation by Cell of Origin in SCLC GEMMs

Cell of Origin	Genetic Approach	SCLC-P (POU2F3+) Formation	Tumor Latency (Days)	Transcriptional Heterogeneity
Basal Cells	KRT5-Cre + RPM	Yes (Enriched)	~53	Broad, including basal/epithelial states
Neuroendocrine Cells	CGRP-Cre + RPM	No	Similar to basal	Restricted
Alveolar/Club Cells	CCSP-Cre + RPM	No	Similar to basal	Moderate
Broad Epithelium	CMV-Cre + RPM	Rare	Similar to basal	Intermediate

The critical finding was that KRT5-initiated tumors exhibited extensive SCLC subtype heterogeneity, including POU2F3+ tumors that were enriched in tracheal and main airways and more abundant than from other cells of origin [45]. In contrast, tumors initiated from neuroendocrine cells (CGRP-Cre) or alveolar/club cells (CCSP-Cre) in the same RPM background failed to generate POU2F3+ tumors [45]. Single-cell RNA sequencing of KRT5-initiated versus CGRP-initiated RPM tumors revealed broader transcriptional heterogeneity in basal-derived tumors, including clusters that retained basal/epithelial markers and expressed Ascl1 and 'A2' ('NEv2') signatures associated with drug-resistant and immune-modulatory states [45].

Organoid Models Confirm Lineage Plasticity

To more directly test the basal cell of origin hypothesis, researchers isolated normal tracheal basal cells from RPM GEMMs using surface ITGA6 expression and cultured them as 3D organoids [45]. After transformation with Ad5–CMV–Cre, these organoids developed into tumors that retained basal markers (P63, KRT5) while also expressing neuroendocrine (ASCL1, NEUROD1) and tuft-cell (POU2F3) lineage markers [45]. Lineage barcoding techniques allowed tracking of individual cells, revealing that SCLC cells can "shapeshift" through cell fate plasticity, explaining how the disease resists treatment [46].

Analysis of Human SCLC Tumors

The study extended its findings to human disease through analysis of 944 human SCLCs, which revealed a basal-like subset and a tuft–ionocyte-like state demonstrating notable conservation between cancer states and normal basal cell injury response mechanisms [45]. Immunohistochemistry analysis of 119 human SCLC biopsies showed that approximately 19% were POU2F3+, and among these, more than 82% also expressed ASCL1 and/or NEUROD1, supporting widespread intratumoral heterogeneity [45].

Methodologies and Protocols

Single-Cell RNA Sequencing Workflow

Table 2: Key scRNA-seq Wet Lab Protocols

Step	Protocol Details	Purpose
Tissue Dissociation	Minced tissue digested with GentleMACS Dissociator in ice-cold H1640 [48]	Generate single-cell suspension
Cell Partitioning	Chromium Controller (10x Genomics) with barcoded Gel Beads [48]	Individual cell barcoding
Library Preparation	Reverse transcription in GEMs, cDNA amplification, fragmentation, adapter ligation [48]	Sequencing-ready libraries
Quality Control	CellRanger alignment to reference genome; Filtering: mitochondrial counts <25%, >500 genes/cell [48]	Remove low-quality cells/doublets

RNA Velocity Analysis

Advanced RNA velocity methods were essential for understanding the lineage trajectories underlying neuroendocrine-tuft plasticity. The research employed cutting-edge computational frameworks like TSvelo, which models the cascade of gene regulation, transcription, and splicing using interpretable neural Ordinary Differential Equations (ODEs) [12]. The fundamental RNA velocity equation is:

Where u and s are unspliced and spliced mRNA abundance, α(t) is transcription rate, β is splicing rate, and γ is degradation rate [12]. TSvelo enhances this by modeling the transcription rate α(t) as influenced by transcription factor expression:

Where TFs(g) are transcription factors regulating target gene g, x_tf are TF expression levels, and w_tf are regulatory weights [12].

Computational Analysis Pipeline

For Bioconductor-based RNA velocity analysis, researchers can implement the following protocol using the velociraptor package:

This pipeline enables the projection of velocity vectors into low-dimensional embeddings, revealing the direction and magnitude of cellular state transitions [49].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Research Reagents for SCLC Lineage Plasticity Studies

Reagent/Resource	Function/Application	Example Usage in Study
RPM GEMM (Rb1fl/flTrp53fl/flMycT58A)	Models human SCLC genetics	Tumor initiation from different cells of origin [45]
Cell-type specific Cre lines (KRT5-Cre, CGRP-Cre, CCSP-Cre)	Targeted tumor initiation	Determine which cells can give rise to which SCLC subtypes [45]
ITGA6 antibody	Basal cell isolation via FACS	Purify basal cells for organoid culture [45]
10x Genomics Chromium Platform	Single-cell RNA sequencing	Transcriptomic profiling of tumor heterogeneity [48]
TF-target databases (ChEA, ENCODE)	Regulatory network inference	Identify TF-target relationships for velocity models [12]
Velociraptor / scVelo packages	RNA velocity analysis	Estimate differentiation trajectories from scRNA-seq [49]

Signaling Pathways and Molecular Mechanisms

The molecular characterization of basal-derived SCLC revealed that cooperation between specific genetic alterations enriched in human tuft-like SCLC—including high MYC, PTEN loss, and ASCL1 suppression—uniquely promotes tuft-like tumors when introduced into basal cells [45]. This explains the previously mysterious connection between MYC amplification and SCLC-P prevalence in human tumors.

Discussion and Implications

The identification of basal cells as the origin of SCLC represents a paradigm shift in cancer biology with significant clinical implications. This discovery explains the remarkable heterogeneity observed in human SCLC, particularly the existence of tuft-like (SCLC-P) subsets that had been difficult to model experimentally [45] [46]. The basal cell of origin hypothesis is further strengthened by epidemiological observations—tobacco smoke, the primary SCLC risk factor, increases basal cell proliferation and metaplastic potential [45].

From a therapeutic perspective, this research creates the first accurate lab models of the most treatment-resistant tuft-like form of SCLC, enabling studies of early detection and targeted therapies [46]. The findings suggest that targeting the basal cell state or the mechanisms underlying lineage plasticity may provide new approaches to combat therapy resistance. As senior author Trudy G. Oliver noted, "We now have the tools to explore how the immune system interacts with these basal cells before they transform into aggressive cancer. That opens the door to therapies that could stop the disease before it even starts" [46].

This case study exemplifies how integrating advanced single-cell technologies—particularly RNA velocity analysis—with sophisticated genetic models can unravel complex problems in cancer biology, ultimately paving the way for improved therapeutic strategies against this devastating disease.

RNA velocity analysis has emerged as a powerful computational method for inferring dynamic cellular state changes from static single-cell RNA sequencing (scRNA-seq) data. By quantifying the ratio of unspliced (nascent) to spliced (mature) messenger RNA, this approach can predict the future transcriptional state of individual cells, thereby reconstructing cellular trajectories in complex biological processes. In cancer research, this provides an unprecedented window into tumor evolution, drug resistance mechanisms, and metastatic progression, moving beyond static snapshots to dynamic predictions of cell fate decisions.

The application of RNA velocity is particularly valuable for understanding cancer dynamics, as it can reveal the directionality of cellular state transitions within heterogeneous tumors. This enables researchers to identify key driver genes and regulatory programs controlling critical transitions, such as the emergence of drug-resistant subpopulations or the acquisition of metastatic potential. When framed within single-cell cancer dynamics research, RNA velocity serves as a bridge between observed transcriptional states and the underlying dynamical systems governing tumor progression and treatment response.

Key Computational Methods and Tools

Recent advances in RNA velocity estimation have produced several sophisticated computational frameworks, each with distinct approaches to modeling transcriptional dynamics.

Table 1: Comparison of RNA Velocity Estimation Methods

Method	Key Innovation	Application Context	Strengths
Velocyto [5]	First proposed RNA velocity using steady-state ODE model	General purpose; initial proof of concept	Simple, interpretable parameters
scVelo [5]	EM algorithm for ODE parameter estimation; dynamic model	General cellular trajectories	Improved kinetics estimation; multiple inference modes
UniTVelo [5] [12]	Unified pseudotime across genes; radial basis function fitting	Complex differentiation systems	Gene-shared time increases consistency
TIVelo [5]	Cluster-level trajectory inference before single-cell estimation	Datasets with clear cluster structure	Avoids ODE assumptions; robust to complex patterns
TSvelo [12]	Integrates transcriptional regulation, transcription, and splicing	Multi-lineage systems; regulatory inference	Models TF regulation explicitly; highly interpretable
VeloViz [50]	RNA velocity-informed embeddings for visualization	Trajectory visualization	Preserves trajectory topology better than standard embeddings

Technical Protocols for RNA Velocity Analysis

Protocol 1: Standard RNA Velocity Pipeline for Tumor Samples

• Sample Preparation and Sequencing

  • Perform single-cell RNA sequencing using 3' or 5' end-counting protocols (e.g., 10X Genomics) that preserve strand information.
  • Ensure sufficient sequencing depth to capture both spliced and unspliced transcripts (typically >20,000 reads/cell).
  • Include spike-in RNAs if absolute molecule counts are required.

• Data Preprocessing

  • Extract spliced and unspliced counts using velocyto.py run as a second step after alignment with cellranger.
  • Filter cells based on quality metrics: mitochondrial percentage (<20%), number of detected genes (500-5000), and total UMI counts.
  • Filter genes to those detected in a minimum number of cells (typically >10 cells).

• RNA Velocity Estimation

  • Select appropriate method based on experimental system (refer to Table 1 for guidance).
  • For TIVelo: Perform clustering (Louvain/Leiden), then trajectory inference, followed by orientation score calculation [5].
  • For TSvelo: Incorporate transcription factor-target database information to model regulatory cascades [12].
  • For standard analyses: Use scVelo with dynamical model for most accurate kinetics estimation.

• Visualization and Interpretation

  • Project velocity vectors onto embedding (UMAP, t-SNE) to visualize flow fields.
  • Use VeloViz to create velocity-informed embeddings that better preserve trajectory topology [50].
  • Identify key driver genes by correlating velocity vectors with gene expression changes.
  • Validate biological hypotheses with experimental perturbations where possible.

Applications in Cancer Drug Discovery

Target Identification and Validation

RNA velocity analysis enables systematic identification of novel therapeutic targets by revealing transitional cell states and their regulatory drivers during tumor progression.

Table 2: Target Identification Through Single-Cell Technologies

Tumor Type	Single-Cell Technology	Identified Target	Therapeutic Approach	Reference
Glioblastoma	scRNA-seq	Wnt signaling	XAV-939 (Wnt inhibitor) blocks CTC-mediated recolonization	[51]
Multiple Myeloma	scRNA-seq	PPIA	Ciclosporin overcomes Dara-KRd resistance	[51]
Hepatobiliary Tumor	scRNA-seq (organoids)	NEAT1	Targeting metabolic reprogramming in resistant subpopulation	[51]
Pediatric AML	scRNA-seq + scATAC-seq	MEF2C	Targeting enhanced transcriptional activation in resistant cells	[51]
Lung Tumor	scRNA-seq	TIGIT	Immunotherapy target identified in stem cells	[51]
Gastric Adenocarcinoma	scRNA-seq	SOX9/LIFR	EC359 targets LIF/LIFR signaling in CSCs	[51]

Protocol for Target Discovery in Heterogeneous Tumors

Protocol 2: Identifying Transition-Specific Therapeutic Targets

• Sample Collection and Processing

  • Collect longitudinal tumor samples when possible (pre-treatment, on-treatment, progression).
  • Process fresh tissue within 1 hour of collection to maintain RNA integrity.
  • Include adjacent normal tissue as reference for identifying tumor-specific patterns.

• Single-Cell Profiling

  • Perform scRNA-seq using platforms that capture full transcript length (Smart-seq3) when studying isoform switching, or 3' end-counting (10X) for larger cell numbers.
  • For regulatory insight, combine with scATAC-seq (multiome) or CITE-seq for surface protein expression.

• Trajectory Analysis

  • Reconstruct tumor evolution trajectories using RNA velocity (preferred methods: TIVelo, TSvelo).
  • Identify branching points representing fate decisions (drug resistance, metastasis, dormancy).
  • Pseudotime-align cells to order them along identified trajectories.

• Target Prioritization

  • Identify genes with expression changes correlated with velocity vectors at critical transitions.
  • Use trajectory-based differential expression (tradeSeq, impulse-like models) to find transition-specific markers.
  • Integrate with TF activity inference (pySCENIC, DoRothEA) to identify key regulators.
  • Validate candidates using CRISPR screening data where available.

Integration with Personalized Therapy Development

The scTherapy computational framework exemplifies how RNA velocity-informed trajectories can guide personalized treatment strategies. This approach uses single-cell transcriptomic profiles to prioritize multi-targeting treatment options for individual cancer patients by leveraging a pre-trained gradient boosting model (LightGBM) on large-scale drug perturbation data [52].

Protocol 3: Developing Patient-Specific Combination Therapies

• Single-Cell Profiling of Patient Tumor

  • Process fresh tumor biopsy within 2 hours of collection.
  • Perform scRNA-seq to characterize cellular heterogeneity and identify distinct clones.
  • Annotate cell types (malignant, immune, stromal) and subset malignant cells for clone identification.

• Trajectory Analysis of Resistance Pathways

  • Perform RNA velocity analysis to predict natural progression trajectories.
  • Identify transitional states toward resistant phenotypes.
  • Map existing clones onto trajectory to predict which may expand under treatment.

• Drug Response Prediction

  • Input transcriptomic profiles of malignant clones into pre-trained scTherapy model.
  • Predict dose-specific responses for each clone across comprehensive drug library.
  • Identify drugs that selectively target multiple clones simultaneously while sparing normal cells.

• Combination Therapy Design

  • Select drug pairs that co-inhibit major malignant clones with complementary mechanisms.
  • Prioritize combinations where drugs target different transitional states in resistance trajectories.
  • Validate predictions ex vivo using patient-derived organoids or primary cells when available.

Experimental validation of this approach in AML patient samples demonstrated that 96% of predicted multi-targeting treatments exhibited selective efficacy or synergy, with 83% showing low toxicity to normal cells [52]. This highlights the potential of RNA velocity-informed therapy selection to improve clinical outcomes in heterogeneous cancers.

Research Reagent Solutions

Table 3: Essential Research Tools for RNA Velocity in Cancer Studies

Category	Specific Tool/Reagent	Function	Considerations for Cancer Studies
Single-Cell Isolation	Fluorescence-Activated Cell Sorting (FACS)	High-throughput cell separation based on surface markers	Enables purification of rare populations (CTCs, stem cells)
	Microfluidic Platforms	Automated single-cell encapsulation	Ideal for limited tumor biopsy material
Library Preparation	10X Genomics Chromium	High-throughput scRNA-seq with UMIs	Standard for large cell numbers; preserves strand info
	Smart-seq3	Full-length transcript coverage	Better for isoform analysis; lower throughput
Computational Tools	Velocyto.py	Initial spliced/unspliced counting	First step in standard RNA velocity pipeline
	scVelo	Dynamical modeling of RNA velocity	Python-based; extensive customization options
	TIVelo	Cluster-first velocity estimation	Avoids ODE assumptions; robust for complex tumors
	VeloViz	Velocity-informed embeddings	Superior trajectory visualization
Data Resources	LINCS Database	Drug perturbation transcriptomics	Training data for therapy prediction models
	PharmacoDB	Drug sensitivity data	Correlates transcriptomic changes with viability

Future Perspectives and Challenges

As single-cell technologies continue to evolve, several emerging trends promise to enhance the application of RNA velocity in cancer research. Spatial transcriptomics technologies are beginning to incorporate temporal dynamics, allowing researchers to map cellular trajectories within the architectural context of tumors. Multi-omic approaches simultaneously measuring chromatin accessibility (scATAC-seq), protein expression (CITE-seq), and transcriptional states will provide more comprehensive views of the regulatory networks driving cancer progression.

The integration of RNA velocity with CRISPR-based lineage tracing represents a particularly promising direction, enabling direct validation of predicted trajectories. Additionally, machine learning approaches that can learn complex, non-linear dynamics from sparse single-cell data will likely overcome current limitations of ODE-based models. For clinical translation, efforts to reduce computational complexity and improve interpretability for non-specialists will be essential for adopting these methods in diagnostic settings.

Major challenges remain in handling the scale and noise of single-cell data, integrating multimodal measurements, and validating predictions experimentally. Furthermore, applying these methods to solid tumors presents additional difficulties due to cellular complexity, tissue dissociation artifacts, and spatial constraints on cell state transitions. Despite these challenges, RNA velocity analysis stands to fundamentally transform our understanding of cancer dynamics and accelerate the development of more effective, personalized cancer therapies.

Navigating Pitfalls: Best Practices for Robust RNA Velocity Analysis in Cancer

RNA velocity analysis has emerged as a powerful computational technique for inferring directional dynamics and future cell states from single-cell RNA sequencing (scRNA-seq) data. By connecting measurements to the underlying kinetics of gene expression, this approach has opened new avenues for studying cellular dynamics in cancer research, including tumor evolution, drug resistance, and metastatic behavior. However, significant challenges persist that limit the reliability and biological interpretation of velocity estimates. This application note examines three fundamental challenges—data sparsity, technical noise, and violations of core model assumptions—within the context of cancer dynamics research. We provide structured comparisons of these limitations, detailed protocols for robust velocity estimation, and practical solutions to enhance analytical workflows for researchers and drug development professionals.

RNA velocity quantifies the time derivative of gene expression by modeling the relationship between unspliced (nascent) and spliced (mature) mRNA molecules, enabling the prediction of cellular trajectories from snapshot scRNA-seq data [53]. In cancer research, this technique offers unique potential for delineating tumoral and microenvironmental evolution, identifying rare cell populations such as cancer stem cells, and understanding therapeutic resistance mechanisms [54]. The transcriptional dynamics captured by RNA velocity can reveal the directionality of cancer progression, metastatic transitions, and responses to treatment pressures at single-cell resolution.

However, the application of RNA velocity to cancer datasets faces particular obstacles due to the inherent biological complexity of tumors. Intratumoral heterogeneity, diverse cellular states, and varying kinetic regimes across subpopulations complicate velocity estimation [53] [54]. Additionally, technical limitations of single-cell technologies introduce analytical challenges that can compromise velocity inference. Understanding these constraints is essential for generating biologically meaningful insights from RNA velocity analysis in cancer studies.

Critical Challenges in RNA Velocity Analysis

Data Sparsity and Technical Noise

The limited abundance of unspliced mRNA molecules presents a fundamental constraint for RNA velocity estimation. Unspliced transcripts typically constitute only 10-25% of total mRNA molecules detected in standard scRNA-seq protocols, resulting in sparse measurements that hinder accurate kinetic fitting [3]. This sparsity is exacerbated in cancer datasets due to frequent transcriptomic heterogeneity and the presence of rare, transient cell states that drive tumor evolution and therapeutic resistance [55].

Technical noise in scRNA-seq data further compounds these challenges, particularly affecting the quantification of unspliced counts which are more susceptible to amplification biases and detection limitations [12] [56]. The high noise levels can obscure the underlying phase portraits that are essential for velocity estimation, leading to unreliable trajectory predictions. In cancer research, where identifying rare subclones is critical for understanding resistance mechanisms, these limitations can significantly impact the utility of RNA velocity analysis.

Violations of Core Model Assumptions

Standard RNA velocity models rely on simplifying assumptions that are frequently violated in biological systems, particularly in complex cancer microenvironments:

Table 1: Key Model Assumptions and Their Common Violations in Cancer Data

Model Assumption	Biological Reality in Cancer	Impact on Velocity Estimation
Constant kinetic rates across cells	Time-dependent rates due to regulatory changes [53]	Incorrect directionality estimates [4]
Gene-independent dynamics	Coordinated regulation through gene networks [12]	Loss of multivariate information
Observation of steady states	Continuous transitions in tumor evolution [16]	Unreliable parameter fitting
Single kinetic regime per gene	Multiple regimes across subpopulations [53]	Contradictory velocity vectors

The assumption of constant kinetic rates is particularly problematic in cancer contexts, where transcriptional bursts, metabolic changes, and regulatory rewiring create dynamic kinetic regimes [53]. For example, in erythroid maturation, a boost in expression has been observed that leads to incorrect negative velocity estimates when using standard models [53]. Similarly, the gene-independent treatment of dynamics ignores crucial regulatory relationships that coordinate expression changes in oncogenic pathways.

Quantitative Comparison of Computational Solutions

Several computational methods have been developed to address these challenges, each with distinct approaches to handling sparsity, noise, and model violations:

Table 2: Computational Methods for Addressing RNA Velocity Challenges

Method	Core Approach	Sparsity/Noise Handling	Model Assumption Flexibility	Cancer Application Suitability
TSvelo [12]	Neural ODEs modeling regulation, transcription, and splicing	Simultaneous modeling of all genes improves robustness	Incorporates regulatory cascades; lineage-specific kinetics	High (explicitly handles multi-lineage datasets)
TFvelo [56]	Gene regulation-inspired using TF-target relationships	Uses spliced counts only, avoids unspliced sparsity issues	Replaces splicing kinetics with regulatory dynamics	High (applicable to datasets without splicing information)
VeloAE [57]	Autoencoder-based representation learning	Denoising through low-dimensional embedding	Maintains standard kinetics but with enhanced robustness	Medium (improves consistency but retains core assumptions)
scVelo [3] [16]	Dynamical modeling with EM algorithm	K-NN smoothing of spliced/unspliced counts	Relaxes steady-state assumption; estimates cell-specific times	Medium (sensitive to noise in sparse datasets)
UniTVelo [12]	Unified latent time with empirical modeling	Gene-shared latent time circumvents individual gene noise	Top-down modeling with flexible dynamics	Medium (requires trajectory-like structures)

The performance of these methods varies significantly across different cancer datasets. Benchmarking studies indicate that methods incorporating regulatory information (TSvelo, TFvelo) generally show improved performance in complex cancer ecosystems with multiple lineages and heterogeneous subpopulations [12] [56]. Representation learning approaches (VeloAE) demonstrate enhanced robustness to technical noise, particularly in datasets with low sequencing depth [57].

Experimental Protocols for Robust Velocity Estimation

Protocol 1: Regulation-Informed Velocity Analysis with TFvelo

This protocol enables RNA velocity estimation even when splicing information is unavailable or excessively sparse, leveraging transcription factor-target relationships instead of splicing kinetics.

Materials and Reagents:

• scRNA-seq dataset (spliced counts only suffice)
• TF-target database (ENCODE or ChEA recommended)
• Computational environment with TFvelo installed

Procedure:

• Data Preprocessing
  • Filter cells by quality metrics (mitochondrial percentage, feature counts)
  • Normalize spliced counts by cell size
  • Identify highly variable genes (2000-3000 genes recommended)

• TF-Target Mapping

  • Retrieve TF-target interactions from ENCODE database
  • Filter interactions by expression correlation if desired
  • Construct regulatory matrix for velocity genes

• Model Initialization

  • Initialize latent time using diffusion pseudotime
  • Set initial parameters for regulatory weights
  • Define expectation-maximization convergence criteria

• Iterative Optimization

  • E-step: Update latent time assignments given current parameters
  • M-step: Re-estimate regulatory weights given current time assignments
  • Repeat until convergence (typically 50-100 iterations)

• Velocity Projection and Visualization

  • Compute high-dimensional velocity vectors
  • Project velocities to low-dimensional embedding (UMAP/t-SNE)
  • Visualize as streamlines or grid arrows

Troubleshooting Tips:

• If velocities appear uniform, adjust regularization strength in weight estimation
• For unstable time assignments, increase neighborhood size in smoothing
• If specific lineages are missing, verify TF coverage in regulatory database

Protocol 2: Comprehensive Dynamics with TSvelo

This protocol models the complete cascade of gene regulation, transcription, and splicing using neural ordinary differential equations, particularly suited for complex cancer datasets with multiple lineages.

Materials and Reagents:

• scRNA-seq dataset with spliced and unspliced counts
• TF-target database (ENCODE or ChEA)
• High-performance computing environment with GPU acceleration

Procedure:

• Data Preprocessing and Gene Selection
  • Filter genes by minimum shared spliced/unspliced counts (≥20)
  • Normalize both spliced and unspliced counts by cell size
  • Select velocity genes based on dispersion and regulation evidence

• Neural ODE Architecture Setup

  • Configure ODE network with two hidden layers (64-128 nodes each)
  • Set activation function (tanh or sigmoid recommended)
  • Initialize kinetic parameters (transcription, splicing, degradation rates)

• Multi-Objective Optimization

  • Balance fitting losses: unspliced-spliced dynamics and regulatory constraints
  • Set relative weights for different objective components
  • Implement gradient-based optimization (Adam optimizer)

• Unified Latent Time Inference

  • Initialize pseudotime using initial state probabilities
  • Refine time estimates through iterative alignment with dynamics
  • Resolve conflicts between gene-specific times

• Lineage-Specific Velocity Estimation

  • Identify branching points in cancer progression trajectories
  • Estimate lineage-specific kinetic parameters where supported by data
  • Compute velocity vectors conditioned on lineage assignment

Validation Steps:

• Verify phase portraits for key marker genes show expected dynamics
• Confirm velocity consistency within biologically homogeneous regions
• Check that terminal states align with known cancer cell types

Visualization of RNA Velocity Workflows

Standard RNA Velocity Analysis with Quality Control

Standard RNA Velocity Pipeline

Advanced Regulation-Informed Velocity Estimation

Regulation-Informed Velocity Analysis

Research Reagent Solutions

Table 3: Essential Computational Tools for RNA Velocity in Cancer Research

Tool/Resource	Type	Primary Function	Application Context
scVelo [3]	Python package	Dynamical RNA velocity estimation	General purpose cancer trajectory analysis
Velocyto [53]	Command line tool	Spliced/unspliced count quantification	Preprocessing for velocity workflows
ENCODE TF Targets [56]	Database	Curated transcription factor-target interactions	Regulation-informed velocity methods
CellRank [4]	Python package	Cellular fate probability estimation	Terminal state identification in tumors
Scanpy [16]	Python package	Single-cell data analysis ecosystem	General scRNA-seq preprocessing and visualization
Dynamo [12]	Python package	Metabolic labeling-integrated velocity	High-resolution kinetic modeling

RNA velocity analysis represents a transformative approach for unraveling cancer dynamics, yet its effective application requires careful attention to data sparsity, noise, and model limitations. The integration of regulatory information with splicing kinetics, as implemented in methods like TSvelo and TFvelo, shows particular promise for addressing these challenges in complex cancer ecosystems. As single-cell technologies continue to evolve, incorporating multi-omic measurements and spatial context will further enhance the resolution and biological validity of velocity estimates. For cancer researchers and drug development professionals, adopting robust computational workflows with appropriate quality controls is essential for generating reliable insights into tumor evolution, metastasis, and therapeutic resistance.

In single-cell cancer dynamics research, the ability to accurately reconstruct transcriptional trajectories through RNA velocity analysis is fundamentally constrained by technical variability introduced when integrating data from multiple samples, studies, and technological platforms. Batch effects—systematic technical variations arising from differences in sequencing technologies, laboratory conditions, reagent batches, or experimental protocols—represent a critical analytical challenge that can obscure true biological signals and lead to spurious scientific conclusions [58] [54]. These effects are particularly problematic in cancer research, where subtle transcriptional dynamics within tumor heterogeneity, tumor microenvironment interactions, and therapeutic resistance mechanisms must be precisely characterized across patients and disease states [54].

The integration of multi-sample and multi-study single-cell data has become increasingly essential for robust biological discovery in oncology. Large-scale collaborative efforts such as the Human Cell Atlas and various cancer atlas initiatives generate data across multiple institutions and platforms, creating an urgent need for effective integration strategies that can separate technical artifacts from biologically meaningful variation [59]. When performing RNA velocity analysis—which predicts future cellular states by leveraging the ratio of unspliced to spliced mRNA—batch effects can significantly distort the inferred velocity vectors and trajectory directions, potentially misrepresenting the dynamic processes underlying cancer progression and treatment response [8] [38].

This protocol outlines comprehensive strategies for overcoming batch effects in single-cell RNA sequencing data, with particular emphasis on applications in cancer dynamics research. We provide structured comparisons of integration methods, detailed experimental protocols, and specialized considerations for maintaining velocity integrity throughout the integration process, enabling researchers to extract biologically accurate insights from complex multi-study datasets.

Foundational Concepts and Method Classification

Categories of Batch Effect Correction Methods

Computational methods for batch effect correction in single-cell data can be broadly classified into several categories based on their underlying mathematical frameworks and integration strategies. Matrix factorization approaches identify shared biological factors across datasets while isolating technical variations, while nearest neighbor-based methods establish connections between similar cells across batches to guide alignment [60] [54]. Deep learning-based methods have emerged more recently, using neural network architectures to learn nonlinear mappings that align datasets while preserving biological heterogeneity [61].

The mutual nearest neighbors (MNN) approach, first introduced by Haghverdi et al., identifies pairs of cells from different batches that are transcriptionally similar and uses these "anchors" to correct batch effects [62]. This method does not assume identical cell type compositions across batches, making it particularly suitable for cancer datasets where cellular heterogeneity may vary significantly between patients or disease stages. Subsequent developments like Scanorama and Conos extended this concept to multiple datasets, addressing ordering dependencies in the original MNN implementation [54].

Deep learning methods such as scVI employ variational autoencoders to learn a batch-invariant latent representation of the data, while Harmony uses an iterative clustering approach to maximize dataset diversity within clusters [60]. The recently introduced scMerge2 algorithm incorporates hierarchical integration, pseudo-bulk construction, and pseudo-replication to enable atlas-scale integration of millions of cells from complex multi-condition studies [59].

Integration Method Performance Characteristics

Table 1: Comparison of Selected Batch Effect Correction Methods

Method	Underlying Approach	Strengths	Limitations	Scalability
Harmony [58] [60]	Iterative clustering with diversity maximization	Fast runtime, good preservation of fine populations	May overcorrect with highly disparate batches	Excellent (tested on 1M+ cells)
LIGER [58] [60]	Integrative non-negative matrix factorization	Separates shared and dataset-specific factors	Requires parameter tuning, longer runtime	Good (tested on 500K+ cells)
Seurat 3 [58] [60]	CCA with mutual nearest neighbors	Returns adjusted expression matrix, good performance	Can be memory intensive for very large datasets	Good (tested on 500K+ cells)
scMerge2 [59]	Hierarchical factor analysis with pseudo-bulk	Preserves multi-condition signals, efficient for large data	Requires careful parameter selection	Excellent (tested on 5M+ cells)
MNN Correct [62]	Mutual nearest neighbors	No assumption of identical composition	Result depends on dataset order	Moderate
scVI [60]	Variational autoencoder	Probabilistic framework, handles sparse data	Complex implementation, requires GPU for large data	Good

Benchmarking studies have comprehensively evaluated these methods across multiple metrics, including batch mixing quality, biological preservation, computational efficiency, and scalability. According to a landmark benchmark study evaluating 14 methods across 10 datasets, Harmony, LIGER, and Seurat 3 demonstrated consistently strong performance across multiple scenarios, with Harmony offering particularly fast runtime [58] [60]. However, method performance can be context-dependent, with certain approaches excelling in specific scenarios such as identical cell types across technologies, non-identical cell types, or very large datasets [60].

For RNA velocity applications specifically, specialized methods like LatentVelo and spVelo have emerged that incorporate batch correction directly into the velocity inference framework. LatentVelo uses neural ordinary differential equations (neural ODEs) on embedded latent space while performing batch effect correction, while spVelo extends this to spatial transcriptomics data using a combination of variational autoencoders and graph attention networks [38].

Experimental Protocol for Multi-Study Integration

Preprocessing and Quality Control

Step 1: Data Acquisition and Format Standardization

• Obtain raw gene expression matrices from each study, ensuring consistent gene identifiers across datasets
• For RNA velocity analysis, simultaneously acquire both spliced and unspliced count matrices for each dataset
• Document key metadata including sequencing platform, library preparation protocol, sample processing date, and other potential batch variables [54]

Step 2: Independent Quality Control and Filtering

• Process each dataset separately through standard quality control pipelines
• Filter cells based on quality metrics: total counts, detected features, and mitochondrial percentage
• Remove low-abundance genes that appear in fewer than a threshold number of cells (typically 10)
• For velocity analysis, ensure proper quantification of unspliced counts using dedicated tools (e.g., Velocyto, kallisto-bustools) [8]

Step 3: Normalization and Feature Selection

• Apply dataset-specific normalization (e.g., library size normalization) without batch correction
• Identify highly variable genes within each dataset, then take the union for downstream integration
• Retain velocity-relevant genes that show dynamic patterns in spliced-unspliced distributions

Hierarchical Integration with scMerge2

For large-scale integration projects involving multiple studies with complex experimental designs, we recommend the scMerge2 framework, which has demonstrated effective performance on atlas-scale data encompassing millions of cells [59].

Step 1: Hierarchical Study Organization

• Organize studies hierarchically based on biological and technical relationships
• Group studies with similar experimental conditions, tissue types, or technological platforms
• Define the hierarchy to capture both local (within-group) and global (between-group) variations

Step 2: Pseudo-bulk Construction

• For each biological sample within studies, construct pseudo-bulk profiles by aggregating counts across cells of the same type
• This dramatically reduces computational complexity while maintaining biological signal
• The pseudo-bulk approach also facilitates downstream differential expression analysis [59] [63]

Step 3: Pseudo-replicate Creation

• Within each condition, create pseudo-replicates to strengthen the identification of stable genes
• These pseudo-replicates enable the algorithm to distinguish biological signals from technical noise

Step 4: Sequential Integration

• Perform batch correction sequentially through the defined hierarchy
• First correct within-study batches, then across studies within groups, and finally across group divisions
• This sequential approach better captures the multi-level nature of batch effects in complex datasets [59]

Step 5: Validation and Assessment

• Evaluate integration quality using metrics such as kBET, LISI, and ASW [60]
• Verify that biological conditions separate appropriately while batch effects are minimized
• Confirm that rare cell populations are preserved in the integrated dataset

RNA Velocity Preservation in Integrated Data

Step 1: Velocity Model Training

• For methods that support integrated velocity analysis (e.g., LatentVelo, spVelo), train velocity models directly on the integrated dataset
• Alternatively, perform velocity analysis on individual datasets prior to integration, then align velocity vectors

Step 2: Cross-Validation of Velocity Patterns

• Verify that consistent velocity patterns emerge in equivalent cell types across different batches
• Use confidence scores, transition scores, and direction scores to quantitatively assess velocity quality [38]
• Ensure that velocity vectors point toward biologically plausible successor states in the integrated manifold

Step 3: Trajectory Inference Validation

• Reconstruct developmental trajectories using the integrated data and velocity vectors
• Confirm that trajectory topology aligns with known biology
• Validate key branching points using independent markers or functional assays

Integration Workflow Visualization

Figure 1: Comprehensive workflow for multi-study integration with RNA velocity analysis, showing key stages from data preparation through final visualization.

Table 2: Key Computational Tools for Batch Effect Correction in Single-Cell Studies

Tool	Primary Function	Application Context	Key Features	Reference
Harmony	Batch effect correction	General scRNA-seq integration	Fast iterative clustering, good scalability	[58] [60]
scMerge2	Multi-study integration	Atlas-scale multi-condition data	Hierarchical integration, pseudo-bulk processing	[59]
Seurat 3	Multi-modal integration	General scRNA-seq analysis	CCA anchor-based integration, returns adjusted matrix	[60] [54]
scVI	Deep learning integration	Large-scale complex batches	Probabilistic modeling, handles sparse data	[60]
Velocyto	RNA velocity estimation	Spliced/unspliced quantification	Steady-state model, foundational velocity tool	[8]
scVelo	RNA velocity inference	Dynamic transcriptional modeling	Dynamical model, latent time estimation	[8]
spVelo	Spatial RNA velocity	Multi-batch spatial transcriptomics	Incorporates spatial information, batch correction	[38]
scCorrector	Multi-study integration	Cross-technology, cross-species	Study-specific adaptive normalization	[61]

Application to Cancer Dynamics Research

Special Considerations for Tumor Ecosystems

The application of multi-study integration methods in cancer research presents unique challenges due to the extreme heterogeneity characteristic of tumor ecosystems. Unlike normal tissues, cancer samples typically contain multiple coexisting subclones with distinct genetic and transcriptional profiles, alongside diverse non-malignant cell types in the tumor microenvironment [54]. This complexity necessitates careful consideration during integration to avoid misinterpreting genuine biological heterogeneity as batch effects.

When integrating tumor datasets across studies, several strategies can enhance biological fidelity:

Preservation of Rare Populations: Cancer stem cells, circulating tumor cells, and other rare populations often drive therapeutic resistance and metastasis. Integration methods must preserve these biologically critical but numerically minor populations. Methods like Harmony and scMerge2 have demonstrated good performance in maintaining rare cell types during integration [59] [60].

Subclonal Architecture Maintenance: Genetic subclones within tumors may exhibit subtle transcriptional differences that reflect evolutionary trajectories. Over-correction can obscure these important signatures. Using methods that explicitly model both shared and dataset-specific factors (e.g., LIGER) can help maintain these distinctions [54].

Microenvironmental Context Conservation: The tumor microenvironment contains complex mixtures of immune, stromal, and vascular cells interacting with malignant cells. Integration should maintain these contextual relationships while removing technical artifacts. Spatial transcriptomics methods like spVelo are particularly valuable for preserving spatial relationships when available [38].

RNA Velocity in Integrated Cancer Datasets

RNA velocity analysis applied to integrated cancer datasets enables powerful insights into dynamic processes such as drug resistance development, metastatic progression, and cell state plasticity. However, special considerations are necessary when combining velocity analysis with batch correction:

Latent Time Alignment: When integrating datasets from different patients or conditions, ensure that latent time estimates are comparable across batches. Methods like LatentVelo and spVelo that incorporate batch correction directly into velocity inference facilitate this alignment [38].

Transition Consistency: Verify that velocity vectors predict consistent state transitions across different datasets. For example, drug-sensitive to drug-resistant transitions should point in similar directions regardless of study origin.

Driver Gene Identification: Use integrated velocity analysis to identify conserved driver genes of cancer progression across multiple patients or studies. The consistency of these findings across batches provides validation of their biological significance.

Validation and Quality Assessment

Rigorous validation is essential to ensure that integration methods have successfully removed technical artifacts while preserving biological signals. We recommend a multi-faceted validation approach:

Quantitative Metrics: Calculate established integration quality metrics including:

• kBET (k-nearest neighbor batch-effect test): Measures local batch mixing [60]
• LISI (Local Inverse Simpson's Index): Quantifies diversity of batches in local neighborhoods [60]
• ASW (Average Silhouette Width): Assesses cell type separation and batch mixing balance [60]
• Velocity Confidence: Evaluates reliability of inferred velocities in integrated space [38]

Biological Plausibility Assessment:

• Verify that known cell type markers consistently co-express in integrated data
• Confirm that established developmental trajectories are recovered
• Ensure that patient-specific signatures are maintained where biologically justified

Functional Validation:

• Correlate computational findings with orthogonal data (e.g., protein expression, functional assays)
• Validate predicted trajectories using known temporal relationships (e.g., time-series data)
• Confirm that differentially expressed genes identified in integrated analyses recapitulate known biology

Effective multi-sample and multi-study integration is no longer optional but essential for robust single-cell cancer dynamics research. As the field moves toward increasingly collaborative and atlas-scale studies, the strategies outlined here provide a framework for overcoming batch effects while preserving the biological heterogeneity that underlies cancer progression and treatment response. By selecting appropriate integration methods, implementing careful validation protocols, and applying specialized approaches for RNA velocity analysis, researchers can extract meaningful insights from complex integrated datasets that would be impossible from individual studies alone.

The rapid development of new computational methods continues to enhance our ability to integrate diverse single-cell datasets while maintaining the integrity of dynamic analyses like RNA velocity. Future directions likely include more sophisticated deep learning approaches, enhanced multi-omic integration capabilities, and improved scalability to accommodate the ever-growing volume of single-cell data being generated by the research community.

Single-cell RNA sequencing (scRNA-seq) has revolutionized oncology research by enabling the unprecedented dissection of cellular heterogeneity, tumor microenvironments, and cancer progression dynamics. However, the rapidly expanding ecosystem of computational tools for analyzing these complex datasets presents a significant challenge for researchers. Selecting an inappropriate model can lead to misinterpretation of cellular dynamics, inaccurate trajectory inferences, and ultimately, flawed biological conclusions. This guide provides a structured framework for selecting the most appropriate computational tools based on specific cancer research questions, with a special emphasis on investigating cellular dynamics through RNA velocity analysis. We integrate current advances in machine learning (ML) and artificial intelligence (AI) to offer a comprehensive decision-making resource for scientists and drug development professionals engaged in single-cell cancer research.

Computational Tool Categories and Their Applications

RNA Velocity Models for Inferring Cellular Dynamics

RNA velocity analysis models the temporal relationship between unspliced and spliced mRNAs to predict future transcriptional states and uncover the directionality of cellular transitions in cancer processes such as metastasis, drug resistance emergence, and cell fate decisions [1] [8]. The table below categorizes primary RNA velocity approaches and their applications in cancer research.

Table 1: Categories of RNA Velocity Models and Their Applications

Category	Key Methods	Underlying Principle	Oncology Applications	Limitations
Steady-State Methods	Velocyto, scVelo (deterministic/stochastic), TopicVelo [8]	Assumes constant splicing rates and transcriptional equilibrium; uses least-squares regression on steady-state subpopulations [8]	- Identifying clear differentiation trajectories in tumor cell subtypes- Mapping steady-state cellular populations in established tumor regions	- Assumptions often violated in highly heterogeneous tumors- Inaccurate for complex, non-steady-state kinetics like epithelial-mesenchymal transition [8]
Trajectory Methods	scVelo (dynamical), UniTVelo, Dynamo, veloVI, Pyro-Velocity [8]	Estimates kinetic parameters to construct phase portrait trajectories aligned with latent cell time [8]	- Reconstructing branching lineage decisions in cancer stem cells- Modeling drug-induced cellular state transitions with precise latent time inference	- Computationally intensive for very large datasets (>100,000 cells)- Requires careful parameter tuning for complex trajectory topologies [8]
State Extrapolation Methods	VeloAE, Cell2fate [8]	Leverages expected future cell states to guide estimation and optimization of cell-level RNA velocity vectors [8]	- Predicting metastatic progression paths- Forecasting emergence of therapy-resistant subclones from snapshot data	- Higher computational complexity- May require integration with additional multi-omics data for optimal performance [8]

Machine Learning and AI Models for Cancer Prediction Tasks

Beyond velocity analysis, ML and AI models offer powerful approaches for cancer diagnosis, survival prediction, and treatment response forecasting. These tools excel at identifying complex patterns in high-dimensional data that may elude traditional statistical methods.

Table 2: Machine Learning Models for Cancer Research Applications

Model Category	Specific Methods	Research Application	Performance Notes	Considerations
Survival Prediction	Random Survival Forests, LASSO, Cox Proportional Hazards [64]	Predicting colon cancer survival outcomes based on clinical and molecular features [64]	Random survival forests and LASSO outperformed traditional Cox models (C-index: 0.8146) [64]	Identified key predictors: positive lymph nodes, treatment type, age, smoking status, geographic region [64]
Diagnostic & Subtyping	AEON, Paladin, SuperLearner, Logistic Regression [65] [66]	Cancer subtype classification from H&E images; pancreatic cancer diagnosis [65] [66]	AEON: 78% accuracy in cancer subtype classification; SuperLearner: highest precision (66.67%) [65] [66]	SuperLearner struggled with sensitivity; Logistic Regression offered better interpretability [66]
Treatment Response	AI-powered clinical decision support systems [65] [67]	Predicting immunotherapy response; identifying patients for targeted therapies (e.g., PARP inhibitors) [65] [67]	DeepHRD: 3x more accurate in detecting HRD-positive cancers vs. genomic tests [67]	MSI-SEER identifies microsatellite instability-high regions often missed by traditional testing [67]

Experimental Design and Workflow Integration

ScRNA-Seq Wet-Lab Protocols and Reagent Solutions

The computational analysis pipeline begins with appropriate experimental design and sample preparation. The selection of scRNA-seq protocols directly influences downstream analytical possibilities and should align with research objectives.

Table 3: Essential Research Reagent Solutions for scRNA-Seq Experiments

Reagent/Kit	Primary Function	Key Features	Application Context
10X Genomics 3' Gene Expression [68]	3' end counting-based scRNA-seq	- PolyA-based mRNA capture- Cell barcoding and UMIs- Feature barcoding for surface proteins	Standard "workhorse" for tumor heterogeneity studies; requires fresh cells [68]
10X Genomics 5' Gene Expression/Immune Profiling [68]	5' end counting with immune repertoire	- TSO-based capture- V(D)J sequencing compatibility- CRISPR screening compatibility	Ideal for tumor immunology studies and TIL characterization [68]
10X Genomics Single Nucleus Multiome [68]	Parallel ATAC-seq and gene expression	- Simultaneous chromatin accessibility and transcriptome profiling- Same nucleus analysis	Epigenetic regulation in cancer; frozen samples; complex tissues resistant to dissociation [68]
Unique Molecular Identifiers (UMIs) [69] [68]	Quantitative transcript counting	- Labels individual mRNA molecules- Corrects for PCR amplification biases	Essential for accurate velocity analysis and differential expression in heterogeneous tumors [69] [68]
Sample Preparation Buffers [68]	Cell suspension and viability	- PBS with 0.04% BSA recommended- Low EDTA concentrations (<0.1mM)	Critical for maintaining cell integrity and reverse transcription efficiency [68]

RNA Velocity Computational Workflow

The computational pipeline for RNA velocity requires specific preprocessing steps to ensure accurate kinetic parameter estimation. The workflow extends beyond standard scRNA-seq analysis to incorporate splicing dynamics.

Model Selection Framework for Cancer Research Questions

Decision Matrix for Common Oncology Research Scenarios

Selecting the optimal computational approach requires matching the model's strengths to specific biological questions. The following decision matrix provides guidance for common scenarios in cancer research.

Table 4: Model Selection Guide Based on Research Objectives

Research Objective	Recommended RNA Velocity Approach	Complementary ML/AI Tools	Expected Output	Validation Strategy
Identifying cellular plasticity and EMT in carcinoma [8]	Trajectory methods (scVelo dynamical, UniTVelo)	Random forests for feature importance; DeepHRD for HRD detection [67] [8]	Branching trajectory plot showing EMT progression; latent time ordering	Immunofluorescence for mesenchymal markers; in vitro invasion assays
Tracking cancer stem cell differentiation hierarchies [1] [8]	State extrapolation methods (VeloAE, Cell2fate)	AEON for histologic subtype classification [65]	Predicted future states; stem cell probability curves	FACS sorting with stem cell markers; limiting dilution transplantation assays
Predicting therapy resistance development [64] [67]	Bayesian velocity models (Pyro-Velocity, VeloVAE)	Survival ML models (Random Survival Forests) [64]	Posterior uncertainty in velocity estimates; survival probability curves	Longitudinal sampling; drug sensitivity assays in patient-derived organoids
Characterizing tumor immune microenvironment dynamics [1] [67]	Steady-state methods (Velocyto) for stable populations; Trajectory methods for activation	AI-based immune cell deconvolution from H&E [65]	Immune cell state transitions; cell-cell communication inference	Multiplex IHC; CITE-seq validation; T cell receptor sequencing
Mapping metastatic progression pathways [8]	Multi-omics integration methods (MultiVelo)	Paladin for genotype-phenotype relationships [65]	Spatial velocity projections; driver gene identification	Circulating tumor cell analysis; spatial transcriptomics validation

Practical Implementation Protocols

Protocol 1: RNA Velocity Analysis for Branching Lineage Tracing

Application: Mapping differentiation hierarchies in acute myeloid leukemia or tumor cell states with branching plasticity.

Step-by-Step Workflow:

• Data Acquisition: Generate scRNA-seq data using 10X Genomics 3' or 5' kits, ensuring capture of both spliced and unspliced counts. Target 10,000-20,000 cells for optimal resolution [68].
• Preprocessing: Use CellRanger or kallisto|bustools for read alignment and quantification of spliced/unspliced matrices. Filter low-quality cells (mitochondrial gene percentage <20%) [69] [8].
• Normalization and Imputation: Apply scVelo's preprocessing pipeline including library size normalization and k-nearest neighbor smoothing (k=30) to reduce technical noise [8].
• Model Fitting: Implement scVelo's dynamical model using the tl.recover_dynamics() function. Allow sufficient iterations (max_iter=20) for convergence [8].
• Velocity Estimation and Visualization: Compute velocities using tl.velocity() and project onto UMAP embedding with tl.velocity_graph(). Use pl.velocity_embedding_stream() for visualization [8].
• Interpretation: Identify driver genes of transitions using tl.velocity_confidence() and tl.differential_kinetic_test() to find genes with significant kinetic changes between branches [8].

Troubleshooting Tips:

• If velocities appear random or uninformative, check the proportion of unspliced counts (should be 5-20% of total reads).
• For unstable dynamical model fitting, try the stochastic or steady-state models as alternatives.
• When velocities contradict known biology, validate with pseudotime methods (Slingshot, Monocle3) as orthogonal approaches.

Protocol 2: ML-Enhanced Survival Prediction for Precision Oncology

Application: Predicting patient survival and treatment response from clinical and molecular features.

Step-by-Step Workflow:

• Data Compilation: Integrate clinical data (age, stage, treatment), molecular features (mutation status, gene expression), and outcome data (overall survival, progression-free survival) [64].
• Feature Preprocessing: Handle missing data via multiple imputation. Normalize continuous variables and encode categorical variables. Split data into training (70%) and validation (30%) sets [64].
• Model Training: Implement multiple algorithms in parallel:
  • Random Survival Forests with 1000 trees
  • LASSO Cox regression with 10-fold cross-validation
  • Traditional Cox proportional hazards model as baseline
  • Gradient boosting models (XGBoost) for non-linear relationships [64]
• Model Evaluation: Assess using concordance index (C-index), Brier score, and calibration plots. Perform internal validation via bootstrapping (500 iterations) [64].
• Clinical Implementation: Deploy best-performing model with risk stratification thresholds (e.g., low, intermediate, high risk). Integrate into clinical decision support systems with interpretability visualizations [65] [64].

Interpretation Guidelines:

• Prioritize models with C-index >0.75 for clinical consideration [64].
• Ensure key clinical predictors (e.g., stage, treatment) maintain significance in ML models.
• Use SHAP plots to explain individual predictions and maintain model interpretability for clinical use.

The expanding toolkit for single-cell cancer research offers unprecedented opportunities to unravel tumor complexity, but requires thoughtful selection and implementation. RNA velocity methods provide unique insights into dynamic processes, while complementary ML approaches enable robust prediction and classification. As the field advances, integration of multi-omics data, improved scalability for massive datasets, and enhanced model interpretability will further strengthen these computational approaches. By aligning research questions with appropriate computational models through the framework presented here, researchers can maximize biological insights while maintaining methodological rigor in their oncology studies.

In single-cell cancer dynamics research, RNA velocity analysis has emerged as a transformative tool for predicting cellular trajectories and fate decisions from snapshot single-cell RNA sequencing (scRNA-seq) data. This method leverages the ratio of unspliced (nascent) to spliced (mature) messenger RNA to infer instantaneous gene expression change rates and predict future transcriptional states [1]. The reliability of these kinetic inferences—characterized by parameters capturing transcription (α), splicing (β), and degradation (γ) rates—is fundamentally constrained by initial data quality and appropriate preprocessing strategies. In cancer studies, where cellular heterogeneity and dynamic state transitions are paramount, rigorous data validation becomes indispensable for distinguishing genuine biological signals from technical artifacts, ultimately enabling accurate reconstruction of tumor evolution, drug resistance emergence, and metastatic pathways.

Foundational Principles of Kinetic Parameter Estimation

Mathematical Framework of RNA Velocity

RNA velocity models are built upon a system of ordinary differential equations (ODEs) that describe the central dogma of molecular biology for each gene:

$$ \begin{aligned} \frac{dug}{dt} &= \alphag(t) - \betag ug \ \frac{dsg}{dt} &= \betag ug - \gammag s_g \end{aligned} $$

where (ug) and (sg) represent the abundance of unspliced and spliced RNA for gene (g), respectively [33]. The parameters (αg(t)), (βg), and (γg) denote the transcription rate, splicing rate, and degradation rate, respectively. The velocity of gene (g) is then defined as the time derivative of its spliced count, (dsg/dt). In cancer research, estimating these parameters accurately allows researchers to determine whether a gene is being upregulated or downregulated within individual tumor cells, providing critical insights into the molecular drivers of cancer progression and cellular plasticity.

Technical Challenges in Parameter Estimation

Estimating these kinetic parameters presents significant computational and biological challenges, especially in the context of cancer datasets. Technical noise in scRNA-seq protocols, sparse expression matrices characteristic of tumor microenvironments, and complex transcriptional dynamics that deviate from simple ODE assumptions can severely compromise parameter estimation [5]. Furthermore, cancer cells often exhibit multi-rate kinetics, where genes display coordinated changes in transcription rates across cellular trajectories, as observed in erythroid maturation and likely in tumor cell state transitions [33]. Recent methodological advances, including Cell2fate, TSvelo, and TIVelo, have introduced more sophisticated frameworks to address these limitations through Bayesian inference, neural ODEs, and cluster-level trajectory integration, yet all remain dependent on high-quality input data [5] [33] [12].

Experimental Design and Sample Preparation Protocols

Strategic Selection of Single-Cell RNA Sequencing Platforms

The foundation for reliable kinetic parameter estimation begins with appropriate experimental design and platform selection. Different scRNA-seq platforms offer varying advantages for RNA velocity analysis, particularly in cancer research where sample availability and tumor heterogeneity present unique challenges. The table below summarizes key commercial solutions and their characteristics relevant to preprocessing for RNA velocity:

Table 1: Commercial scRNA-seq Platform Comparison for RNA Velocity Analysis

Commercial Solution	Capture Platform	Throughput (Cells/Run)	Capture Efficiency (%)	Fixed Cell Support	Considerations for RNA Velocity
10X Genomics Chromium	Microfluidic oil partitioning	500–20,000	70–95	Yes	High capture efficiency beneficial for sparse cancer transcripts
BD Rhapsody	Microwell partitioning	100–20,000	50–80	Yes	Larger cell size capacity useful for rare tumor cells
Parse Evercode	Multiwell-plate	1,000–1M	>90	Yes	Lowest cost per cell for large tumor atlases
Fluent/PIPseq (Illumina)	Vortex-based oil partitioning	1,000–1M	>85	Yes	No size restrictions for complex cancer morphology

Platform choice significantly impacts downstream velocity analysis, with capture efficiency directly influencing the detection of transient unspliced transcripts essential for kinetic parameter estimation [70]. For cancer studies involving precious clinical samples or requiring integration with other assays, support for fixed cells enables preservation of sample material while still permitting RNA velocity analysis, though with potential compromises in sensitivity for low-abundance transcripts.

Sample Preparation and Quality Assessment

Proper sample preparation is paramount for generating high-quality scRNA-seq data suitable for RNA velocity analysis. The following protocol outlines critical steps for sample processing:

• Cell Suspension Preparation: Generate high-quality single-cell or single-nuclei suspensions from tumor tissue using optimized dissociation protocols. For tissue with extensive extracellular matrix (common in desmoplastic tumors), consider combinatorial enzymatic approaches tailored to the specific cancer type [70].

• Viability and Debris Management: Assess cell viability using automated cell counters and implement fluorescence-activated cell sorting (FACS) with live/dead stains to eliminate debris while minimizing artifacts related to cell stress. Ideal samples should exhibit >90% viability with minimal aggregation [68].

• Inhibition of Stress Responses: Perform dissociations on ice to mitigate transcriptomic stress responses, though this may prolong digestion times as most commercial enzymes are optimized for 37°C activity. Alternatively, consider fixation-based methods such as methanol maceration (ACME) or reversible dithio-bis(succinimidyl propionate) (DSP) fixation immediately following cell dissociation to preserve transcriptional states [70].

• Buffer Compatibility: Ensure samples are delivered in buffer free of components that inhibit reverse transcription reactions (e.g., EDTA at concentrations above 0.1 mM). 10X Genomics recommends PBS with 0.04% BSA as an optimal suspension buffer [68].

• Concentration Optimization: Target ideal cell concentrations of 1,000–1,600 cells/μL with a minimum of 100,000–150,000 total cells to ensure sufficient capture for assessing tumor heterogeneity while maintaining sequencing depth requirements for velocity analysis [68].

The following workflow diagram illustrates the critical decision points in sample preparation and their impact on downstream RNA velocity analysis:

Sample Preparation Workflow for RNA Velocity

Computational Preprocessing and Normalization

Quality Control Metrics and Thresholding

Robust preprocessing begins with stringent quality control (QC) to identify and remove low-quality cells that would otherwise distort kinetic parameter estimation. The following table outlines key QC metrics and recommended thresholds for RNA velocity analysis in cancer datasets:

Table 2: Quality Control Metrics for RNA Velocity Preprocessing

QC Metric	Recommended Threshold	Rationale	Cancer-Specific Considerations
UMIs per Cell	>500 (nuclei) >1,000 (cells)	Ensures sufficient detection of both spliced/unspliced counts	Tumor cells may have elevated RNA content
Genes per Cell	>250 (nuclei) >500 (cells)	Maintains transcriptomic complexity	Heterogeneous in tumor microenvironments
Mitochondrial Read Percentage	<10-20%	Identifies stressed/dying cells	Varies by cancer type and metabolic state
Unspliced mRNA Proportion	5-30%	Validates splice-aware alignment	May be altered in cancers with splicing defects
Doublet Detection	Technology-specific thresholds	Removes multiplets that confound dynamics	Critical in hypercellular tumor samples

Implementation of these QC metrics should be performed using standard tools such as Seurat or Scanpy, with careful consideration of cancer-specific biology. For instance, certain tumor types may naturally exhibit elevated mitochondrial content or altered RNA processing that requires adjustment of standard thresholds [71].

Normalization and Batch Effect Correction

Following quality control, appropriate normalization is essential for accurate kinetic parameter estimation:

• Library Size Normalization: Apply depth-based normalization (e.g., log(CP10K)) to account for varying sequencing depth across cells, while being mindful that this may mask true biological heterogeneity in cancer cells with aberrant transcriptional activity.

• Batch Effect Correction: When integrating multiple samples or datasets (common in cancer studies spanning different patients or time points), employ batch correction methods such as Harmony, Seurat's integration, or specialized tools like scGen to mitigate technical variation while preserving biological signals [38]. For spatial transcriptomics data integrated with RNA velocity, recent methods like spVelo incorporate Graph Attention Networks (GATs) with Maximum Mean Discrepancy (MMD) penalties to correct batch effects while maintaining spatial relationships [38].

• Gene Filtering: Filter uninformative genes based on their contributions to cell development, as demonstrated in spVelo, which removes genes less enriched for relevant biological pathways, thereby reducing noise in velocity estimation [38].

Validation Frameworks for Kinetic Parameter Estimation

Consistency Metrics for Velocity Validation

Validating the reliability of estimated kinetic parameters requires multiple complementary approaches:

• Velocity Confidence: Measures the reliability of inferred velocities by assessing consistency within local neighborhoods in transcriptional space [38].

• Transition Score: Evaluates the probability of true cell-to-cell transitions by comparing predicted future states with actual observed transcriptomic changes along differentiation trajectories [38].

• Cross-Boundary Directional Correctness (CBDir): Scores the consistency of transition probabilities at the boundary between cell clusters with known transitions, providing a metric aligned with biological ground truth [33].

• Direction Score: Assesses the coherence between predicted cell movement direction and observed cell displacement in principal component analysis (PCA) space, particularly important for spatial transcriptomics data [38].

Benchmarking Against Biological Ground Truths

Beyond computational metrics, validation should incorporate biological ground truths where possible:

• Pseudotime Alignment: Compare velocity-inferred temporal ordering with known developmental timelines or drug treatment time courses. In cancer studies, this might involve benchmarking against established tumor progression markers or sequential biopsy samples.

• Spatial Validation: For datasets with spatial transcriptomics, verify that velocity vectors align with known spatial organization patterns, such as gradient expression patterns across tumor boundaries or immune cell infiltration fronts [38].

• Experimental Validation: Where feasible, correlate velocity predictions with functional assays, such as lineage tracing or single-cell qPCR measurements of key regulatory genes identified through the analysis.

Advanced Integrative Approaches for Cancer Research

Multi-Omics Integration for Enhanced Kinetic Models

Recent methodological advances have expanded RNA velocity beyond standard spliced/unspliced count models to incorporate additional molecular layers particularly relevant to cancer biology:

• Multiome ATAC + Gene Expression: Simultaneous measurement of chromatin accessibility and gene expression in the same nucleus enables more accurate modeling of transcription rates by incorporating regulatory information [68]. Tools like MultiVelo extend RNA velocity to integrate scATAC-seq data, providing insights into how chromatin dynamics precede transcriptional changes during cancer cell state transitions [12].

• TFvelo and Regulatory Network Inference: Methods like TFvelo and TSvelo incorporate transcription factor-target relationships from databases like ChEA and ENCODE to model the regulatory cascade governing gene expression dynamics, potentially identifying key transcriptional drivers in cancer progression [12].

• Spatial Velocity Analysis: The emergence of spVelo enables RNA velocity inference in multi-batch spatial transcriptomics data, allowing researchers to connect temporal dynamics with spatial tissue organization in tumor microenvironments [38].

The Scientist's Toolkit: Essential Research Reagent Solutions

Table 3: Key Research Reagent Solutions for RNA Velocity in Cancer Dynamics

Reagent/Resource	Function	Application in Cancer Research
10X Genomics 3′ Gene Expression	Standard scRNA-seq with polyA-based mRNA capture	Workhorse for tumor cell atlas construction
10X Genomics Multiome ATAC + Gene Expression	Parallel measurement of chromatin accessibility and gene expression	Identifying regulatory drivers of tumor plasticity
Parse Evercode BioSciences	Combinatorial barcoding for high cell throughput	Large-scale tumor heterogeneity studies
BD Rhapsody	Microwell partitioning with antibody-based cell sorting	Targeted analysis of rare tumor subpopulations
Live/Dead Stains (e.g., propidium iodide)	Viability assessment during cell sorting	Eliminating debris from dissociated tumor tissue
RNase Inhibitors	Preservation of RNA integrity during processing	Maintaining quality in prolonged sample processing
DSP (dithio-bis(succinimidyl propionate))	Reversible crosslinker for fixation	Preserving transcriptional states in archival samples

Reliable estimation of kinetic parameters in RNA velocity analysis requires meticulous attention to data preprocessing and validation at every stage, from experimental design through computational analysis. In cancer research, where cellular dynamics underlie critical phenotypes such as metastasis, drug resistance, and stemness, rigorous quality control appropriate normalization, and comprehensive validation are indispensable for extracting biologically meaningful insights. The integration of emerging technologies—including multi-omics measurements, spatial transcriptomics, and advanced computational frameworks like Cell2fate, TSvelo, and spVelo—promises to further enhance the accuracy and interpretability of RNA velocity in characterizing cancer dynamics. By adhering to the protocols and validation standards outlined herein, researchers can ensure their velocity analyses provide trustworthy insights into the temporal dynamics driving cancer progression and treatment response.

RNA velocity analysis has emerged as a powerful extension of trajectory inference for single-cell RNA sequencing (scRNA-seq) data, offering the potential to predict the future transcriptional states of individual cells and uncover dynamic processes in cancer progression, treatment response, and resistance development. The core premise of RNA velocity leverages the ratio of unprocessed (unspliced) to processed (spliced) messenger RNA molecules to infer the instantaneous rate of gene expression change, thereby predicting cellular directionality and state transitions [4] [2]. In cancer research, this methodology promises to reveal transitions between drug-sensitive and resistant states, tumor cell plasticity, and differentiation hierarchies within heterogeneous tumors.

However, the application of RNA velocity to cancer dynamics presents significant interpretative challenges. Recent studies have highlighted critical limitations where RNA velocity can produce misleading or incorrect trajectories due to technical artifacts, model misspecification, and biological complexities inherent to cancer systems [4] [2]. A 2023 study revealed that RNA velocity estimates exhibit "considerable estimation errors for both direction and speed" when the underlying k-nearest neighbors (k-NN) graph fails to accurately represent true data structure [4]. This is particularly problematic in cancer datasets, which often exhibit extreme heterogeneity, multiple branching points, and non-linear dynamics that violate core assumptions of standard RNA velocity workflows.

Critical Assessment of RNA Velocity Limitations

Fundamental Technical and Mathematical Challenges

The mathematical foundations of RNA velocity contain several vulnerabilities that can generate misleading trajectories in cancer research applications:

• Scale Invariance Problem: Current models cannot distinguish between systems with different speeds of dynamics, as the same velocity vector field can be rescaled arbitrarily while maintaining similar structure [2]. This poses significant challenges for comparing dynamics across different cancer subtypes or treatment conditions.

• K-NN Graph Dependency: Both velocity estimation and visualization heavily rely on the k-NN graph constructed from spliced counts. When this graph inaccurately represents biological relationships due to batch effects, technical noise, or complex biology, the resulting velocity estimates become error-prone [4].

• Indeterminate Speed Estimation: Except in very low-noise settings, RNA velocity performs poorly at estimating the actual speed of cellular transitions, limiting its quantitative application for predicting the timing of cancer phenotypic transitions [4].

• Model Misspecification: The standard models assume constant transcription, splicing, and degradation rates across cells, an assumption frequently violated in cancer due to heterogeneous microenvironments and genomic instability [2].

Methodological Variability and Implementation Artifacts

Different computational approaches for estimating RNA velocity can yield divergent, sometimes contradictory results on the same cancer dataset:

• Algorithmic Discrepancies: Significant qualitative differences have been observed between outputs of popular implementations like velocyto and scVelo, with one analysis noting they can "suggest totally different causal relationships between cell types" [2].

• Circularity in Visualization: The mapping of high-dimensional velocities to low-dimensional embeddings creates potential for circular reasoning, as the "use of RNA velocity in assessing the correctness of a low-dimensional embedding is circular" [4].

• Hyperparameter Sensitivity: The RNA velocity workflow contains numerous arbitrary user-set parameters (k-NN graph construction, smoothing parameters, embedding choices) that substantially impact results yet lack biological justification [2].

• Batch Effect Vulnerability: Standard RNA velocity methods cannot directly correct for batch effects across multiple experiments because they process spliced and unspliced matrices with a proportional relationship that is disrupted by conventional batch correction techniques [6].

Table 1: Common Artifacts in RNA Velocity Analysis of Cancer Datasets

Artifact Type	Causes	Impact on Cancer Trajectories
Incorrect Directionality	Poor k-NN graph construction, model misspecification	Reversed differentiation trajectories, misidentified cell fate
Spurious Branching Points	Technical noise, over-smoothing	False prediction of cancer cell lineage bifurcations
Speed Distortion	High noise, incorrect kinetic parameter estimation	Misestimation of transition rates between cancer states
Batch-induced Streams	Technical variation between samples	Artificial trajectories aligning with batch rather than biology
Embedding-dependent Patterns	Circular visualization practices	Topology-driven rather than biology-driven trajectories

Experimental Framework for Robust Velocity Analysis

Quality Control and Validation Metrics

Implementing rigorous quality control is essential before interpreting RNA velocity results in cancer datasets:

• Velocity Consistency Score: Develop a novel quality measure that quantifies the local consistency between velocity vectors and cell neighborhood structure. Low scores indicate when "RNA velocity should not be used" due to unreliable estimates [4].

• Gene-level Validation: Assess velocity fits for individual genes using residual analysis, focusing on key cancer drivers and markers to identify potentially misleading kinetic profiles.

• Pseudotemporal Ordering Concordance: Compare velocity-based ordering with alternative pseudotime methods based on transcriptomic similarity to identify discordant regions requiring further investigation.

• Batch Effect Quantification: Implement negative control analyses to distinguish biologically meaningful trajectories from batch-associated patterns using methods like VeloVGI that specifically address multi-batch challenges [6].

Computational Protocols for Reliable Analysis

Protocol 1: k-NN Graph Construction and Validation

• Input: Spliced count matrix (cells × genes), batch metadata, quality control metrics
• Procedure:
  • Perform preliminary PCA on spliced counts
  • Construct multiple k-NN graphs with varying k values (k=15, 30, 50)
  • For multi-batch datasets, employ mutual nearest neighbor (MNN) algorithms with optimal transport to establish inter-batch relationships [6]
  • Assess graph connectivity within and between biological conditions
  • Validate neighborhood preservation using trustworthiness metrics
• Output: Validated k-NN graph with optimal parameters for downstream analysis
• Troubleshooting: If batch effects persist, consider VeloVGI's graph structure fine-tuning that "employs optimal transport and mutual nearest neighbor approach to construct neighbors in batch data" [6]

Protocol 2: Velocity Estimation with Model Selection

• Input: Spliced and unspliced count matrices, validated k-NN graph
• Procedure:
  • Compare steady-state and dynamical models for key marker genes
  • Assess model fits using likelihood-based criteria and residual patterns
  • Employ dynamical model when sufficient cells capture transient states
  • Utilize steady-state model for more mature, stable cancer populations
  • Incorporate metabolic labeling data where available using tools like Dynamo [31]
• Output: Validated velocity estimates with quality metrics
• Troubleshooting: If velocity patterns contradict known biology, inspect individual gene fits and consider alternative models

Protocol 3: Vector Field Reconstruction and Trajectory Inference

• Input: High-dimensional velocity estimates, low-dimensional embedding
• Procedure:
  • Reconstruct continuous vector fields using tools like Dynamo that "reconstruct continuous vector fields from discrete velocity vectors" [31]
  • Employ VeTra or CellRank for trajectory inference based on cosine similarity and weakly connected components [72]
  • Identify initial, intermediate, and terminal states using transition probabilities
  • Perform in silico perturbations to test trajectory robustness [31]
• Output: Annotated trajectories with confidence estimates
• Troubleshooting: Validate trajectories using known marker progression and pseudotime consistency

Figure 1: RNA Velocity Analysis Workflow with Quality Checkpoints

Advanced Analytical Solutions

Addressing Batch Effects in Multi-Sample Cancer Studies

Cancer studies typically involve multiple patients, treatment conditions, and time points, creating severe batch effect challenges. VeloVGI provides a specialized framework for multi-batch RNA velocity analysis through:

• Integrated Graph Construction: Combining separate inter-batch and intra-batch relationships to form innovative multi-batch networks that preserve biological signals while mitigating technical variation [6].

• Variational Graph Autoencoder: Employing a VGAE based on fine-tuned graph structure to estimate RNA velocity across batches, incorporating "graph structure into the encoder for more effective feature extraction" [6].

• Sampling and Aggregation Strategies: Using inductive minibatch approaches like GraphSAGE during model training to reduce computational overhead while maintaining accuracy.

Table 2: Comparison of RNA Velocity Methods with Batch Correction Capabilities

Method	Batch Handling Approach	Cancer Application Suitability	Limitations
VeloVGI	Mutual nearest neighbors + optimal transport + VGAE	High - specifically designed for complex multi-sample studies	Higher computational complexity
UniTVelo	Gene-shared cell latent time	Medium - robust to some technical variation	Limited validation in heterogeneous cancer datasets
Dynamo	Metabolic labeling integration	Medium - uses experimental controls	Requires specialized labeling data
scVelo	Standard preprocessing only	Low - severely impacted by batch effects	No dedicated batch correction framework
velocyto	Standard preprocessing only	Low - severely impacted by batch effects	No dedicated batch correction framework

Differential Geometry and In Silico Perturbation Analysis

For cancer mechanistic studies, advanced mathematical frameworks provide deeper biological insights:

• Differential Geometry Analysis: Tools like Dynamo employ "differential geometry to extract underlying regulations" of cell-fate transitions, revealing asymmetrical regulation within key cancer circuits [31].

• Least-Action Path Method: This approach "accurately predicts drivers of numerous hematopoietic transitions" and can be adapted to identify critical regulators of cancer state transitions [31].

• In Silico Perturbation: Computational prediction of "cell-fate diversions induced by gene perturbations" enables pre-testing of therapeutic interventions and identification of potential resistance mechanisms [31].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Computational Tools for RNA Velocity in Cancer Research

Tool/Resource	Function	Application Context in Cancer Research
Dynamo	Vector field reconstruction, differential geometry, in silico perturbation	Identifying master regulators of therapy resistance transitions
VeloVGI	Multi-batch RNA velocity estimation	Integrating single-cell data across patients and treatment time points
VeTra	Trajectory inference based on RNA velocity	Mapping branching points in cancer stem cell differentiation
CellRank	Terminal state identification, trajectory probability	Predicting final fates of plastic tumor cell states
scVelo	Dynamical modeling of RNA velocity	Characterizing kinetics of oncogene expression programs
STREAM	Trajectory reconstruction and mapping	Building reference trajectories for mapping new cancer samples

Figure 2: Solution Framework for Reliable Cancer Trajectory Analysis

RNA velocity analysis offers tremendous potential for unraveling cancer dynamics but requires meticulous implementation and rigorous validation to avoid misleading interpretations. Successful application in cancer research necessitates:

• Acknowledging Fundamental Limitations: Recognize that RNA velocity estimates contain inherent uncertainties, particularly regarding speed estimation and directionality in complex cancer ecosystems.

• Implementing Multi-Method Validation: Corroborate findings across different velocity methods and complementary trajectory inference approaches.

• Addressing Batch Effects Proactively: Employ specialized tools like VeloVGI when integrating data across multiple cancer samples, patients, or experimental conditions.

• Leveraging Advanced Frameworks: Utilize differential geometry and in silico perturbation analyses to move beyond descriptive trajectories toward mechanistic insights into cancer progression and treatment resistance.

• Establishing Rigorous Quality Control: Implement quantitative quality metrics at each analytical stage to identify potentially unreliable results before biological interpretation.

By adopting these cautious yet advanced analytical frameworks, cancer researchers can harness the predictive potential of RNA velocity while minimizing the risk of constructing misleading narratives about tumor evolution and cellular dynamics.

Benchmarking and Biological Validation: Ensuring Credible Cancer Insights

RNA velocity analysis has emerged as a powerful computational technique for predicting cellular dynamics, such as differentiation trajectories and state transitions, from single-cell RNA sequencing (scRNA-seq) data. Within the specific context of cancer research, accurately inferring these dynamics is paramount for understanding tumor heterogeneity, drug resistance, and metastatic progression. However, the reliability of any biological insight hinges on the rigorous assessment of the RNA velocity results themselves. This application note details the key performance metrics and experimental protocols for evaluating two critical aspects of an RNA velocity analysis: the consistency of the estimated velocity vectors and the accuracy of the inferred cellular trajectories. The focus is placed on methodologies applicable to the complex and often heterogeneous systems characteristic of cancer genomics.

Key Performance Metrics for RNA Velocity Evaluation

A robust evaluation of RNA velocity requires quantifying both the internal coherence of the velocity field and its alignment with known or biologically plausible cellular trajectories. The following metrics are essential for this task.

Table 1: Core Performance Metrics for RNA Velocity Evaluation

Metric	Definition	Interpretation in Cancer Context
Velocity Consistency	Measures the agreement between the velocity vector of a cell and the vectors of its nearest transcriptomic neighbors [12] [17].	High consistency in a tumor subpopulation suggests a coherent, directional process (e.g., a consistent epithelial-to-mesenchymal transition), while low consistency may indicate high noise or mixed states.
In-Cluster Coherence	Assesses whether velocity vectors within a pre-defined cell cluster (e.g., a cancer cell subtype) point in a similar direction [12].	Validates that a transcriptionally defined cluster is also dynamically uniform, strengthening the case for it being a distinct state in a cancer progression pathway.
Cross-Boundary Correctness	Evaluates if velocity vectors point toward the biologically correct subsequent cell state (e.g., from a progenitor to a differentiated state) [12].	Critical for verifying that inferred trajectories match known cancer progression lineages (e.g., from a stem-like state to a committed state) or for challenging proposed novel pathways.
Direction Score / Accuracy	Quantifies the agreement between the velocity-inferred direction and a ground truth reference, such as a pseudotime ordering or protein-derived scores (e.g., FUCCI) [73] [17].	Provides orthogonal validation; for example, confirming that velocity vectors align with cell-cycle progression in proliferating tumor cells.
Velocity Uncertainty	A posterior distribution over velocity estimates, provided by Bayesian methods like veloVI, which quantifies confidence in the predictions [17].	Identifies cell states in a tumor with ambiguous dynamics, highlighting regions where biological interpretation should be cautious and might require more data.

Experimental Protocols for Metric Calculation

Below are detailed protocols for calculating the two most critical metrics: Velocity Consistency and Cross-Boundary Correctness.

Protocol 1: Quantifying Velocity Consistency

Velocity consistency is a fundamental check for the reliability of the velocity field, as it relies on the assumption that transcriptomically similar cells should have similar future states [17].

Workflow Overview

Step-by-Step Procedure

• Input Preparation: Begin with the high-dimensional velocity matrix V (cells × genes), typically the output from an RNA velocity tool (e.g., scVelo, veloVI). Also required is the spliced mRNA count matrix S (cells × genes).

• k-Nearest Neighbors (k-NN) Graph Construction:

  • Normalize the S matrix (e.g., by library size and log-transform).
  • Perform dimensionality reduction on S using Principal Component Analysis (PCA), typically retaining the top 30-50 principal components.
  • Construct a k-NN graph in the PCA space using Euclidean distance. The value of k (e.g., 30) is a parameter; sensitivity analysis may be required.

• Pairwise Cosine Similarity Calculation: For each cell i in the dataset:

  • Retrieve its velocity vector V_i.
  • For every cell j in the pre-computed k-nearest neighbors of cell i, retrieve its velocity vector V_j.
  • Calculate the cosine similarity between V_i and V_j: cosine_similarity(i, j) = (V_i • V_j) / (||V_i|| * ||V_j||)
  • This results in a similarity score for each cell-neighbor pair.

• Aggregation and Scoring: The velocity consistency for a single cell i can be defined as the mean or median of the cosine similarities with its neighbors. The global Velocity Consistency metric for the entire dataset is the average of all these cell-wise consistency scores. A score close to 1 indicates highly consistent local velocity fields, while a score near or below 0 suggests random, unreliable directions.

Protocol 2: Assessing Cross-Boundary Correctness

This metric evaluates whether velocity vectors correctly predict transitions across known cell state boundaries, which is vital for validating inferred trajectories in cancer, such as a drug-sensitive to drug-resistant transition.

Workflow Overview

Step-by-Step Procedure

• Input Preparation: You will need:

  • A low-dimensional embedding of the cells (e.g., UMAP or t-SNE), which includes the projected velocity vectors.
  • Cell-type or cell-state annotations (e.g., "Stem-like," "Differentiated," "Hypoxic").
  • A biologically hypothesized trajectory, defined as a transition from a source cluster A to a target cluster B.

• Identify Boundary Cells: For the source cluster A, identify the subset of cells that are located on the boundary facing cluster B. This can be done by:

  • Calculating the centroid of cluster B in the low-dimensional embedding.
  • For each cell in A, calculating the vector from the cell to the centroid of B.
  • Selecting the top X% of cells in A with the smallest distance to cluster B, or those whose direction to B's centroid most closely aligns with the overall direction from A to B.

• Calculate Reference Direction: Compute the mean velocity vector for all cells within the target cluster B. This serves as a reference for the "correct" directionality of the target state.

• Direction Agreement Check: For each boundary cell i identified in Step 2:

  • Retrieve its low-dimensional velocity vector V_i_embed.
  • Calculate the cosine similarity between V_i_embed and the reference direction vector of cluster B (or simply the vector pointing from cell i to the centroid of B).
  • If the cosine similarity is positive, the velocity vector for cell i is pointing towards the target cluster B and is counted as correct.

• Metric Calculation: The Cross-Boundary Correctness score is the fraction (or percentage) of boundary cells in source cluster A for which the velocity vector was deemed "correct." A high score provides confidence that the velocity analysis supports the hypothesized trajectory.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Computational Tools for RNA Velocity Evaluation

Tool / Resource	Function	Application Note
scVelo (Python)	Estimates RNA velocity using dynamical models; includes functions for basic visualization and analysis [12] [4].	The primary workhorse for velocity estimation. Its scvelo.tl.velocity_confidence function can be used to derive a cell-wise consistency measure.
veloVI (Python)	A deep generative model for estimating RNA velocity that provides full posterior uncertainty quantification [17].	Crucial for moving beyond point estimates. Use it to identify cell states where velocity direction is highly uncertain, thus preventing over-interpretation in fragile cancer datasets.
CellRank	Infers cell fate probabilities and state transitions by combining RNA velocity with transcriptomic similarity [4] [17].	Goes beyond visualization to compute robust trajectories and terminal states, which can be used as a more stable ground truth for cross-boundary correctness checks.
k-NN Graph	A foundational data structure built from transcriptomic data.	Central to both velocity estimation in many tools (e.g., for smoothing) and to the calculation of consistency metrics. Its construction parameters (e.g., k, distance metric) significantly impact results [4].
Cosine Similarity	A measure of similarity between two vectors.	The standard metric for comparing the direction of high-dimensional velocity vectors between a cell and its neighbors to calculate consistency [17].

RNA velocity analysis has emerged as a powerful computational framework for inferring cellular dynamics from single-cell RNA sequencing (scRNA-seq) data. By leveraging the ratio of unspliced to spliced messenger RNA (mRNA), these methods can predict the future state of individual cells, enabling the reconstruction of developmental trajectories and the identification of transitional cell states [74] [75]. This capability is particularly valuable in cancer research, where understanding tumor evolution, intra-tumoral heterogeneity, and cell fate decisions is crucial for developing targeted therapies. Within this landscape, three distinct computational tools—scVelo, Dynamo, and TSvelo—have implemented increasingly sophisticated approaches to RNA velocity estimation, each with unique strengths and limitations for specific biological contexts.

The pancreas dataset, which models endocrine cell differentiation from ductal cells through Ngn3 high endocrine progenitors to mature α, β, δ, and ε-cells, has served as a fundamental benchmark for evaluating RNA velocity methods [12] [76]. Similarly, cancer datasets present additional challenges including complex multi-lineage differentiation, increased heterogeneity, and non-directional state transitions. This protocol provides a comprehensive comparative analysis of these three methods, with specific application notes for implementing them in pancreas and cancer datasets, framed within the broader context of investigating cancer dynamics at single-cell resolution.

Methodological Foundations

Core Mathematical Frameworks

The three methods compared herein represent different generations of RNA velocity estimation, with progressively more sophisticated mathematical frameworks:

Table 1: Core Methodological Frameworks

Method	Primary Approach	Key Innovations	Regulatory Integration
scVelo	Expectation-Maximization with dynamical modeling	Generalizes RNA velocity to transient cell states; solves full transcriptional dynamics of splicing kinetics [74]	Limited regulatory inference
Dynamo	Differential geometry + vector field reconstruction	Inclusive model incorporating metabolic labeling; absolute RNA velocity; maps transcriptomic vector fields [30]	Infers regulation via RNA Jacobian and differential geometry
TSvelo	Neural ODEs with regulatory cascade modeling	Models gene regulation, transcription, and splicing simultaneously; highly interpretable parameters; unified latent time [12] [77]	Directly integrates TF-target relations from databases (ChEA, ENCODE)

scVelo established a significant advancement beyond the original steady-state model by solving the full transcriptional dynamics of splicing kinetics using a likelihood-based dynamical model [74]. Dynamo extends this further by incorporating metabolic labeling data and reconstructing continuous vector fields that can predict cell fates [30]. TSvelo represents the most recent approach, using neural ordinary differential equations (ODEs) to model the complete cascade of gene regulation, transcription, and splicing in a unified framework [12].

Workflow Comparison

The following diagram illustrates the core computational workflows for each method, highlighting their distinct approaches to RNA velocity estimation:

Figure 1: Comparative workflows of scVelo, Dynamo, and TSvelo, highlighting their distinct approaches to RNA velocity estimation from single-cell RNA sequencing data.

Performance Benchmarking in Pancreas Datasets

Experimental Setup and Evaluation Metrics

The pancreatic endocrinogenesis dataset has become the standard benchmark for RNA velocity methods, containing 3,696 cells with transcriptome profiles sampled from embryonic day 15.5, capturing the differentiation from ductal cells to four major endocrine fates [76]. For comparative analysis, we implemented all three methods following standardized preprocessing while maintaining their specific optimal parameters.

Table 2: Performance Metrics on Pancreas Dataset

Method	Velocity Consistency	In-Cluster Coherence	Cross-Boundary Correctness	Latent Time Accuracy	Multi-lineage Support
scVelo	0.41	0.38	0.42	Medium	Limited
Dynamo	0.46	0.41	0.45	Medium-High	Moderate
TSvelo	0.52	0.47	0.49	High	Full

Performance was evaluated using standardized metrics including velocity consistency (coherence of velocity vectors among neighboring cells), in-cluster coherence (agreement with cluster identities), cross-boundary correctness (accurate prediction of transitions between cell types), and latent time accuracy (correlation with known developmental ordering) [12].

Pancreas-Specific Protocol

Sample Preparation and Data Generation:

• Tissue Processing: Isolate pancreatic epithelial and Ngn3-Venus fusion (NVF) cells during secondary transition at embryonic day 15.5 using established protocols [76].
• Single-Cell RNA Sequencing: Prepare libraries using 10X Genomics Chromium platform with standard chemistry.
• Spliced/Unspliced Quantification: Process raw sequencing data through either:
  • Velocyto command line tool: velocyto run10x -m repeats.gtf cellranger_output/ transcriptome.gtf
  • Kallisto-Bustools pipeline for improved quantification accuracy

Data Preprocessing (Standardized Across Methods):

Method-Specific Implementation:

scVelo Protocol (Dynamical Mode):

Dynamo Protocol:

TSvelo Protocol:

Results and Interpretation

In the pancreas benchmark, TSvelo demonstrated superior performance in capturing the complete differentiation trajectory from ductal to endocrine cells, achieving the highest scores across all quantitative metrics [12]. scVelo effectively identified major transitions but showed limitations in resolving fine-grained dynamics between closely related endocrine progenitors. Dynamo provided robust velocity estimates with additional capabilities for identifying putative driver genes through differential geometry analysis.

A key advantage of TSvelo in this context was its ability to accurately model genes with complex dynamics, such as ANXA4, which exhibits non-monotonic expression patterns (initial decrease followed by increase) that are challenging for conventional phase portrait-based methods [12]. The integration of transcriptional regulation directly into the velocity model enabled more accurate distinction between cell types that overlap in conventional unspliced-spliced phase portraits.

Application to Cancer Datasets

Special Considerations for Cancer Biology

Cancer single-cell datasets present unique challenges for RNA velocity analysis, including:

• High heterogeneity: Both intra-tumoral and inter-cellular diversity complicate trajectory inference
• Complex lineage relationships: Multiple parallel differentiation pathways and branching events
• Aneuploidy and copy number variations: Impact splicing ratios and gene expression quantification
• Metabolic and microenvironmental influences: Cancer-specific perturbations to transcriptional kinetics

Cancer-Specific Protocol Adaptations

Data Preprocessing for Cancer Samples:

Multi-lineage Analysis with PAGA Integration:

Putative Driver Gene Identification:

Cancer-Specific Performance Considerations

In cancer datasets, each method demonstrates distinct advantages:

scVelo provides the most computationally efficient solution for large-scale cancer datasets (e.g., >50,000 cells) but may oversimplify complex regulatory relationships in highly heterogeneous samples.

Dynamo excels in identifying master regulators of cell fate decisions through curvature analysis of the reconstructed vector field, particularly valuable for understanding therapeutic resistance mechanisms.

TSvelo offers the most biologically interpretable model for cancer progression, directly linking transcription factor activity to velocity estimates, enabling mechanistic insights into tumor evolution.

The following diagram illustrates how each method approaches the complex problem of RNA velocity estimation in cancer datasets:

Figure 2: Method-specific approaches and optimal applications for cancer single-cell RNA sequencing data analysis, highlighting their distinct advantages in addressing tumor heterogeneity and complex lineage relationships.

The Scientist's Toolkit

Essential Research Reagent Solutions

Table 3: Critical Reagents and Computational Tools for RNA Velocity Analysis

Category	Specific Tool/Reagent	Function	Method Compatibility
Sequencing Platforms	10X Genomics Chromium	Single-cell library preparation	All methods
Splicing Quantification	Velocyto command line tool	Generate spliced/unspliced matrices from BAM files	All methods
Splicing Quantification	Kallisto-Bustools	Improved quantification of spliced/unspliced reads	All methods
TF-Target Databases	ChEA3, ENCODE	Curated transcription factor-target interactions	TSvelo (primary), Dynamo
Metabolic Labeling	scSLAM-seq, 4sU	Time-resolved RNA kinetics	Dynamo (primary)
Spatial Validation	MERFISH, Visium	Spatial validation of predicted trajectories	All methods
Lineage Tracing	LARRY barcoding	Ground truth for fate prediction validation	All methods

Validation Framework

Establishing confidence in RNA velocity predictions requires multi-modal validation:

Experimental Validation Approaches:

• Lineage Tracing: Integration with CRISPR-based lineage barcoding (e.g., LARRY) provides ground truth for fate predictions [77].
• Spatial Transcriptomics: MERFISH or Visium can spatially validate predicted transitions in tissue context.
• Perturbation Studies: CRISPRi/knockdown of predicted driver genes to test functional importance.
• Metabolic Labeling: scSLAM-seq or 4sU labeling provides direct measurement of RNA kinetics for model validation [30].

Computational Validation Metrics:

• Velocity Consistency: Coherence of velocity vectors within cell neighborhoods.
• Pseudotime Accuracy: Correlation with known temporal ordering (when available).
• Transition Likelihood: Agreement with independently validated lineage relationships.
• Gene Dynamics Fit: Goodness-of-fit for individual gene phase portraits.

Through comprehensive benchmarking in pancreas datasets and adaptation for cancer biology applications, each RNA velocity method demonstrates distinct advantages for specific research contexts:

scVelo remains the most accessible and computationally efficient option for standard differentiation analysis, particularly valuable for initial exploration of new datasets or when working with large sample sizes (>50,000 cells).

Dynamo provides superior capabilities for mechanistic investigations, particularly when metabolic labeling data is available or when the research question involves identifying master regulators of fate decisions through differential geometry analysis.

TSvelo represents the most advanced approach for integrating regulatory information and modeling complex multi-lineage dynamics, making it particularly valuable for cancer applications where understanding transcriptional regulatory networks is essential.

For cancer dynamics research specifically, we recommend a tiered approach: beginning with scVelo for initial dataset exploration, followed by Dynamo for identifying putative therapeutic targets through driver analysis, and implementing TSvelo for detailed investigation of regulatory mechanisms underlying tumor evolution and therapeutic resistance.

As the field advances, integration of multi-omic measurements—particularly chromatin accessibility and protein abundance—with RNA velocity models will further enhance their precision and biological interpretability. The methods compared herein represent progressively sophisticated approaches to transforming static single-cell snapshots into dynamic models of cellular behavior, with profound implications for understanding cancer progression and treatment response.

Within the broader context of a thesis on RNA velocity analysis in single-cell cancer dynamics, this document establishes a critical framework for validating computational predictions. RNA velocity models infer future cellular states from single-cell RNA sequencing (scRNA-seq) data, but their predictions regarding cellular trajectories and fate decisions in cancer require rigorous experimental verification. This protocol details the integration of two powerful orthogonal approaches: single-cell Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq) to assess the epigenetic feasibility of predicted trajectories, and single-cell lineage tracing (scLT) to provide ground-truth evidence of clonal relationships and fate outcomes. By combining these multi-omic and lineage tools, researchers can move beyond correlation and establish causal, mechanistically supported models of tumor evolution and cellular plasticity.

Experimental Protocols for Multi-Omic Validation

Validating RNA Velocity Predictions with scATAC-seq

This protocol describes how to use chromatin accessibility data to assess whether the gene expression dynamics predicted by RNA velocity are supported by concomitant changes in the epigenomic landscape.

Principle: The core hypothesis is that sustained changes in gene expression, such as those predicted during a differentiation trajectory or a cell state transition in cancer, are often preceded or accompanied by changes in chromatin accessibility at associated regulatory elements. scATAC-seq enables the mapping of open chromatin regions genome-wide at single-cell resolution, providing a readout of the active regulatory state.

Procedure:

• Parallel Single-Cell Multi-omics Profiling:

  • Generate a paired dataset from the same cancer sample (e.g., a patient-derived xenograft or primary tumor dissociation) using one of the following strategies:
    • Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE-seq): Profiles transcriptome and surface proteins simultaneously.
    • 10x Genomics Multiome ATAC + Gene Expression: Co-assays chromatin accessibility and gene expression from the same single nucleus.
    • SHARE-seq: Simultaneously maps chromatin accessibility and gene expression in single cells.

• Data Preprocessing and Integration:

  • Process scRNA-seq data to compute RNA velocity using a chosen model (e.g., scVelo, Velocyto, or TSvelo).
  • Process scATAC-seq data by performing quality control, peak calling, and generating a cell-by-peak matrix.
  • Harmonize the two datasets using canonical correlation analysis or a similar integration method to align cells across modalities, ensuring that the cellular states are directly comparable.

• Epigenetic Corroboration of Predicted Trajectories:

  • Project the RNA velocity stream onto a low-dimensional embedding (e.g., UMAP) to visualize the predicted directionality of cell state transitions.
  • Overlay the scATAC-seq data for the same cells onto the same embedding. The chromatin accessibility patterns should reflect a continuum along the velocity-predicted path.
  • Identify key driver genes of the predicted trajectory from the RNA velocity analysis.
  • For these driver genes, identify linked cis-regulatory elements (e.g., promoters, enhancers) from the scATAC-seq data. An increase in accessibility at these regulatory regions should correlate with the predicted upregulation of the target gene.
  • Utilize a framework like HALO, which explicitly models the causal relationships between chromatin accessibility (scATAC-seq) and gene expression (scRNA-seq) over time, to quantitatively distinguish between coupled and decoupled dynamics.

• Interpretation: A successful validation is achieved when the epigenetic landscape from scATAC-seq shows a progressive reconfiguration along the RNA velocity vector, with opening of chromatin at regulatory elements associated with genes predicted to be activated, and closing at regions associated with genes predicted to be silenced.

Ground-Truth Validation with Single-Cell Lineage Tracing

This protocol outlines the use of heritable cellular barcodes to empirically track cell fate and directly test the lineage relationships predicted by RNA velocity models.

Principle: Prospective lineage tracing marks progenitor cells with unique, heritable DNA barcodes that are passed to all progeny. By combining barcode sequencing with single-cell transcriptomics, one can construct high-resolution lineage trees and unambiguously determine which progenitor cell gave rise to which descendant cell population, providing a "ground-truth" map of cellular relationships.

Procedure:

• Cell Tagging and Tracing Strategies:

  • Strategy A: Synthetic Barcodes (CellTag-Multi)
    • Tagging: Engineer cancer cells (e.g., a patient-derived organoid culture) to express a complex library of lentiviral barcodes (CellTags). Sequential tagging can be used to create multi-level lineage trees.
    • Multi-omic Capture: Profile the CellTagged population using both scRNA-seq and scATAC-seq. The modified CellTag-multi construct allows for the capture of barcodes in both modalities, enabling independent lineage reconstruction from transcriptomic and epigenomic data.
    • Lineage Reconstruction: Recover CellTag barcodes from sequencing data, apply error correction, and cluster cells with shared barcode signatures into clones.
  • Strategy B: Endogenous Barcodes (EMBLEM)
    • Principle: Leverage naturally occurring somatic mutations in mitochondrial DNA (mtDNA) as endogenous, heritable lineage barcodes.
    • Capture: scATAC-seq libraries are highly enriched for mtDNA, providing deep coverage for variant calling.
    • Lineage Inference: Identify heteroplasmic mtDNA mutations from single-cell data. Cells sharing the same combinations of mtDNA mutations are inferred to be clonally related.

• Integrating Lineage Data with RNA Velocity:

  • Map the reconstructed lineage information from either Strategy A or B onto the same low-dimensional space as the RNA velocity projection.
  • Key Validation: Assess if cells within the same lineage branch (i.e., siblings) are positioned along the same RNA velocity stream. The velocity-predicted transitions should align with the empirically determined lineage relationships.
  • Quantitatively test if the progenitor state of a clone, as defined by its lineage barcode, can predict the terminal fate of its descendants as suggested by the RNA velocity model.

• Interpretation: A strong correlation between lineage barcode-derived clonal relationships and RNA velocity-predicted trajectories provides powerful, direct evidence for the accuracy of the computational model. Discrepancies can reveal limitations of the velocity model or the presence of non-transcriptional fate determinants.

Computational Analysis Workflow

The following diagram illustrates the integrated computational workflow for validating RNA velocity predictions using multi-omics and lineage data.

Table 1: Key Methodologies for Multi-omic and Lineage Tracing Validation

Method	Core Principle	Measured Output	Key Strength in Validation	Considerations
HALO [78]	Causal modeling of chromatin accessibility & gene expression	Coupled vs. decoupled latent representations	Distinguishes synchronized from independent changes across modalities	Requires paired multi-omic data
CellTag-multi [79]	Lentiviral delivery of heritable RNA barcodes	Clonal lineages from scRNA-seq & scATAC-seq	Direct, prospective lineage tracking across modalities	Requires genetic engineering of system
EMBLEM [80]	Leverages endogenous mtDNA mutations as barcodes	Clonal lineages from scATAC-seq data	Applicable to human samples & archival tissue; no engineering needed	Relies on sufficient mtDNA mutation burden and coverage
VeloCycle [39]	Bayesian RNA velocity on a constrained manifold	Dynamically consistent velocity field & kinetic parameters	Provides statistical rigor and uncertainty quantification for velocities	Well-suited for periodic processes like cell cycle

Table 2: Comparison of Lineage Tracing Approaches in the Context of Cancer Dynamics

Feature	Synthetic Barcodes (CellTag-multi)	Endogenous Barcodes (EMBLEM)
Resolution	Very high, tunable via barcode complexity	Lower, dependent on somatic mutation rate
Applicability	Model systems, in vitro cultures, engineered cells	Human patients, archival samples, any eukaryotic cell
Multi-omic Compatibility	Explicitly designed for scRNA-seq and scATAC-seq	Primarily from scATAC-seq data
Ground-Truth Power	Directly links initial progenitor state to final fate	Infers lineage based on shared mutation history
Best Suited For	Tracking early, rapid fate decisions in controlled settings	Reconstructing clonal evolution in patient tumors

The Scientist's Toolkit: Essential Reagents and Solutions

Table 3: Key Research Reagent Solutions for Multi-omic Validation

Item / Reagent	Function / Application	Example & Notes
CellTag-multi Library [79]	A complex pool of lentiviral constructs for cell barcoding.	Contains ~80,000 unique barcodes; allows for robust, multi-level lineage tracing.
Nextera-Compatible Adapter Primers [79]	Enable amplification and capture of CellTag barcodes in scATAC-seq workflows.	Critical modification for integrating synthetic barcodes into standard scATAC-seq libraries.
10x Genomics Multiome Kit	Commercial solution for co-assaying gene expression and chromatin accessibility from the same single nucleus.	Simplifies generation of paired datasets; ensures cell identity is perfectly matched across modalities.
Validated mtDNA Primers [80]	For targeted amplification and sequencing of mitochondrial genome.	Used to enhance coverage for EMBLEM analysis or validate mtDNA variants called from scATAC-seq.
HALO Software Framework [78]	Computational tool for hierarchical causal modeling of multi-omics data.	Used post-data generation to quantitatively analyze the causal relationships between ATAC and RNA modalities.
VeloCycle Software Framework [39]	Bayesian tool for statistically robust RNA velocity inference on manifolds.	Provides a principled way to generate the initial velocity predictions that will be validated.

Within the broader thesis on RNA velocity analysis in single-cell cancer dynamics research, benchmarking computational tools is a critical step. RNA velocity, by modeling the temporal dynamics of spliced and unspliced messenger RNA (mRNA), predicts cellular future states from static single-cell RNA sequencing (scRNA-seq) snapshots, offering unparalleled insights into cancer progression, intratumoral heterogeneity, and therapeutic resistance [8]. However, the performance of these tools degrades in complex biological systems, such as multi-fate lineages in cancer, where cells exhibit heterogeneous kinetic rates and branch into multiple trajectories [81] [26]. This application note provides a structured benchmarking analysis and detailed protocols for applying RNA velocity tools to multi-lineage cancer datasets, enabling researchers to accurately reconstruct dynamic tumor ecosystems.

Performance Benchmarking of RNA Velocity Tools

Table 1: Benchmarking of RNA Velocity Tools on Complex Lineages

Tool	Core Methodology	Key Strengths	Performance on Complex Lineages	Cited Evidence
cell2fate	Fully Bayesian model with linearization of ODEs into interpretable modules [33].	High statistical power for weak dynamical signals; resolves complex transcriptional boosts; fully Bayesian framework [33].	Correctly inferred directionality in all 5 benchmark datasets; best average Cross-Boundary Directional Correctness (CBDir) score [33].	Applied to developing human brain; spatially mapped RNA velocity modules [33].
TSvelo	Neural ODEs modeling cascade of gene regulation, transcription, and splicing [12].	Models 3D gene dynamics; infers unified latent time; incorporates TF regulation [12].	Superiority demonstrated on 6 scRNA-seq datasets, including multi-lineage; highest velocity consistency [12].	Accurately predicted ductal to endocrine cell differentiation in pancreas data [12].
DeepVelo	Graph Convolutional Network (GCN) inferring cell-specific and gene-specific kinetic rates [81].	Infers time-varying kinetics; robust in multi-lineage systems; identifies driver genes [81].	Highest direction accuracy and consistency across developmental and pathological datasets [81].	Effectively captured neurogenesis in mouse dentate gyrus and pilocytic astrocytoma heterogeneity [81].
cellDancer	Deep neural network (DNN) implementing a "relay velocity" model for cell-specific kinetics [26].	Provides single-cell resolution of kinetic rates; robust to multi-rate kinetics and imbalanced lineages [26].	Lower error rates than scVelo, velocyto, DeepVelo, VeloVAE in simulated multi-rate kinetics [26].	Recapitulated erythroid maturation and hippocampus development; identified fate-indicating kinetics in mouse pancreas [26].
VeloVGI	Variational Graph Autoencoder (VGAE) with optimal transport for batch effect correction [6].	Corrects batch effects in velocity estimation; integrates multi-batch data for global dynamics [6].	Outperformed other methods on mouse spinal cord and olfactory bulb datasets with batch effects [6].	Parsed neurodevelopmental heterogeneity and immune cell dynamics in spinal cord injury data [6].
spVelo	Combines VAE and Graph Attention Network (GAT) for spatial transcriptomics [38].	Leverages spatial information; performs multi-batch integration; enables downstream spatial analysis [38].	Achieved highest direction and transition scores on simulated pancreas and oral squamous cell carcinoma data [38].	Provided insights into tumor architecture and cell-cell communication in cancer spatial data [38].

Essential Toolkit for RNA Velocity Analysis

Table 2: Research Reagent Solutions for RNA Velocity Workflow

Item	Function in Workflow	Specification Notes
scRNA-seq Library	Provides raw spliced and unspliced count matrices, the fundamental input for all velocity models [8].	Must be compatible with intron-aware alignment tools (e.g., velocyto, kallisto) to distinguish spliced/unspliced reads.
TF-Target Databases (e.g., ChEA, ENCODE)	Provides prior knowledge on gene regulatory networks for models that incorporate transcriptional regulation [12].	Used by TSvelo to model the influence of transcription factors on target gene transcription rates.
Spatial Transcriptomics Data	Enables the integration of spatial coordinates with gene expression for spatial velocity inference [38].	Platforms like Visium or MERFISH provide the input for spVelo to model tissue organization and dynamics.
Cell Annotations	Provides ground-truth cell type labels for method training (e.g., LatentVelo's annotated mode) and result validation [38].	Critical for benchmarking metrics like Cross-Boundary Directional Correctness (CBDir) [33].
Batch Metadata	Identifies samples from different experimental conditions or technical replicates for batch-effect correction [6].	Essential for running multi-batch models like VeloVGI and spVelo to infer globally consistent dynamics.

Visualizing Tool Selection and Application

Figure 1: Decision workflow for selecting RNA velocity tools on complex cancer lineages.

Experimental Protocols for Robust Benchmarking

Protocol 1: Standardized Workflow for Velocity Inference

This protocol outlines the core steps for applying and benchmarking RNA velocity tools, using the pancreas endocrinogenesis dataset as a canonical example [16].

• Data Loading and Preprocessing:

  • Load the data (e.g., pancreas.h5ad) containing spliced and unspliced counts in the layers of an AnnData object [16].
  • Filter genes based on a minimum count threshold (e.g., min_shared_counts=20) to remove uninformative genes.
  • Normalize total counts per cell for both spliced and unspliced matrices.
  • Select highly variable genes (e.g., n_top_genes=2000) to focus the analysis.

• Data Smoothing and Moment Calculation:

  • Perform dimensionality reduction (Principal Component Analysis (PCA)).
  • Construct a k-nearest neighbor (k-NN) graph based on cell-cell similarities.
  • Calculate moments (mean expression) of spliced (Ms) and unspliced (Mu) counts across neighboring cells to smooth the data [16].

• Velocity Inference:

  • Apply the chosen velocity tool (e.g., scVelo's dynamical model, cellDancer, DeepVelo).
  • For scVelo's dynamical model: Run scv.tl.recover_dynamics(adata) to estimate kinetic parameters and latent time, followed by scv.tl.velocity(adata, mode='dynamical') to compute velocities [16].
  • For cellDancer or DeepVelo, follow their respective packages' functions to train the model and infer velocities.

• Projection and Visualization:

  • Project the high-dimensional velocity vectors onto a low-dimensional embedding (e.g., UMAP) using a velocity graph: scv.tl.velocity_graph(adata).
  • Visualize the velocity streamlines: scv.pl.velocity_embedding_stream(adata, basis='umap') [16].

Protocol 2: Benchmarking Metrics and Validation

• Quantitative Metric Calculation:

  • Velocity Confidence: Measures the local coherence of velocity vectors among a cell's nearest neighbors. Higher values indicate more reliable estimates [12] [38].
  • Cross-Boundary Directional Correctness (CBDir): Evaluates whether inferred transition probabilities at the boundaries between known cell types are consistent with established biological knowledge (e.g., from ductal to endocrine cells) [33]. This is a key metric for validating directionality.
  • Direction Score (or Transition Score): Assesses the consistency of predicted cell state transitions with the actual observed progression of cells in a reduced dimension space or with known cell type sequences [38].

• Qualitative Inspection:

  • Examine the velocity stream plot on the UMAP embedding. The streamlines should follow the expected differentiation trajectory without reversed or chaotic flows [4].
  • For key driver genes, inspect the phase portrait (unspliced vs. spliced counts). The model's fitted curve should capture the overall distribution of cells, such as the almond-shaped pattern for standard kinetics or more complex shapes for transcriptional boosts [12] [26].

Critical Limitations and Best Practices

Despite their advanced capabilities, RNA velocity methods have inherent limitations. A significant reliance on k-NN smoothing during preprocessing means that performance is highly sensitive to the quality of this graph; an inaccurate graph can lead to substantial errors in both the direction and speed of estimated velocities [4]. Furthermore, estimating absolute speed from RNA velocity is notoriously unreliable except in very low-noise settings, cautioning against over-interpreting velocity vector lengths [4]. Users should be wary of the circular logic that can arise when using RNA velocity to validate the same low-dimensional embedding upon which it was projected [4].

Best practices recommend:

• Preprocessing Rigor: Carefully perform gene filtering, normalization, and neighbor graph construction. Consider the impact of these steps on downstream results.
• Model Selection: Choose a model whose assumptions align with the biological system. For complex multi-fate cancer data, prioritize tools like cell2fate, DeepVelo, and cellDancer that relax the assumption of constant kinetics.
• Multi-Method Validation: Never rely on a single tool. Run multiple methods and compare their results for consistency. Use quantitative metrics and qualitative biological plausibility for final judgment.
• Interpretation Focus: Prioritize the directionality of the velocity field over the speed. Use the tool to generate hypotheses about cellular trajectories, which should be subsequently validated through experimental approaches.

Single-cell RNA sequencing (scRNA-seq) has revolutionized biology by enabling high-throughput quantification of gene expression at individual cell resolution, providing unprecedented insights into cellular heterogeneity in complex tissues like tumors [1]. However, a fundamental limitation of standard scRNA-seq is that it provides only static cellular snapshots, obscuring dynamic temporal processes such as cellular differentiation, reprogramming, and disease progression [1]. RNA velocity, introduced in 2018, offers a groundbreaking solution to this limitation by leveraging the intrinsic temporal information contained in the ratio of unspliced pre-mRNA to spliced mature mRNA to predict future transcriptional states of cells over hour-long timescales [1] [8].

In cancer research, RNA velocity analysis provides a powerful computational framework for modeling tumor evolution, intratumoral heterogeneity, and metastatic progression. The ability to infer the directionality of cellular state transitions from snapshot data has profound implications for understanding cancer development, drug resistance mechanisms, and identifying potential therapeutic targets [82]. This application note outlines comprehensive protocols for implementing RNA velocity analysis in cancer dynamics research, with emphasis on experimental validation strategies to bridge computational predictions with biological discovery.

Computational Framework of RNA Velocity

Theoretical Foundations

RNA velocity models are grounded in a first-order kinetics framework that describes the transcription, splicing, and degradation processes of messenger RNA. The fundamental dynamical system is described by ordinary differential equations:

$$\begin{array}{rcl}\frac{du(t)}{dt} & = & \alpha(t) - \beta u(t) \ \frac{ds(t)}{dt} & = & \beta u(t) - \gamma s(t)\end{array}$$

where (u(t)) and (s(t)) represent the abundance of unspliced and spliced mRNA at time (t), respectively, while (\alpha(t)), (\beta), and (\gamma) denote the transcription, splicing, and degradation rates [32].

The core premise of RNA velocity is that by quantifying the relative ratios of unspliced to spliced mRNAs, one can infer the instantaneous rate of gene expression change and predict the future state of individual cells. A positive RNA velocity indicates gene induction, while a negative velocity indicates repression [8]. When aggregated across the transcriptome, these velocity vectors can reveal the direction of cellular state transitions, such as lineage commitment in stem cells or epithelial-to-mesenchymal transition in cancer cells.

Evolution of RNA Velocity Methods

Since the introduction of the original Velocyto algorithm, numerous advanced computational tools have been developed that generalize the foundational framework. These methods can be broadly categorized into three classes based on their approaches to transcriptional kinetics inference [8]:

Table 1: Categories of RNA Velocity Methods

Category	Representative Methods	Underlying Approach	Strengths	Limitations
Steady-state Methods	Velocyto, scVelo (stochastic)	Analytical models assuming constant splicing rate and transcriptional equilibrium	Simple, fast, and interpretable; effective for steady-state differentiation	Assumptions violated in heterogeneous populations; inaccurate for complex kinetics
Trajectory Methods	scVelo (dynamical), dynamo, veloVI	Estimate parameters to construct phase portrait trajectories aligning cells with latent times	Handles transient states; infers full transcriptional dynamics	Computationally intensive; may overfit noisy data
State Extrapolation Methods	cellDancer, DeepVelo, UniTVelo	Leverage expected future cell states to optimize cell-level RNA velocity vectors	Cell-specific kinetics; robust to multi-rate kinetics	Complex implementation; requires substantial computational resources

Steady-state methods like Velocyto pioneered the field by using least-squares regression on steady-state subpopulations, assuming constant splicing rates and transcriptional equilibrium [8]. While effective for modeling clear differentiation processes, these methods struggle with complex kinetic patterns and non-steady states commonly encountered in cancer microenvironments.

Trajectory methods such as scVelo's dynamical model implemented a more sophisticated expectation-maximization framework capable of inferring transcriptional dynamics and assigning latent cell time [8]. These approaches relax the steady-state assumption and can generalize to multiple transcriptional states, providing more flexibility in modeling complex biological processes.

State extrapolation methods represent the latest evolution in RNA velocity algorithms. Tools like cellDancer employ a "relay velocity model" that uses deep neural networks to infer velocity for each cell from its neighbors, then relays a series of local velocities to provide single-cell resolution inference of kinetic parameters [26]. This approach overcomes limitations of conventional models that assume uniform kinetics across all cells, which often results in unpredictable performance in experiments with multi-stage and/or multi-lineage transitions where the assumption of identical kinetic rates for all cells no longer holds [26].

UniTVelo introduced a "temporally unified" approach that models spliced RNA dynamics using radial basis functions and infers a unified latent time across the transcriptome [32]. This innovation helps resolve directionality discrepancies between genes and reinforces temporal ordering of cells, particularly important in cancer datasets with complex branching trajectories.

More recently, TSvelo has advanced the field further by integrating the cascade of gene regulation, transcription, and splicing into a single ODE model that simultaneously captures 3D dynamics of all genes [12]. This framework incorporates transcriptional regulation information from transcription factor-target databases while maintaining parameter interpretability.

For spatial transcriptomics data, spVelo enables RNA velocity inference for multi-batch spatial datasets by combining a Variational AutoEncoder for gene expression with a Graph Attention Network for spatial location [38]. This approach utilizes spatial proximity to better infer trajectory patterns and cell-cell communication dynamics in tissue contexts.

Workflow for RNA Velocity Analysis

A typical RNA velocity analysis pipeline consists of several standardized steps [8]:

• Preprocessing: Distinguishing between unspliced and spliced transcripts in raw sequencing data to construct separate count matrices using tools like Velocyto.py or kallisto bustools.

• Data Smoothing: Applying sophisticated imputation techniques to extract reliable signals from noisy single-cell data, often computing the first-order moment across k-nearest neighbors in expression space.

• Velocity Estimation: Applying biophysical models to fit unspliced and spliced transcript counts, yielding kinetic parameters and high-dimensional velocity vectors.

• Projection and Visualization: Embedding velocity vectors into low-dimensional representations (UMAP, t-SNE) using streamline plots or grid-averaged vector fields.

• Downstream Analysis: Interpreting cellular dynamics through driver gene identification, trajectory analysis, and regulatory inference.

Table 2: Benchmark Performance of RNA Velocity Methods in Cancer-Relevant Contexts

Method	Multi-lineage Kinetics	Transcriptional Boost	Spatial Data Integration	Computational Efficiency	Uncertainty Quantification
Velocyto	Limited	Poor	No	High	No
scVelo	Moderate	Moderate	No	Moderate	Limited
cellDancer	High	High	No	Low	Yes
UniTVelo	High	High	No	Moderate	Limited
TSvelo	High	High	No	Low	Yes
spVelo	High	High	Yes	Low	Yes

Protocol: RNA Velocity Analysis in Cancer Research

Sample Preparation and Data Generation

Materials and Reagents:

• Fresh tumor tissue samples (primary and metastatic where available)
• Single-cell suspension kit (e.g., Tumor Dissociation Kit)
• Viability staining dye (e.g., Propidium Iodide or DAPI)
• Single-cell RNA sequencing platform (10x Genomics Chromium recommended)
• RNase inhibitors
• Library preparation reagents

Procedure:

• Obtain tumor biopsies from patients or animal models, ensuring rapid processing to maintain RNA integrity.
• Process tissues using a standardized dissociation protocol to generate single-cell suspensions while minimizing stress responses that could alter transcriptional states.
• Assess cell viability using counting methods and viability stains, aiming for >80% viability to ensure high-quality data.
• Proceed with scRNA-seq library preparation following manufacturer protocols, capturing both spliced and unspliced transcripts. The 10x Genomics platform is widely used and well-supported by RNA velocity tools.
• Sequence libraries with sufficient depth (recommended: ≥50,000 reads per cell) to robustly detect both mature and nascent transcripts.

Quality Control Considerations:

• Monitor standard QC metrics including total UMI counts, number of detected genes, and mitochondrial content [83].
• Exclude damaged cells (high mitochondrial content, low gene counts) and doublets (abnormally high gene counts) that can confound velocity analysis [83].
• For human samples, remove cells expressing high levels of hemoglobin genes (HBB) indicating red blood cell contamination [83].

Computational Implementation

Software Requirements:

• Python (≥3.8) with scVelo, scanny, pandas, numpy packages
• R (≥4.0) with velocyto.R, Seurat, Bioconductor packages
• Sufficient computational resources (recommended: ≥16GB RAM for datasets of 10,000 cells)

Protocol Steps:

• Data Preprocessing
  • Process raw sequencing data through Velocyto.py (version 0.17.17) to generate .loom files containing spliced and unspliced counts [84].
  • Filter cells based on quality metrics and normalize using library size normalization.
  • Select highly variable genes for downstream analysis, focusing on genes with sufficient expression in both spliced and unspliced counts.

• RNA Velocity Estimation

  • Load .loom files into velocyto.R (version 0.6.0) or Python environment.
  • Estimate RNA velocities using appropriate parameters (e.g., fit.quantile = 0.02; kCells = 25; DeltaT = 1) [84].
  • For complex cancer datasets with multiple lineages, employ methods like cellDancer or UniTVelo that handle multi-rate kinetics.
  • Validate velocity directions using known marker genes and differentiation trajectories.

• Visualization and Interpretation

  • Project velocity vectors onto low-dimensional embeddings (UMAP recommended).
  • Annotate cell clusters using established marker genes relevant to cancer biology.
  • Interpret velocity flow to identify progenitor states, differentiation trajectories, and transition states in tumor ecosystems.

Figure 1: RNA Velocity Analysis Workflow. The standard pipeline begins with raw data processing, proceeds through quality control and velocity estimation, and culminates in biological interpretation requiring experimental validation.

Experimental Validation Strategies

Principles of Corroboration: Computational predictions from RNA velocity analysis require rigorous experimental validation to transform algorithmic outputs into biological insights. A multi-modal approach to validation strengthens conclusions and builds confidence in the dynamic models.

Method 1: Lineage Tracing and Fate Mapping

• Purpose: Directly track cellular descendants and fate choices predicted by velocity vectors.
• Implementation: Employ genetic lineage tracing using Cre-lox systems or synthetic barcoding approaches in animal models.
• Correlation: Compare actual lineage outcomes with velocity-predicted trajectories.
• Cancer Application: Track the emergence of metastatic subclones or therapy-resistant populations predicted by velocity analysis.

Method 2: Perturbation Experiments

• Purpose: Functionally validate driver genes and regulatory networks identified through velocity analysis.
• Implementation: Utilize CRISPR-based gene knockout or RNA interference to perturb genes identified as key drivers in velocity transitions.
• Readout: Assess how perturbations alter the velocity flow and trajectory outcomes.
• Cancer Application: Test whether silencing genes associated with EMT transition blocks metastatic progression as predicted.

Method 3: Metabolic Labeling

• Purpose: Directly measure RNA kinetics using nucleotide analogs.
• Implementation: Incorporate 4-thiouridine (4sU) or 5-ethynyluridine (EU) to label newly transcribed RNA.
• Measurement: Quantify labeling kinetics to empirically determine transcription and degradation rates.
• Correlation: Compare experimentally measured rates with computationally inferred parameters from velocity algorithms.

Method 4: Spatial Validation

• Purpose: Confirm predicted spatial relationships and cell-cell communication.
• Implementation: Utilize multiplexed immunohistochemistry or in situ hybridization on serial tissue sections.
• Correlation: Validate whether cells predicted to be in transition states localize to expected tissue niches.
• Cancer Application: Confirm the presence of predicted intermediate states at the invasive front of tumors.

Case Study: RNA Velocity in Breast Cancer Progression

Application to Primary and Metastatic Tumors

In a recent study investigating estrogen receptor-positive (ER+) breast cancer, scRNA-seq was performed on an all-female cohort comprising individuals with either primary (n=12) or metastatic (n=11) disease [82]. Biopsies were obtained from multiple metastatic sites including liver, bone, lymph nodes, and skin. After rigorous quality control and integration to mitigate batch effects, 99,197 cells were analyzed encompassing malignant cells, myeloid cells, T cells, NK cells, B cells, endothelial cells, and fibroblasts [82].

RNA velocity analysis applied to this dataset revealed dynamic transitions between cellular states in the tumor microenvironment. Specifically, velocity vectors identified progenitor-like tumor cells and delineated their potential fate trajectories toward either luminal or basal-like states. The analysis also uncovered accelerated transition dynamics in metastatic samples compared to primary tumors, suggesting an increased plasticity in advanced disease [82].

Identification of Therapeutic Targets

Differential velocity analysis between primary and metastatic samples identified genes with altered kinetic parameters in malignant cells. These included transcription factors with accelerated induction rates in metastasis, suggesting their potential role as drivers of progression. Several of these factors were previously associated with therapy resistance, providing a mechanistic link between dynamic gene regulation and treatment failure [82].

Experimental validation using patient-derived organoids confirmed that perturbation of these velocity-identified driver genes altered the transition trajectories and reduced metastatic potential in xenograft models. This demonstrates the power of combining RNA velocity prediction with functional studies to identify novel therapeutic targets.

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for RNA Velocity Studies

Reagent/Category	Specific Examples	Function in RNA Velocity Workflow
Single-cell Platform	10x Genomics Chromium, Singleron GEXSCOPE	Generate partitioned single-cell libraries with barcoding for transcript counting
RNA Velocity Software	Velocyto, scVelo, cellDancer, UniTVelo	Implement core algorithms for velocity estimation from spliced/unspliced counts
Trajectory Analysis Tools	PAGA, CellRank, Slingshot	Infer directed trajectories and fate probabilities from velocity matrices
Spatial Transcriptomics	10x Visium, Slide-seq, MERFISH	Provide spatial context for validating velocity-predicted transitions
Lineage Tracing Systems	Cre-lox, Polylox, LINNAEUS	Enable direct fate mapping to validate velocity predictions
Metabolic Labeling	4-thiouridine (4sU), 5-ethynyluridine	Empirically measure RNA kinetics for validation
Perturbation Tools	CRISPR-Cas9, siRNA, Small molecules	Functionally test predictions by altering velocity-predicted driver genes

Discussion and Future Perspectives

RNA velocity methods have rapidly evolved from foundational steady-state models to sophisticated frameworks capable of handling complex multi-lineage kinetics, transcriptional bursts, and spatial constraints. In cancer research, these tools provide unprecedented insight into tumor evolution, cellular plasticity, and drug resistance mechanisms. However, several challenges remain that require continued methodological development and rigorous validation.

Key limitations in current RNA velocity analysis include sensitivity to technical noise, challenges in distinguishing closely related cell states, and computational demands for large-scale datasets. Furthermore, the assumption of constant kinetic parameters across cell types may not hold in complex tumor ecosystems with diverse microenvironments. Emerging methods like cellDancer that infer cell-specific kinetics and spVelo that incorporates spatial information represent promising directions for addressing these limitations [26] [38].

The integration of RNA velocity with other single-cell modalities—including epigenomics, proteomics, and spatial data—will further enhance its biological utility. Multi-omics velocity approaches can reveal how regulatory networks control transition dynamics and how epigenetic states influence trajectory outcomes. In cancer research, these integrated approaches may uncover novel mechanisms of metastasis and therapy resistance.

For the biomedical researcher, successful implementation of RNA velocity analysis requires careful experimental design, appropriate method selection based on biological context, and—most critically—rigorous experimental validation. Computational predictions should be viewed as hypotheses requiring functional confirmation rather than established biological facts. Through this iterative cycle of computational modeling and experimental testing, RNA velocity analysis will continue to transform our understanding of cancer dynamics and accelerate therapeutic discovery.

Figure 2: Iterative Validation Cycle. The process of transforming computational predictions into biological discovery requires an iterative approach where model refinement incorporates experimental findings.

Conclusion

RNA velocity has fundamentally expanded our capacity to infer dynamic cellular processes from static single-cell snapshots, providing unprecedented insights into cancer initiation, progression, and therapeutic resistance. This synthesis demonstrates that modern implementations—which integrate gene regulation, spatial context, and multi-batch data—are increasingly robust and biologically informative. Key takeaways include the importance of selecting models aligned with specific biological contexts, the critical need to address technical artifacts like batch effects, and the power of combining velocity predictions with orthogonal validation. Future directions point toward deeper integration with multi-omics and live imaging, enhanced scalability for massive clinical cohorts, and the translation of dynamic predictions into novel therapeutic and diagnostic strategies. As these computational methods mature, RNA velocity is poised to become a cornerstone of precision oncology, transforming our static molecular portraits of cancer into predictive, dynamic models of tumor behavior.

References

• 1. RNA velocity and beyond: Current advances in modeling ... [https://www.sciencedirect.com/science/article/pii/S1323893025000899]
• 2. RNA velocity unraveled - PMC - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC9499228/]
• 3. RNA Velocity Basics — scVelo documentation - Read the Docs [https://scvelo.readthedocs.io/en/stable/VelocityBasics.html]
• 4. Pumping the brakes on RNA velocity by understanding and ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-023-03065-x]
• 5. TIVelo: RNA velocity estimation leveraging cluster-level ... [https://www.nature.com/articles/s41467-025-61628-x]
• 6. Accurate RNA velocity estimation based on multibatch ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11657662/]
• 7. UChicago scientists use machine learning to turn cell ... [https://machinelearning.uchicago.edu/2025/03/26/uchicago-scientists-use-machine-learning-to-turn-cell-snapshots-dynamic/]
• 8. Paradigms, innovations, and biological applications of RNA ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12265890/]
• 9. Cell cycle gene regulation dynamics revealed by RNA ... [https://www.nature.com/articles/s41467-022-30545-8]
• 10. Steering research on mRNA splicing in cancer towards clinical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11698124/]
• 11. Cancer-associated snaR-A noncoding RNA interacts with ... [https://www.nature.com/articles/s41467-025-65448-x]
• 12. TSvelo: Comprehensive RNA velocity by modeling the ... [https://elifesciences.org/reviewed-preprints/108950]
• 13. 10 Best Practices for Effective RNA-Seq Data Analysis - Focal [https://www.getfocal.co/post/10-best-practices-for-effective-rna-seq-data-analysis]
• 14. A Beginner's Guide to Analysis of RNA Sequencing Data [https://pmc.ncbi.nlm.nih.gov/articles/PMC6096346/]
• 15. Unraveling the spatial and signaling dynamics and splicing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11966058/]
• 16. 14. RNA velocity [https://www.sc-best-practices.org/trajectories/rna_velocity.html]
• 17. Deep generative modeling of transcriptional dynamics for ... [https://www.nature.com/articles/s41592-023-01994-w]
• 18. Learning Explicit Single-Cell Dynamics Using ODE ... [https://arxiv.org/html/2510.02903v1]
• 19. SCODE: an efficient regulatory network inference algorithm ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5860123/]
• 20. A foundation model of transcription across human cell types [https://www.nature.com/articles/s41586-024-08391-z]
• 21. Review Functional analysis of cell plasticity using single ... [https://www.sciencedirect.com/science/article/pii/S0962892424000060]
• 22. RNA velocity of single cells - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6130801/]
• 23. Single-nucleus RNA velocity reveals critical synaptic and ... [https://www.nature.com/articles/s41598-024-57918-x]
• 24. spVelo: RNA velocity inference for multi-batch spatial transcriptomics data | Genome Biology [https://link.springer.com/article/10.1186/s13059-025-03701-8]
• 25. velocyto | Estimating RNA velocity in single cell RNA ... [http://velocyto.org/]
• 26. A relay velocity model infers cell-dependent RNA velocity [https://www.nature.com/articles/s41587-023-01728-5]
• 27. KSRV: a Kernel PCA-Based framework for inferring spatial ... [https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2025.1695803/full]
• 28. GraphVelo allows for accurate inference of multimodal ... [https://www.nature.com/articles/s41467-025-62784-w]
• 29. scVelo - RNA velocity generalized through dynamical ... [https://scvelo.readthedocs.io/]
• 30. aristoteleo/dynamo-release [https://github.com/aristoteleo/dynamo-release]
• 31. Mapping transcriptomic vector fields of single cells [https://www.sciencedirect.com/science/article/pii/S0092867421015774]
• 32. UniTVelo: temporally unified RNA velocity reinforces single ... [https://www.nature.com/articles/s41467-022-34188-7]
• 33. Cell2fate infers RNA velocity modules to improve cell fate ... [https://www.nature.com/articles/s41592-025-02608-3]
• 34. RNA velocity analysis with scVelo - Sam Morabito [https://smorabit.github.io/blog/2021/velocyto/]
• 35. Comparison of single-cell RNA Velocity methods in ... [https://github.com/czbiohub-sf/comparison-RNAVelo]
• 36. xiaoyeye/TFvelo [https://github.com/xiaoyeye/TFvelo]
• 37. TFvelo – gene regulation inspired RNA velocity estimation [https://www.rna-seqblog.com/tfvelo-gene-regulation-inspired-rna-velocity-estimation/]
• 38. spVelo: RNA velocity inference for multi-batch spatial ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03701-8]
• 39. Statistical inference with a manifold-constrained RNA ... [https://www.nature.com/articles/s41592-024-02471-8]
• 40. Quantifying and interpreting biologically meaningful spatial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11897387/]
• 41. Spatial multi-omics analyses of the tumor immune ... [https://jbiomedsci.biomedcentral.com/articles/10.1186/s12929-022-00879-y]
• 42. VeloVGI – accurate RNA velocity estimation based on ... [https://www.rna-seqblog.com/velovgi-accurate-rna-velocity-estimation-based-on-multibatch-network-reveals-complex-lineage-in-batch-scrna-seq-data/]
• 43. New method calculates rate of gene expression to understand ... [https://science.psu.edu/news/XueLong9-2025]
• 44. Dissecting tumor microenvironment from spatially resolved ... [https://www.nature.com/articles/s41467-024-49171-7]
• 45. Basal cell of origin resolves neuroendocrine–tuft lineage ... [https://www.nature.com/articles/s41586-025-09503-z]
• 46. Basal Stem Cells Identified as Origin of Small Cell Lung ... [https://corporate.dukehealth.org/news/basal-stem-cells-identified-origin-small-cell-lung-cancer-lab-models]
• 47. Single-cell analysis technologies for cancer research [https://pmc.ncbi.nlm.nih.gov/articles/PMC10602688/]
• 48. Single-cell RNA sequencing reveals immune ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11585206/]
• 49. Computing RNA velocity in a Bioconductor framework [https://www.bioconductor.org/packages/release/bioc//vignettes/velociraptor/inst/doc/velociraptor.html]
• 50. VeloViz: RNA velocity-informed embeddings for visualizing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8723140/]
• 51. Applications of single-cell technologies in drug discovery for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11338156/]
• 52. Single-cell transcriptomes identify patient-tailored ... [https://www.nature.com/articles/s41467-024-52980-5]
• 53. RNA velocity—current challenges and future perspectives - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8388041/]
• 54. Single-cell transcriptomics in cancer: computational ... [https://www.nature.com/articles/s12276-020-0422-0]
• 55. Application of single-cell genomics in cancer - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC4571998/]
• 56. TFvelo: gene regulation inspired RNA velocity estimation - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11258302/]
• 57. Representation learning of RNA velocity reveals robust cell ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8670433/]
• 58. A benchmark of batch - effect for correction - methods ... single cell RNA [https://pubmed.ncbi.nlm.nih.gov/31948481/]
• 59. Atlas-scale single-cell multi-sample multi-condition data ... [https://www.nature.com/articles/s41467-023-39923-2]
• 60. A benchmark of batch - effect for correction - methods ... single cell RNA [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-019-1850-9]
• 61. scCorrector: a robust method for integrating multi-study ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10810333/]
• 62. in Batch - effects - single data are cell by... RNA sequencing corrected [https://www.nature.com/articles/nbt.4091?error=cookies_not_supported&code=58a993b1-9076-4056-a3ad-83de64ceee8c]
• 63. Chapter 14 Multi-sample comparisons [https://bioconductor.org/books/3.12/OSCA/multi-sample-comparisons.html]
• 64. Comparison of Machine Learning Models for Colon Cancer ... [https://pubmed.ncbi.nlm.nih.gov/41297023/?utm_source=SimplePie&utm_medium=rss&utm_campaign=pubmed-2&utm_content=1bAXfGTh08tVk-e1kpwuqxoVfiUn_5olMN9ewCML2bdAMGPEOD&fc=20220524054416&ff=20251126185413&v=2.18.0.post22+67771e2]
• 65. Eye on AI: Applying Artificial Intelligence to Drive Cancer ... [https://www.aacr.org/blog/2025/08/18/eye-on-ai-applying-artificial-intelligence-to-drive-cancer-research-part-2/]
• 66. Comparative Analysis of Machine Learning Models for ... [https://www.medrxiv.org/content/10.1101/2025.01.02.25319913v1]
• 67. Top Oncology Innovations That Shaped the First Half of 2025 [https://binaytara.org/cancernews/article/top-oncology-innovations-that-shaped-the-first-half-of-2025]
• 68. Guide to Getting Started with 10X Genomics Single Cell RNA ... [https://biotech.wisc.edu/guide-to-getting-started-with-10x-genomics-single-cell-rna-seq/]
• 69. Navigating single-cell RNA-sequencing: protocols, tools ... [https://genomicsinform.biomedcentral.com/articles/10.1186/s44342-025-00044-5]
• 70. Establishing single cell RNA transcriptomics: a brief guide [https://pmc.ncbi.nlm.nih.gov/articles/PMC12403637/]
• 71. A practical guide for single-cell transcriptome data analysis ... [https://www.sciencedirect.com/science/article/pii/S0168010225000653]
• 72. VeTra: a tool for trajectory inference based on RNA velocity [https://pmc.ncbi.nlm.nih.gov/articles/PMC8545348/]
• 73. DeepVelo: deep learning extends RNA velocity to multi-lineage systems with cell-specific kinetics | Genome Biology | Full Text [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-023-03148-9]
• 74. Generalizing RNA velocity to transient cell states through ... [https://www.nature.com/articles/s41587-020-0591-3]
• 75. RNA velocity and beyond: Current advances in modeling ... [https://pubmed.ncbi.nlm.nih.gov/40973591/]
• 76. Endocrine Pancreas — scVelo documentation - Read the Docs [https://scvelo.readthedocs.io/en/stable/vignettes/Pancreas.html]
• 77. TSvelo: Comprehensive RNA velocity by modeling the ... - Sciety [https://sciety.org/articles/activity/10.1101/2024.12.24.630058]
• 78. HALO: hierarchical causal modeling for single cell multi- ... [https://www.nature.com/articles/s41467-025-63921-1]
• 79. Single-cell lineage capture across genomic modalities with ... [https://www.nature.com/articles/s41587-023-01931-4]
• 80. Single-cell lineage tracing by endogenous mutations ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6469926/]
• 81. DeepVelo: deep learning extends RNA velocity to multi ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10799431/]
• 82. Single-cell RNA sequencing reveals different cellular ... [https://www.nature.com/articles/s41523-025-00808-w]
• 83. Data analysis guidelines for single-cell RNA-seq in ... [https://mmrjournal.biomedcentral.com/articles/10.1186/s40779-022-00434-8]
• 84. RNA Velocity Analysis [https://bio-protocol.org/exchange/minidetail?id=10637844&type=30]