Benchmarking Single-Cell CNV Detection: A Comprehensive Guide to Algorithm Performance and Selection

Abstract

This article provides a comprehensive analysis of the current landscape of computational tools for detecting copy number variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data. Drawing from recent large-scale benchmarking studies, we explore the foundational principles of CNV inference, categorize methodological approaches, and present performance evaluations across diverse datasets and conditions. Key findings reveal significant performance differences among popular tools like CaSpER, CopyKAT, and inferCNV, with optimal selection dependent on specific research goals, data types, and technical parameters including sequencing depth, platform selection, and tumor purity. We offer practical troubleshooting guidance and optimization strategies to address common challenges including batch effects, reference selection, and data quality issues. This resource equips researchers and drug development professionals with evidence-based recommendations for selecting and implementing CNV detection methods in cancer genomics and clinical research applications.

The Landscape of Single-Cell CNV Detection: Principles, Challenges, and Biological Significance

Copy Number Variations (CNVs), the gain or loss of genomic regions, are a hallmark of cancer and play a crucial role in tumor evolution and heterogeneity by amplifying oncogenes or inactivating tumor suppressor genes [1] [2]. The emergence of single-cell RNA-sequencing (scRNA-seq) has provided an unprecedented opportunity to study this genetic heterogeneity within tumors. Consequently, several computational methods have been developed to infer CNVs from scRNA-seq data, expanding its application to study genetic heterogeneity using transcriptomic data [3] [2]. This guide objectively benchmarks the performance of these single-cell CNV detection algorithms based on recent, comprehensive studies, providing researchers with data-driven insights for tool selection.

Benchmarking Single-Cell CNV Detection Algorithms

Multiple independent benchmarking studies have evaluated the performance of popular scCNV inference tools, revealing that their performance varies significantly depending on the specific research application, scRNA-seq platform, and data quality [3] [2] [4].

The table below summarizes the overall findings from these benchmarking efforts, highlighting the recommended use cases for each method.

Tool Name	Primary Methodology	Performance & Recommended Use Cases	Key Limitations
CopyKAT [3] [2] [4]	Statistical model, segmentation approach [3]	- Overall CNV Inference: Top performer for sensitivity/specificity [2] [4].- Subclone Identification: Excels at identifying tumor subpopulations [2] [4].	Performance affected by batch effects in multi-platform data [2].
CaSpER [3] [2] [4]	Hidden Markov Model (HMM) integrating gene expression and allele frequency (AF) [3]	- Overall CNV Inference: Top performer for sensitivity/specificity, especially in large droplet-based datasets [3] [2] [4].	Requires higher runtime [3].
InferCNV [1] [2] [4]	Hidden Markov Model (HMM) on expression [3] [2]	- Subclone Identification: Excels in identifying tumor subpopulations from a single platform [2] [4].- Sensitivity: High sensitivity with sufficient sequenced cells [4].	- Does not directly predict tumor cells (infers CNV scores) [1].- Highly affected by batch effects [2] [4].
SCEVAN [1] [3]	Segmentation approach [3]	- Prediction: Moderate sensitivity but may overestimate tumor cells [1].	Overestimates true number of tumor cells; requires epithelial filtering [1].
sciCNV [1] [2]	Calculates expression disparity and concordance scores [2]	- Subclone Identification: Good performance in subclone identification from a single platform [2].	- Does not directly predict tumor cells (computes CNV scores) [1].- Score distribution may not clearly separate malignant/non-malignant cells [1].
HoneyBADGER [2] [4]	HMM + Bayesian method; can use expression and allele frequency [2]	- Batch Resilience: Allelic version more resilient to batch effects [4].- Sensitivity: Lower sensitivity for rare tumor populations [4].	Lower overall sensitivity [4].

Quantitative Performance Data from Benchmarking Studies

A major benchmarking study published in Nature Communications (2025) evaluated six tools on 21 scRNA-seq datasets using ground truth from whole-genome or whole-exome sequencing [3]. The study assessed the ability to correctly identify ground truth CNVs and euploid cells. The results, summarized in the table below, show that methods incorporating allelic information (like CaSpER and Numbat) performed more robustly for large droplet-based datasets, though they required higher runtime [3].

Method	Data Type Used	Performance on Euploid (PBMC) Dataset	Impact of Reference Dataset	Runtime & Memory
InferCNV [3] [2]	Expression only [3]	Performance varies with reference choice [3].	Significant impact [3].	Varies by method and data size [3].
CopyKAT [3]	Expression only [3]	Performance varies with reference choice [3].	Significant impact [3].	Varies by method and data size [3].
SCEVAN [3]	Expression only [3]	Performance varies with reference choice [3].	Significant impact [3].	Varies by method and data size [3].
CONICSmat [3]	Expression only (per chromosome arm) [3]	Performance varies with reference choice [3].	Significant impact [3].	Varies by method and data size [3].
CaSpER [3]	Expression + Allele Frequency [3]	More robust performance [3].	Lower impact; more robust [3].	Higher runtime requirements [3].
Numbat [3]	Expression + Allele Frequency [3]	More robust performance [3].	Lower impact; more robust [3].	Higher runtime requirements [3].

Another study in Precision Clinical Medicine (2025) focused on five tools, evaluating their sensitivity and specificity on a breast cancer cell line (HCC1395) versus a matched normal B-cell line across four scRNA-seq platforms (10x, C1, C1 HT, ICELL8) [2]. The following table synthesizes the key findings.

Tool	Sensitivity & Specificity (Overall)	Performance on 10x Data (0.5M reads/cell)	Performance on Full-Length Data (C1, ICELL8)
HoneyBADGER [2]	Lower than top performers [2].	N/A	N/A
InferCNV [2]	Varied, not top tier for overall inference [2].	Good performance for subclone identification [2].	Good performance for subclone identification [2].
sciCNV [2]	Varied, not top tier for overall inference [2].	Good performance for subclone identification [2].	Good performance for subclone identification [2].
CaSpER [2]	Among the best (with CopyKAT) [2].	Good sensitivity and specificity [2].	Good sensitivity and specificity [2].
CopyKAT [2]	Among the best (with CaSpER) [2].	Good sensitivity and specificity [2].	Good sensitivity and specificity [2].

Note: N/A indicates that specific data for this combination was not detailed in the provided search results.

Experimental Protocols for Key Benchmarking Studies

The conclusions drawn above are based on rigorous experimental designs. The following workflow diagrams and protocols outline the methodologies used in the major benchmarking studies.

Experimental Workflow for scRNA-seq CNV Caller Benchmarking

Description: This workflow, based on the Nature Communications study [3], involved applying six CNV callers to 21 diverse scRNA-seq datasets. Performance was benchmarked against a ground truth from whole-genome or whole-exome sequencing using metrics like correlation, AUC, and F1 score. The study also evaluated performance on a euploid PBMC dataset and the impact of reference dataset selection [3].

Protocol for Multi-Platform and Clinical Validation

Description: This protocol, from the Precision Clinical Medicine study [2], first evaluated tools on mixed cell line samples to assess subclone identification accuracy using metrics like the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI). Findings were then validated using a clinical dataset from a small cell lung cancer (SCLC) patient, with scRNA-seq results corroborated by single-cell whole-exome sequencing (scWES) and bulk whole-genome sequencing (WGS) [2].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table lists key materials and resources used in the benchmarking experiments, which are also essential for conducting robust scCNV analysis.

Item Name	Function/Description	Example/Reference
Reference Euploid Cells	A set of normal cells used for expression normalization in CNV inference. Critical for accuracy.	Matched B-cell line (HCC1395BL) [2]; PBMCs from healthy donor [3].
Orthogonal Validation Data	Data from a different modality (not scRNA-seq) used as a ground truth to validate CNV calls.	single-cell or bulk Whole-Genome Sequencing (scWGS/WGS); Whole-Exome Sequencing (WES) [3] [2].
Benchmarking Pipeline	A computational workflow to systematically test and compare CNV callers on new datasets.	Snakemake pipeline from Colomemaria et al. (https://github.com/colomemaria/benchmarkscrnaseqcnv_callers) [3].
Batch Effect Correction Tool	Software to correct for technical variation between datasets from different platforms or batches.	ComBat [4].
Cell Type Annotation Tools	Methods to classify cell types (e.g., tumor vs. normal) which is necessary for selecting reference cells.	Louvain clustering and marker gene expression [3].

The benchmarking data leads to several clear recommendations for researchers:

• For overall CNV inference where the goal is accurate detection of gains and losses, CaSpER and CopyKAT are the most reliable choices [2] [4].
• For subclone identification within a tumor from a single scRNA-seq platform, InferCNV and CopyKAT deliver superior results [2] [4].
• The selection of reference euploid cells is a critical parameter that significantly influences the results of expression-based methods; careful annotation of normal cell types is essential [1] [3].
• Be cautious of batch effects, which can severely impact subclone identification when integrating datasets from different scRNA-seq platforms. Consider using batch correction tools or allele-frequency-based methods like the allelic version of HoneyBADGER for more robust integration [2] [4].

In conclusion, there is no single "best" tool for all scenarios. The choice of algorithm must be guided by the specific biological question, the scRNA-seq technology used, and the available computational resources. As the field progresses, the development of more accurate and robust algorithms remains a priority.

The study of copy number variations (CNVs) is crucial for understanding cancer evolution, tumor heterogeneity, and therapeutic resistance. While single-cell DNA sequencing represents the gold standard for CNV detection, single-cell RNA sequencing (scRNA-seq) has emerged as a powerful alternative that enables simultaneous analysis of genomic alterations and transcriptional states from the same dataset [3]. This dual-information capability has driven the development of numerous computational methods designed to infer CNVs from transcriptomic data, creating a critical need for comprehensive benchmarking to guide researchers through the rapidly expanding methodological landscape.

Several independent benchmarking studies conducted in 2025 have systematically evaluated these computational tools, revealing dramatic differences in their performance, strengths, and limitations [3] [2] [5]. These evaluations provide critical insights for researchers seeking to select appropriate methods for specific experimental contexts, from basic research to clinical applications. This review synthesizes these recent findings to offer evidence-based guidance for leveraging scRNA-seq data in CNV analysis, with a particular focus on practical implementation considerations for cancer research and drug development.

Performance Comparison of scRNA-seq CNV Callers

Computational methods for inferring CNVs from scRNA-seq data employ diverse algorithmic strategies, which can be broadly categorized into expression-based approaches and allele-frequency-integrated approaches. Expression-based methods (InferCNV, CopyKAT, SCEVAN, CONICSmat, sciCNV) operate on the principle that genes in amplified genomic regions show elevated expression, while those in deleted regions show reduced expression compared to diploid regions [3]. In contrast, allele-frequency methods (CaSpER, Numbat, HoneyBADGER) integrate single nucleotide polymorphism (SNP) information from scRNA-seq reads with expression signals, implementing hidden Markov models to call CNVs [3] [2].

Recent benchmarking studies have evaluated these tools across multiple dimensions, including CNV prediction accuracy, ability to identify euploid cells, subclone detection performance, computational efficiency, and robustness to technical variables. The most comprehensive analysis evaluated six popular methods (InferCNV, CopyKAT, SCEVAN, CONICSmat, CaSpER, Numbat) across 21 scRNA-seq datasets generated from different technologies (droplet-based and plate-based) and organisms (human and mouse) [3]. Another independent study published in June 2025 compared five methods (HoneyBADGER, inferCNV, sciCNV, CaSpER, and CopyKAT) across multiple scRNA-seq platforms and included clinical validation [2].

Quantitative Performance Metrics

Table 1: Performance Metrics of scRNA-seq CNV Callers Based on Independent Benchmarking Studies

Method	Primary Strategy	CNV Calling Accuracy	Subclone Identification	Euploid Detection	Computational Demand
CaSpER	Expression + Allele frequency	High (AUC: 0.72-0.89) [2]	Moderate [2]	Good [3]	High runtime [3]
CopyKAT	Expression-based segmentation	High (AUC: 0.75-0.91) [2]	Excellent [2] [4]	Moderate [3]	Moderate [3]
InferCNV	HMM on expression	Moderate [3]	Excellent [2] [4]	Good [3]	Moderate [3]
SCEVAN	Segmentation-based	Variable across datasets [3]	Good [3]	Moderate [5]	Low-Moderate [3]
Numbat	Expression + Allele frequency	Moderate-High [3]	Good [3]	Good [3]	High runtime [3]
HoneyBADGER	HMM + Bayesian	Moderate [2]	Moderate [2]	Not comprehensively evaluated	Moderate [2]
sciCNV	Expression disparity scoring	Moderate [2]	Good (single platform) [2]	Challenging [5]	Low [2]

Table 2: Performance in Specific Research Contexts Based on Application Studies

Method	Tumor Cell Identification	Rare Population Detection	Batch Effect Sensitivity	Optimal Use Case
CaSpER	High sensitivity [2]	Moderate [2]	Moderate [2]	Large droplet-based datasets [3]
CopyKAT	Moderate sensitivity, overestimates tumor cells [5]	Good [2]	High sensitivity [2]	Subclone identification in homogeneous data [2]
InferCNV	Does not directly predict tumor cells [5]	Excellent with sufficient cells [4]	High sensitivity [2]	Subclone resolution in complex tumors [2]
SCEVAN	Moderate sensitivity, overestimates tumor cells [5]	Good [3]	Moderate [3]	Automated tumor/normal classification [5]
Numbat	Good [3]	Good [3]	Lower sensitivity to batch effects [3]	Datasets with good SNP coverage [3]
HoneyBADGER	Allele-based version more robust [4]	Poor [4]	Low (allele-based) [4]	Multi-platform integrated analysis [2]
sciCNV	Does not directly predict tumor cells [5]	Poor [4]	High sensitivity [2]	Low-computational budget analyses [2]

Performance metrics reveal that no single method outperforms others across all evaluation criteria. CaSpER and CopyKAT consistently demonstrate superior performance in overall CNV inference accuracy, while InferCNV and CopyKAT excel specifically in subclone identification tasks [2] [4]. Methods incorporating allelic information (CaSpER, Numbat) generally perform more robustly for large droplet-based datasets but require higher computational resources [3].

The benchmarking analyses further highlight the significant impact of technical and biological variables on method performance. Specifically, sequencing depth, read length, choice of reference dataset, and tumor purity substantially influence accuracy metrics [3] [2]. For example, a study evaluating CNV identification in endometrial cancer found that SCEVAN and CopyKAT demonstrated moderate sensitivity but significantly overestimated the true number of tumor cells, emphasizing the importance of complementary validation through epithelial marker expression [5].

Experimental Design and Methodologies in Benchmarking Studies

Benchmarking Frameworks and Validation Strategies

The 2025 benchmarking studies employed rigorous experimental designs to evaluate method performance under controlled conditions and real-world scenarios. The most comprehensive assessment utilized 21 scRNA-seq datasets with orthogonal CNV validation through either single-cell or bulk whole-genome sequencing (scWGS/WGS) or whole exome sequencing (WES) [3]. This design enabled direct comparison between computationally inferred CNVs and experimentally determined ground truth across diverse biological contexts, including cancer cell lines, primary tumors, and diploid reference samples.

Evaluation metrics were carefully selected to assess different aspects of performance. Threshold-independent metrics included correlation analysis and area under the curve (AUC) scores, with separate evaluations for gain versus all and loss versus all regions [3]. Additionally, partial AUC values were calculated to focus on biologically meaningful threshold ranges [3]. For classification performance, F1 scores were computed based on optimal gain and loss thresholds identified through systematic testing of biologically meaningful values [3].

Table 3: Key Experimental Resources in scRNA-seq CNV Benchmarking

Resource Type	Specific Examples	Function in Experimental Design
Reference Datasets	PBMCs from healthy donors [3], HCC1395/HCC1395BL cell lines [2]	Provide diploid controls for normalization and baseline establishment
Cell Line Mixtures	5 human lung adenocarcinoma line mixtures [2], Gastric cancer spike-ins [6]	Enable controlled evaluation of subclone detection accuracy
Orthogonal Validation	scWGS, bulk WGS, WES [3], Array CGH [6], Karyotyping [7]	Provide ground truth for CNV calls and method validation
Software Pipelines	Snakemake benchmarking pipeline [3], SCOPE [6], SCYN [6]	Standardize analysis workflows and enable reproducible comparisons
Sequencing Platforms	10x Genomics, Fluidigm C1, ICELL8, Drop-seq, CEL-seq2 [2]	Assess platform-specific performance and technical variability

scRNA-seq CNV Calling Workflow

The following diagram illustrates the general computational workflow for inferring CNVs from scRNA-seq data, as implemented across the benchmarked methods:

This workflow highlights two critical steps that significantly impact performance: reference selection and algorithm selection. The benchmarking studies consistently demonstrated that the choice of reference diploid cells for normalization profoundly influences CNV calling accuracy, with performance varying substantially depending on whether internal or external references were used [3]. For cancer cell lines where matched normal cells are unavailable, the selection of appropriate external reference datasets becomes particularly important [3].

Factors Influencing Method Performance

Technical and Biological Considerations

Benchmarking studies have identified several key factors that significantly impact the performance of scRNA-seq CNV callers:

• Sequencing Depth and Platform: Methods perform differently across scRNA-seq platforms (10x Genomics, Fluidigm C1, ICELL8, Drop-seq) with varying sensitivity to sequencing depth [2] [8]. CaSpER and CopyKAT maintain more consistent performance across platforms, while sciCNV and HoneyBADGER show greater platform-specific variability [2].

• Tumor Purity and Composition: Complex tumor samples with low purity or high stromal contamination present challenges for all methods, though allele-frequency integrated approaches generally show better robustness in these scenarios [3] [9]. Methods vary in their ability to distinguish euploid from aneuploid cells, with several tools overestimating tumor cells in heterogeneous samples [5].

• CNV Size and Complexity: Large chromosomal alterations are more reliably detected than focal CNVs across all methods [3]. The number and type of CNVs in the sample (gains versus losses) also influence detection accuracy, with most methods showing better performance for gained regions [3].

• Batch Effects: When integrating datasets across multiple scRNA-seq platforms, batch effects significantly degrade the performance of most methods for subclone identification unless corrected using tools like ComBat [2]. The allele-based version of HoneyBADGER demonstrates greater resilience to such technical variability [4].

Analytical Considerations for Experimental Design

The following diagram illustrates the key decision points and considerations for designing scRNA-seq CNV studies:

Successful implementation of scRNA-seq CNV analysis requires careful selection of experimental and computational resources. The following table details key reagents and their functions based on the benchmarking studies:

Table 4: Essential Research Reagents and Resources for scRNA-seq CNV Studies

Category	Specific Resource	Function & Importance
Reference Cells	PBMCs from healthy donors [3]	Critical normalization control for identifying aneuploid cells in tumor samples
Validated Cell Lines	HCC1395/HCC1395BL pair [2]	Provide matched tumor-normal system for method validation and optimization
Spike-in Controls	Gastric cancer cell spike-ins [6], Lung adenocarcinoma mixtures [2]	Enable controlled assessment of detection sensitivity and specificity
Orthogonal Validation	scWGS, bulk WGS, Array CGH [3]	Establish ground truth for benchmarking and performance verification
Analysis Pipelines	Snakemake benchmarking pipeline [3], SCYN [6]	Standardize analytical workflows and ensure reproducible results
Computational Infrastructure	High-memory computing nodes [3]	Essential for processing large datasets, especially allele-frequency methods

The comprehensive benchmarking of scRNA-seq CNV callers reveals a complex performance landscape where method selection must be guided by specific research objectives and experimental constraints. For general CNV detection in large droplet-based datasets, CaSpER and CopyKAT emerge as top performers with balanced sensitivity and specificity [3] [2]. When subclone identification is the primary goal, particularly in complex tumor ecosystems, InferCNV and CopyKAT provide superior resolution of cellular subpopulations [2] [4].

The integration of allelic information with expression signals generally enhances robustness, though at increased computational cost [3]. Importantly, all methods show performance dependence on technical factors including sequencing depth, platform selection, and reference quality, underscoring the importance of appropriate experimental design. Future method development should address current limitations in detecting focal CNVs, managing batch effects in multi-platform studies, and improving accessibility for researchers without specialized computational expertise.

As single-cell genomics continues its transition toward clinical applications, accurate CNV detection from scRNA-seq data will play an increasingly important role in cancer diagnostics, biomarker discovery, and therapeutic monitoring. The benchmarking frameworks and performance insights summarized here provide a foundation for selecting optimal analytical strategies while highlighting critical areas for future methodological innovation.

Expression-Based vs. Allele-Frequency Based Approaches

Copy number variations (CNVs) are genomic alterations involving the gain or loss of DNA segments, playing crucial roles in cancer development and tumor heterogeneity [3] [2]. Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool not only for characterizing cellular transcriptomes but also for inferring CNVs, enabling the simultaneous analysis of genetic alterations and their functional consequences in the same cells [3] [4]. Computational methods for CNV detection from scRNA-seq data primarily fall into two methodological categories: those utilizing only gene expression levels and those incorporating allelic imbalance information [3]. This guide provides a comprehensive comparison of these approaches, supported by recent benchmarking studies, to assist researchers in selecting optimal tools for their specific applications.

Methodological Categories and Representative Tools

Core Computational Principles

Expression-based approaches operate on the fundamental assumption that genes located in genomically amplified regions exhibit higher expression levels, while those in deleted regions show reduced expression compared to diploid regions [3]. These methods employ sophisticated normalization strategies using reference diploid cells, followed by various computational techniques to infer CNV patterns from expression outliers.

Allele-frequency based approaches integrate both gene expression data and single nucleotide polymorphism (SNP) information called from scRNA-seq reads [3]. These methods leverage minor allele frequency (AF) patterns that deviate from expected heterozygous distributions in regions with copy number alterations, providing an orthogonal signal to validate expression-based inferences.

Table 1: Classification of Single-Cell CNV Detection Methods

Method Category	Representative Tools	Core Algorithm	Primary Output
Expression-Based	InferCNV [3] [2], CopyKAT [3] [2], SCEVAN [3], CONICSmat [3]	Hidden Markov Models [3], Segmentation approaches [3], Mixture Models [3]	Discrete CNV calls or normalized expression scores [3]
Allele-Frequency Integrated	CaSpER [3] [2], Numbat [3], HoneyBADGER [2] [4]	Hidden Markov Models integrating expression and allele frequency [3] [2]	CNV predictions with allelic imbalance support

The following diagram illustrates the fundamental workflows for these two methodological approaches:

Performance Benchmarking and Experimental Data

Experimental Design in Benchmarking Studies

Recent independent benchmarking studies have employed rigorous experimental designs to evaluate the performance of single-cell CNV detection methods. The primary evaluation schemes include:

• Sensitivity and specificity analysis using scRNA-seq datasets from cancer cell lines with matched normal B-cell lines from the same donor, generated across multiple scRNA-seq platforms (10x Genomics, Fluidigm C1, C1 HT, and ICELL8) [2] [4].

• Subclone identification accuracy assessed using mixed scRNA-seq datasets comprising three or five human lung adenocarcinoma cell lines with known genetic profiles, mimicking tumor subpopulations [2] [4].

• Clinical validation employing scRNA-seq data from 92 primary and 39 relapse small cell lung cancer cells, with orthogonal validation using single-cell whole exome sequencing (scWES) and bulk whole genome sequencing (WGS) [2] [4].

These benchmarking studies evaluated performance using multiple metrics including sensitivity, specificity, area under the curve (AUC), F1 scores, and clustering accuracy metrics such as Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) [3] [2].

Table 2: Quantitative Performance Comparison Across Methodological Categories

Performance Metric	Top Performing Tools	Performance Characteristics	Method Category
Overall CNV Detection	CaSpER [2] [4], CopyKAT [2] [4]	Balanced sensitivity and specificity across platforms [2]	Allele-Frequency Integrated & Expression-Based
Subclone Identification	InferCNV [2] [4], CopyKAT [2] [4]	High accuracy in distinguishing tumor subpopulations [2]	Expression-Based
Rare Population Detection	InferCNV [4] [10]	Strong sensitivity with sufficient cell numbers [4] [10]	Expression-Based
Batch Effect Resilience	HoneyBADGER (allele-based) [4] [10]	More resilient to technical batch effects [4] [10]	Allele-Frequency Integrated
Runtime Efficiency	Expression-based methods (generally) [3]	Lower computational requirements [3]	Expression-Based

Impact of Experimental Factors

Benchmarking studies have identified several critical factors that significantly impact method performance:

• Reference dataset selection: The choice of euploid reference cells for normalization substantially affects CNV calling accuracy, particularly for expression-based methods [3].

• Sequencing depth and platform: Allele-frequency integrated methods generally require higher sequencing depths for reliable SNP calling, while expression-based methods show variable performance across different scRNA-seq platforms [2].

• Dataset size: Methods incorporating allelic information perform more robustly for large droplet-based datasets but require higher computational resources [3].

• Batch effects: When combining datasets across multiple platforms, batch effects significantly impact most methods, particularly expression-based approaches, unless corrected using specialized tools like ComBat [2].

Experimental Protocols for Method Evaluation

Standardized Benchmarking Pipeline

The benchmarking study published in Nature Communications provides a reproducible Snakemake pipeline for evaluating scRNA-seq CNV callers (https://github.com/colomemaria/benchmarkscrnaseqcnv_callers) [3]. The key methodological steps include:

• Data Preprocessing: Raw scRNA-seq data from 21 different datasets (13 human cancer cell lines, 6 human primary tumor samples, 1 mouse primary tumor, and 1 human diploid dataset) were processed using consistent quality control metrics [3].

• Ground Truth Establishment: Orthogonal CNV measurements from (sc)WGS or WES data were used to establish validation sets. For plate-based datasets where scRNA-seq and scWGS were measured in the same cells, cell-by-cell comparison was performed [3].

• Pseudobulk Analysis: For most datasets where ground truth was not measured in the same cells as scRNA-seq, per-cell results were combined into an average CNV profile (pseudobulk) before comparison with validation data [3].

• Metric Calculation: Threshold-independent evaluation metrics included correlation analysis, area under the curve (AUC) scores, and partial AUC values with biologically meaningful thresholds. Sensitivity and specificity values for gains and losses were calculated using optimal thresholds determined by multi-class F1 scores [3].

Evaluation Metrics and Statistical Analysis

The following diagram illustrates the comprehensive evaluation workflow used in benchmarking studies:

Performance evaluation included both threshold-dependent and threshold-independent metrics. For AUC calculations, predictions were evaluated separately for gain versus all and loss versus all, resulting in two scores. Partial AUC values were calculated with maximal sensitivity defined by baseline scores to focus on biologically meaningful thresholds [3].

Table 3: Key Research Reagents and Computational Resources for scCNV Analysis

Resource Category	Specific Items	Function/Purpose	Example Sources/References
Reference Cell Lines	HCC1395/HCC1395BL (breast cancer/B-lymphocyte) [2]	Paired tumor-normal model for method validation	Coriell Institute [2]
	25 Coriell cell lines with known CNVs [11]	Validation set for CNV detection performance	Coriell Institute [11]
scRNA-seq Platforms	10x Genomics, Fluidigm C1, ICELL8, C1 HT [2]	Generate scRNA-seq data for CNV inference	Multiple manufacturers [2]
Orthogonal Validation	scWGS, bulk WGS, WES [3] [2]	Establish ground truth for benchmarking	Various platforms [3] [2]
Benchmark Datasets	21 scRNA-seq datasets [3]	Comprehensive performance evaluation	Public repositories [3]
	Mixed lung adenocarcinoma cell lines [2] [4]	Subclone identification assessment	Tian et al. dataset (GSE118767) [2]
Computational Resources	High-performance computing clusters	Handle memory-intensive algorithms (especially allele-based methods)	Institutional resources [3]
	Reproducible workflow tools	Snakemake pipeline for standardized benchmarking [3]	https://github.com/colomemaria/benchmarkscrnaseqcnv_callers [3]

The comparative analysis of expression-based and allele-frequency integrated approaches for single-cell CNV detection reveals distinct strengths and limitations for each methodological category. Expression-based methods generally offer computational efficiency and robust performance in subclone identification, while allele-frequency integrated approaches provide enhanced robustness in large datasets and resilience to certain technical artifacts at the cost of higher computational requirements [3] [2].

Selection of the optimal approach should consider specific research goals, experimental design, and computational resources. For studies focusing primarily on subpopulation identification in datasets from single platforms, expression-based methods like InferCNV and CopyKAT offer excellent performance [2] [4]. For large-scale studies integrating multiple datasets or requiring high confidence in CNV calls, allele-frequency integrated methods like CaSpER may be preferable despite their computational intensity [3] [2].

Future methodological developments will likely focus on hybrid approaches that optimally leverage both expression and allele frequency signals while addressing current limitations in computational efficiency and batch effect sensitivity.

The detection of Copy Number Variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data has emerged as a powerful, indirect approach to study genetic heterogeneity in complex tissues, most notably cancer. Several computational methods have been developed to infer CNVs from transcriptomic data, operating on the core assumption that genes in amplified genomic regions show elevated expression, while those in deleted regions show reduced expression [3]. However, the path from raw scRNA-seq data to accurate CNV profiles is fraught with technical challenges that directly impact the reliability and biological interpretability of the results. Independent benchmarking studies have become essential for guiding researchers through the complex landscape of method selection and application [3] [12] [2].

This guide objectively compares the performance of leading scRNA-seq CNV callers, focusing on three interconnected technical hurdles: data normalization, technical noise, and resolution limitations. By synthesizing evidence from recent, large-scale benchmarking efforts, we provide a data-driven framework for selecting the optimal CNV inference method based on specific experimental conditions and research goals. The insights presented here are critical for researchers, scientists, and drug development professionals seeking to leverage scRNA-seq data for genomic studies.

Performance Benchmarking of scRNA-seq CNV Callers

The computational methods for inferring CNVs from scRNA-seq data can be broadly categorized into those using only gene expression information and those that integrate expression with allelic imbalance data from single nucleotide polymorphisms (SNPs) [3]. Each method employs distinct strategies for normalization, noise reduction, and segmentation.

Table 1: Classification and Key Characteristics of scRNA-seq CNV Callers

Method	Core Data Input	Primary Algorithm	Reported Resolution	Output Type
InferCNV	Gene Expression	Hidden Markov Model (HMM)	Per gene or segment	Subclone groups [3]
CopyKAT	Gene Expression	Segmentation	Per gene or segment	Per cell [3]
SCEVAN	Gene Expression	Segmentation	Per gene or segment	Subclone groups [3]
CONICSmat	Gene Expression	Mixture Model	Per chromosome arm	Per cell [3]
CaSpER	Expression + Allelic Frequency	HMM + Multiscale Smoothing	Per gene or segment	Per cell [3] [2]
Numbat	Expression + Allelic Frequency	Hidden Markov Model (HMM)	Per gene or segment	Subclone groups [3]
HoneyBADGER	Expression + Allelic Frequency	Bayesian HMM	Per gene or segment	Subclone groups [2]
sciCNV	Gene Expression	Expression Disparity Score	Per gene or segment	Subclone groups [2]

Comparative Performance Across Benchmarking Studies

Recent independent evaluations have revealed significant performance variations among methods, influenced by dataset-specific factors including technology platform, sequencing depth, and the choice of reference cells.

Table 2: Quantitative Performance Summary from Benchmarking Studies

Method	Overall CNV Accuracy	Subclone Identification	Performance with Batch Effects	Sensitivity on Rare Populations	Key Strengths
CaSpER	Top performer [13] [2]	Moderate	Highly affected in multi-platform data [2]	Information Missing	Robust for large droplet-based datasets [3]
CopyKAT	Top performer [12] [13] [2]	Excellent [12] [2]	Highly affected in multi-platform data [2]	Information Missing	Accurate subpopulation characterization [2]
InferCNV	Moderate	Excellent [12] [13] [2]	Highly affected in multi-platform data [2]	Strong with sufficient cells [13]	Identifies tumor subclones effectively [13]
sciCNV	Lower than top performers	Good [2]	Highly affected in multi-platform data [2]	Falls short [13]	Effective on single-platform data [13]
HoneyBADGER	Lower than top performers	Falls short [13]	More resilient (allele-based) [13] [2]	Falls short [13]	Resilient to batch effects [13]

Experimental Protocols in Benchmarking Studies

Standardized Evaluation Framework

The benchmarking methodologies employed in recent studies provide a template for rigorous CNV caller validation. Understanding these protocols is essential for interpreting performance claims and designing new evaluations.

Detailed Methodological Components

Dataset Selection and Ground Truth Establishment

Benchmarking studies utilized diverse datasets representing various biological contexts and technical platforms. The 2025 Nature Communications benchmarking included 21 scRNA-seq datasets encompassing human cancer cell lines (gastric, colorectal, breast, melanoma), primary tumor samples (leukemia, basal cell carcinoma, multiple myeloma), and diploid controls (PBMCs) [3]. Technologies included both droplet-based (17 datasets) and plate-based (4 datasets) platforms [3].

Ground truth CNV profiles were established using orthogonal genomic measurements, including single-cell or bulk whole-genome sequencing ((sc)WGS) and whole exome sequencing (WES) [3]. For cell line mixtures, the known proportions of different lines provided a truth standard for subclone identification accuracy [2]. This approach enabled both pseudobulk comparisons (averaging CNV profiles across cells) and, for plate-based datasets where scRNA-seq and scWGS were measured in the same cells, direct cell-by-cell comparisons [3].

Method Application and Parameterization

All evaluated methods were run according to their respective tutorials or with default parameters as specified in the benchmarking studies [3]. A critical standardized aspect was the selection of reference cells for normalization. To ensure fair comparison, normal (diploid) cells were manually annotated for each sample using Louvain clustering and known marker genes, with the same healthy cells used as reference across all methods [3]. For cancer cell lines where no directly matched reference exists, external reference datasets with healthy cells from similar cell types were selected [3].

Performance Quantification

Multiple complementary metrics were employed to evaluate different aspects of CNV calling performance:

• Threshold-independent metrics: Correlation and Area Under the Curve (AUC) scores evaluated the agreement between inferred and ground truth CNVs, with separate analyses for gains versus all and losses versus all [3]. Partial AUC values were calculated to focus on biologically meaningful threshold ranges [3].

• Classification accuracy: Optimal gain and loss thresholds were determined using a multi-class F1 score, from which sensitivity and specificity values were derived [3].

• Subclone identification: Metrics including Adjusted Rand Index (ARI), Fowlkes-Mallows index (FM), Normalized Mutual Information (NMI), and V-Measure quantified the accuracy of cellular subpopulation identification compared to known cell line identities in mixed samples [2].

• Computational efficiency: Runtime and memory requirements were assessed to evaluate practical scalability [3].

Critical Technical Hurdles in Detail

The Normalization Challenge

Normalization presents a fundamental hurdle for scRNA-seq CNV detection, as methods must distinguish true copy-number-driven expression changes from overwhelming technical and biological confounding factors.

The normalization challenge is particularly acute because global-scaling methods inherited from bulk RNA-seq analysis make assumptions that are frequently violated in single-cell data [14]. These methods assume that the expected read count for a gene in a cell is proportional to a gene-specific expression level and a cell-specific scaling factor representing technical effects [14]. However, scRNA-seq data exhibits unique features including high sparsity (zero-inflation) and substantial technical noise that complicate this relationship [14] [15].

The choice of reference cells for normalization profoundly impacts results. Methods that include allelic information (CaSpER, Numbat) generally perform more robustly for large droplet-based datasets, potentially because allele frequency data provides an orthogonal signal that is less dependent on reference selection [3]. Benchmarking revealed that dataset-specific factors including dataset size, the number and type of CNVs in the sample, and reference dataset choice significantly influence performance [3].

Technical Noise and Data Sparsity

The high levels of technical noise and dropout characteristics of scRNA-seq data directly challenge CNV detection sensitivity and specificity. The sparsity of scRNA-seq data—manifesting as a high proportion of zero read counts—arises from both biological reasons (genuine lack of expression in subpopulations) and technical reasons (dropouts where expressed genes fail to be detected) [14].

The benchmarking studies found that performance variations between methods were significantly influenced by sequencing depth and read length [12] [2]. Methods incorporating allelic information generally require higher runtime but demonstrate improved robustness to technical noise in larger datasets [3]. The ability to distinguish signal from noise also depends on the underlying algorithm, with HMM-based approaches (InferCNV, Numbat) and segmentation methods (CopyKAT, SCEVAN) employing different denoising strategies [3].

Batch effects represent a particularly pernicious form of technical noise. When combining datasets across different scRNA-seq platforms, most expression-based CNV inference methods (InferCNV, CaSpER, sciCNV, CopyKAT) were highly affected in their subpopulation identification accuracy [2]. Only the allele-based version of HoneyBADGER demonstrated relative resilience to these batch-related distortions [13] [2].

Resolution Limitations

The resolution of CNV detection—both in terms of genomic scale and cellular minority populations—represents a third critical hurdle. Resolution limitations manifest in several dimensions:

• Genomic resolution: Methods vary in their reporting resolution, from CONICSmat (chromosome arm level) to other methods that report per gene or across segments of multiple genes [3]. Focal CNVs affecting small genomic regions are particularly challenging to detect reliably from scRNA-seq data due to the sparse nature of gene coverage.

• Cellular resolution: The ability to identify rare tumor subpopulations varies significantly between methods. InferCNV showed strong sensitivity for detecting rare populations when sufficient cells were sequenced, while sciCNV and HoneyBADGER fell short in this regard [13].

• Ploidy estimation: Methods struggle with accurate ploidy determination, particularly in highly aneuploid samples. The benchmarking revealed that no method consistently outperformed others across all datasets, with performance being highly context-dependent [3].

Table 3: Key Research Reagents and Computational Tools for scRNA-seq CNV Analysis

Resource Category	Specific Item	Function/Role in CNV Analysis
Experimental Platforms	10X Genomics Chromium	Droplet-based scRNA-seq platform for high-throughput cell capture [15]
	Fluidigm C1	Automated microfluidic system for plate-based scRNA-seq [15] [2]
	SMART-seq2/3	Full-length transcript protocol for higher sensitivity [15]
Reference Data	Human PBMC scRNA-seq	Common source of normal reference cells for blood-derived samples [3]
	Cell line atlases (e.g., HCC1395BL)	Matched "normal" control cell lines for benchmarking [2]
	External diploid references	Healthy cells from similar tissues for normalizing cell line data [3]
Bioinformatics Tools	InferCNV	Widely-used HMM-based method for subclone identification [3] [12]
	CopyKAT	High-performance method for tumor subpopulation characterization [12] [2]
	CaSpER	Integrates expression and allele frequency for robust calling [3] [2]
	Seurat/Scanpy	Standard scRNA-seq preprocessing and cell type annotation [3]
Validation Methods	scWGS (single-cell Whole Genome Sequencing)	Gold-standard orthogonal validation for CNV profiles [3]
	Bulk WES/WGS	Ground truth establishment for pseudobulk comparisons [3] [2]
	Chromosomal Microarray	Traditional CNV detection for validation [7]
Benchmarking Resources	Benchmarking pipeline [3]	Snakemake workflow for method comparison on new datasets [3]
	Mixed cell line datasets	Controlled samples with known proportions for accuracy assessment [2]

The comprehensive benchmarking of scRNA-seq CNV callers reveals that method selection involves navigating critical trade-offs across the three technical hurdles of normalization, noise, and resolution. Based on the consolidated findings:

• For large droplet-based datasets, CaSpER and CopyKAT generally provide the most balanced performance, with CaSpER particularly benefiting from its integration of allelic information for normalization robustness [3] [13] [2].

• For precise subclone identification in data from a single platform, InferCNV and CopyKAT deliver superior performance, making them ideal for studying tumor heterogeneity [12] [13] [2].

• When analyzing datasets combined across multiple platforms, researchers should anticipate significant batch effects on most expression-based methods and consider allele-based approaches or implement batch correction strategies [2].

• For euploid samples or studies requiring null detection, careful validation is essential, as methods vary in their ability to correctly identify the absence of CNVs [3].

The field continues to evolve with new methods like msCNVS [7] and SCOPE [16] emerging, though these were not included in the comprehensive benchmarks discussed here. Researchers should consult the benchmarking pipeline provided by Colomé-Tatché et al. [3] to determine the optimal method for their specific datasets and biological questions. As single-cell genomics progresses toward clinical applications, addressing these technical hurdles will be paramount for reliable biomarker discovery and therapeutic monitoring.

A Practical Guide to scRNA-seq CNV Callers: From Theory to Implementation

Copy number variations (CNVs), defined as the gain or loss of genomic regions, are fundamental drivers of disease, particularly in cancer, where they contribute to tumor initiation, progression, and therapeutic resistance [3]. The inherent heterogeneity of tumors means that distinct cellular subclones with unique CNV profiles coexist within the same sample, complicating treatment and influencing clinical outcomes. Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful technology that enables researchers to dissect this complexity by capturing gene expression at the individual cell level. Computational methods that infer CNVs from scRNA-seq data leverage the principle that genes within amplified genomic regions tend to show elevated expression, while those in deleted regions show reduced expression compared to diploid regions.

Several tools have been developed to decode CNV signals from transcriptomic data, allowing scientists to study genetic and functional heterogeneity simultaneously from a single assay [3] [17]. However, these methods employ distinct algorithmic strategies, have different input requirements, and demonstrate variable performance across diverse datasets. This guide provides an objective, data-driven comparison of six prominent tools—InferCNV, CopyKAT, CaSpER, Numbat, SCEVAN, and CONICSmat—framed within the context of comprehensive benchmarking studies. By synthesizing empirical evidence from large-scale evaluations, we aim to equip researchers, scientists, and drug development professionals with the insights needed to select the optimal CNV detection tool for their specific experimental context and biological questions.

Methodologies of scRNA-seq CNV Callers

The computational tools for inferring CNVs from scRNA-seq data can be broadly classified into two categories based on their input data and underlying algorithms.

Algorithmic Approaches and Input Requirements

• Expression-Based Methods: This category includes tools that rely solely on gene expression data. They operate on the core assumption that copy number amplifications lead to upregulated gene expression, while deletions result in downregulation within the affected genomic regions.

  • InferCNV utilizes a corrected moving average over gene windows and employs a Hidden Markov Model (HMM) for CNV calling [3] [17].
  • CopyKAT applies an integrative Bayesian segmentation approach to infer CNV profiles [17] [4].
  • SCEVAN implements a multi-channel segmentation algorithm designed to distinguish tumor cells from normal cells and identify clonal subpopulations [17].
  • CONICSmat estimates CNVs based on a Mixture Model and reports results per chromosome arm [3].

• Expression + Allelic Information Methods: These more advanced tools integrate gene expression data with allelic imbalance information derived from single nucleotide polymorphisms (SNPs).

  • Numbat employs a haplotype-aware Hidden Markov Model that integrates signals from gene expression, allelic ratio from scRNA-seq reads, and population-derived haplotypes to infer allele-specific CNAs and copy-number neutral loss of heterozygosity (cnLOH) [17] [3].
  • CaSpER utilizes a multiscale signal-processing framework that integrates gene expression and allelic shift signal profiles for CNV calling [17] [3].

Benchmarking Experimental Designs

Independent benchmarking studies have evaluated these tools using rigorous frameworks to assess their real-world performance. The key aspects of these experimental designs include:

• Ground Truth Validation: Performance was assessed by comparing scRNA-seq CNV predictions to orthogonal CNV measurements obtained from either (single-cell) whole-genome sequencing ((sc)WGS) or whole-exome sequencing (WES) data [3] [17]. In some studies, single-cell multi-omics datasets enabling simultaneous interrogation of DNA and RNA within the same cell provided the most accurate validation [17].

• Diverse Dataset Composition: Benchmarking studies utilized multiple scRNA-seq datasets spanning various contexts, including:

  • Human cancer cell lines (e.g., gastric, colorectal, breast, melanoma)
  • Primary tumor samples (e.g., acute lymphoblastic leukemia, basal cell carcinoma, multiple myeloma, colorectal cancer, glioma)
  • Diploid control samples (e.g., peripheral blood mononuclear cells - PBMCs)
  • Different sequencing technologies (droplet-based and plate-based) [3] [17]

• Performance Metrics: Multiple quantitative metrics were employed to evaluate different aspects of performance:

  • CNV Prediction Accuracy: Correlation, area under the curve (AUC), partial AUC, and F1 scores for gains and losses [3]
  • Tumor/Normal Classification: Sensitivity, specificity, and F1 scores for distinguishing malignant from normal cells [17]
  • Subclone Identification: Accuracy in reconstructing clonal architecture [3] [4]
  • Computational Efficiency: Runtime and memory requirements [3] [18]

The following diagram illustrates a typical benchmarking workflow for evaluating scRNA-seq CNV callers:

Performance Comparison Across Benchmarking Studies

Quantitative Performance Metrics

Independent benchmarking studies have systematically evaluated CNV callers across multiple dimensions. The table below summarizes key performance metrics from these comprehensive assessments:

Table 1: Comprehensive Performance Comparison of scRNA-seq CNV Callers

Tool	Algorithm Type	Tumor/Normal Classification F1 Score	CNV Profile Accuracy	Subclone Identification	Aneuploidy Detection in Normal Cells	Runtime & Memory
Numbat	Expression + Allelic	0.95-0.99 [17]	High [17]	Good [3]	High sensitivity [17]	High runtime [3] [18]
CopyKAT	Expression	0.80-0.90 [17]	High [4] [12]	Excellent [4] [12]	Moderate [17]	Fast, low memory [3] [18]
CaSpER	Expression + Allelic	0.75-0.85 [17]	High [4] [12]	Moderate [3]	Moderate [17]	High runtime [3]
InferCNV	Expression	0.70-0.85 [17]	Variable [3]	Good [4]	Low [17]	Moderate to high runtime [3]
SCEVAN	Expression	0.65-0.80 [17]	Moderate [3]	Good [3]	Best for breakpoint detection [17]	Fast [18]
CONICSmat	Expression	Not reported	Low resolution [3]	Poor [3]	Not reported	Fastest, low memory [3] [18]

Tool Performance Across Different Applications

The benchmarking studies reveal that each tool has distinct strengths depending on the specific analytical task:

• For Overall CNV Inference: CaSpER and CopyKAT consistently delivered the most balanced CNV inference results across multiple datasets and sequencing platforms [4] [12]. However, their effectiveness varied with sequencing depth and platform type.

• For Tumor/Normal Cell Classification: Numbat demonstrated superior performance in distinguishing tumor cells from normal cells, achieving F1 scores of 0.95-0.99 across multiple solid tumor types [17]. Among expression-only methods, CopyKAT achieved the best classification performance.

• For Subclone Identification: InferCNV and CopyKAT excelled in identifying tumor subpopulations, particularly when analyzing data from a single platform [4]. SCEVAN showed the best performance in clonal breakpoint detection [17].

• For Specialized Detection: Numbat showed high sensitivity in detecting copy-number neutral loss of heterozygosity (cnLOH) [17], while SCEVAN performed well in identifying aneuploidy in non-malignant cells within the tumor microenvironment [17].

Experimental Factors Influencing Tool Performance

Critical Experimental Considerations

Benchmarking studies have identified several experimental factors that significantly impact the performance of scRNA-seq CNV callers:

• Reference Dataset Selection: All methods require a set of euploid reference cells for normalization. The choice of reference significantly affects performance [3] [18]. When using T-cells from the same dataset as reference, good performance was observed for all methods. However, when using external references (e.g., Monocytes or T-cells from another dataset), Numbat and CaSpER outperformed other methods, likely due to their incorporation of allelic information [18].

• Tumor Purity: The ratio of tumor to normal cells in the sample dramatically affects performance. Numbat consistently outperformed other tools across a wide range of tumor/normal cell ratios (from 1:100 to 100:1), while InferCNV exhibited errors in classification when tumor purity was high, sometimes incorrectly centering CN gains and losses in tumor cells as the baseline [17].

• Sequencing Depth: As sequencing depth decreases, the overall classification accuracy for all tools drops significantly. One study showed that when median unique molecular identifiers (UMIs) per cell were down-sampled to ~10k, 3k, and 1k, all tools showed reduced F1 scores, with Numbat experiencing the most pronounced drop [17].

• Inclusion of Tumor Microenvironment (TME) Cells: For samples with imbalanced tumor versus normal ratios, including TME cells (immune, endothelial, and fibroblast cells) significantly improved the accuracy of tumor cell prediction for SCEVAN and InferCNV [17].

The following diagram illustrates how these factors influence the CNV calling workflow and results:

Table 2: Key Research Reagent Solutions for scRNA-seq CNV Analysis

Resource Category	Specific Examples	Function in CNV Analysis
Reference Datasets	Healthy PBMCs, matched normal tissue samples	Provides euploid baseline for normalization of gene expression signals [3] [18]
Validation Technologies	(sc)WGS, WES, single-cell multi-omics	Generates ground truth data for benchmarking CNV predictions [3] [17]
Platform-Specific Kits	10x Genomics Chromium, SMART-seq2	Produces scRNA-seq data with characteristics that affect tool performance [4] [19]
Computational Infrastructure	High-performance computing clusters	Enables running computationally intensive tools like Numbat and CaSpER [3] [18]
Bioinformatics Pipelines	Snakemake workflow, Conda environments	Ensures reproducible installation and execution of CNV callers [18]

Based on the comprehensive benchmarking evidence, we recommend:

• For most applications with available allelic information: Numbat demonstrates the best overall performance across multiple evaluation criteria, including tumor/normal classification and CNV profile accuracy [17].

• When only expression matrix is available: CopyKAT is recommended, as it outperforms other expression-only methods in overall CNV inference and subclone identification [17] [4] [12].

• For specific applications:

  • For clonal breakpoint detection: SCEVAN shows superior performance [17]
  • For copy-number neutral LOH detection: Numbat provides high sensitivity [17]
  • For rapid analysis of large datasets: CONICSmat and CopyKAT offer fast runtime with minimal memory requirements [3] [18]

• Critical implementation considerations:

  • Use matched reference cells from the same sample when possible [18]
  • Include tumor microenvironment cells for analysis of samples with high tumor purity [17]
  • Ensure adequate sequencing depth (median UMIs >10k per cell) for optimal performance [17]
  • Account for batch effects when combining datasets across different platforms [4] [12]

No single scRNA-seq CNV caller outperforms all others in every scenario. The choice of tool should be guided by the specific research question, data characteristics, and analytical requirements. As the field evolves, we anticipate that continued benchmarking efforts will further refine these recommendations and drive improvements in computational methods for detecting copy number variations from single-cell transcriptomic data.

The characterization of copy number variations (CNVs) at single-cell resolution is crucial for deciphering tumor heterogeneity, identifying rare subclones, and understanding cancer evolution. Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for inferring CNVs, allowing researchers to connect genetic alterations with transcriptional phenotypes from the same dataset. Several computational methods have been developed to tackle the challenge of inferring CNVs from scRNA-seq data, employing distinct algorithmic approaches including Hidden Markov Models (HMMs), segmentation techniques, and mixture models [3].

These methods operate on the fundamental principle that genes located in genomic regions with copy number gains tend to show higher expression levels, while those in deleted regions show lower expression compared to diploid regions. However, the indirect nature of this inference requires sophisticated normalization strategies and robust statistical models to distinguish true CNV signals from technical noise and biological variation [3]. This review provides a comprehensive comparison of the leading scCNV detection methods, focusing on their underlying algorithms, performance characteristics, and optimal use cases based on recent benchmarking studies.

Algorithmic Approaches and Method Classification

Algorithm Categories and Representatives

Single-cell CNV detection methods can be broadly categorized by their underlying computational frameworks. Hidden Markov Models (HMMs) are probabilistic models that treat the genome as a sequence of hidden states (copy number states) with probabilistic transitions between them. Segmentation approaches partition the genome into non-overlapping segments with homogeneous copy number profiles. Mixture models assume the data is generated from a mixture of probability distributions, each representing a distinct cell subpopulation or copy number state [3] [20].

Table 1: Classification of Single-Cell CNV Detection Methods by Algorithmic Approach

Algorithmic Approach	Representative Methods	Core Methodology	Key Advantages
Hidden Markov Models (HMMs)	InferCNV, CaSpER, Numbat	Uses probabilistic transitions between hidden states (CNV states) across the genome	Robust to noise, models spatial dependencies along genome
Segmentation Approaches	copyKat, SCEVAN, SCYN	Partitions genome into segments with homogeneous CNV profiles using change-point detection	Efficient for large datasets, identifies clear breakpoints
Mixture Models	CONICSmat, CopyMix	Models data as mixture of distributions representing cell subpopulations	Simultaneously clusters cells and infers CNV profiles

Technical Implementation Details

Methods also differ in their data requirements and output resolutions. Some tools, like CaSpER and Numbat, incorporate allelic information from single nucleotide variants (SNPs) in addition to expression data, which can improve accuracy but requires higher sequencing depth [3]. Output resolution varies from chromosome-arm level (CONICSmat) to gene-level or segmented regions (InferCNV, copyKat, SCEVAN) [3]. Some methods provide discrete CNV calls, while others output continuous scores that require thresholding for interpretation [3].

Performance Benchmarking and Comparative Analysis

Comprehensive Performance Evaluation

Recent large-scale benchmarking studies have systematically evaluated the performance of scCNV detection methods across diverse datasets. A study published in Nature Communications assessed six popular methods on 21 scRNA-seq datasets using ground truth CNV measurements from orthogonal techniques such as single-cell or bulk whole-genome sequencing ((sc)WGS) or whole-exome sequencing (WES) [3]. Performance was evaluated using metrics including correlation with ground truth, area under the curve (AUC) values for gain versus all and loss versus all classifications, and F1 scores [3].

Table 2: Comprehensive Performance Comparison of scCNV Detection Methods

Method	Algorithm Type	Overall CNV Accuracy	Subclone Identification	Runtime Efficiency	Key Strengths	Optimal Use Cases
CaSpER	HMM + Allelic Info	High [2] [13]	Moderate [2]	Moderate [3]	Integrates expression and allele frequency	Large droplet-based datasets
CopyKAT	Segmentation	High [2] [13]	High [2] [13]	High [3]	Robust subclone identification	Tumor subpopulation detection
InferCNV	HMM	Moderate [2]	High [2] [13]	Low [3]	Flexible HMM framework	Single-platform studies
SCEVAN	Segmentation	Variable [5]	Variable [5]	Moderate	Automatic malignant cell detection	Datasets with clear normal reference
Numbat	HMM + Allelic Info	Moderate [3]	High [3]	Low [3]	Allele-aware modeling	Datasets with sufficient SNP information
SCICNV	Not Specified	Moderate [2]	Moderate [2]	Moderate	Expression disparity scoring	Full-length transcript datasets

Platform-Specific and Context-Dependent Performance

Method performance shows significant dependence on sequencing platform and data characteristics. Methods incorporating allelic information (CaSpER, Numbat) generally perform more robustly on large droplet-based datasets but require higher computational resources [3]. A 2025 benchmarking study found that CaSpER and CopyKAT delivered the most balanced CNV inference results across platforms, though their effectiveness varied with sequencing depth and platform type [2] [13]. For subclone identification, inferCNV and CopyKAT excelled with data from a single platform [2] [4].

Batch effects significantly impact most methods when combining datasets across different scRNA-seq platforms. Unless corrected using tools like ComBat, batch effects can severely degrade performance, though the allele-based version of HoneyBADGER has shown more resilience to such technical variations [2]. For detecting rare tumor populations, inferCNV demonstrates strong sensitivity, particularly with sufficient cell numbers, while sciCNV and HoneyBADGER generally fall short in this application [13].

Experimental Design and Methodologies

Benchmarking Frameworks and Validation Strategies

Robust benchmarking of scCNV detection methods requires careful experimental design incorporating orthogonal validation. The Nature Communications study utilized 21 scRNA-seq datasets comprising 13 human cancer cell lines, six human primary tumor samples, one mouse primary tumor sample, and one human diploid dataset (PBMCs) [3]. Seventeen datasets were generated with droplet-based technologies and four with plate-based technology [3]. Ground truth CNV profiles were obtained from either (sc)WGS or WES data [3].

To enable comparison between scRNA-seq inferences and ground truth, the per-cell results from scRNA-seq methods were combined to create an average CNV profile (pseudobulk) before comparison [3]. For plate-based datasets where scRNA-seq and scWGS were measured in the same cells, direct cell-by-cell comparison was performed [3]. Threshold-independent evaluation metrics included correlation analysis and AUC scores, with separate evaluations for gain versus all and loss versus all classifications [3].

Diagram 1: Benchmarking workflow for scCNV detection methods. The process begins with scRNA-seq data and reference cells, progresses through CNV prediction using different algorithms, and concludes with performance evaluation against orthogonal ground truth data.

Critical Experimental Considerations

The choice of reference euploid cells significantly impacts method performance. For primary tissue samples, the common assumption is that tissues contain mixtures of tumor and normal cells, with the latter serving as reference [3]. Some methods require user-provided cell type annotations to specify reference cells, while others offer automatic detection of normal cells [3]. For cancer cell lines where no directly matched reference cells exist, researchers must select matched external reference datasets with healthy cells from similar cell types [3].

Performance evaluation requires careful metric selection. The Nature Communications study used partial AUC values with biologically meaningful thresholds to account for method-specific baseline scores [3]. Sensitivity and specificity values for gains and losses were obtained using optimal thresholds determined via multi-class F1 score optimization [3]. For subclone identification, studies often use metrics such as Adjusted Rand Index (ARI), Fowlkes-Mallows index (FM), Normalized Mutual Information (NMI), and V-Measure to compare estimated tumor subpopulations against known cell line identities [2].

Research Reagent Solutions and Computational Tools

Successful implementation of scCNV detection methods requires specific computational tools and resources. The following table outlines key solutions used in benchmarking studies and their functions in the analysis workflow.

Table 3: Essential Research Reagents and Computational Tools for scCNV Analysis

Tool/Resource	Function	Application Context
Reference Genome	Provides genomic coordinate system	Essential for all methods for gene positioning
Orthogonal Validation Data (scWGS/WES)	Ground truth for performance evaluation	Benchmarking studies and method validation
Cell Type Annotations	Identifies normal reference cells	Critical for normalization in most methods
Harmony/ComBat	Batch effect correction	Essential when integrating datasets from multiple platforms
Snakemake Pipeline	Workflow management	Reproducible benchmarking of multiple methods [3]
SCSsim	Single-cell sequencing simulator	Generating synthetic data for controlled evaluations [21]

Implementation and Practical Considerations

For researchers implementing these methods, several practical considerations emerge from benchmarking studies. Computational requirements vary significantly, with methods incorporating allelic information generally requiring more runtime and memory [3]. The availability of high-quality reference cells is paramount, as reference choice dramatically impacts result quality [3]. Dataset size also influences performance, with some methods scaling better to large cell numbers than others [3].

Diagram 2: Decision workflow for scCNV detection method selection. The process guides researchers from data input through algorithm selection to final interpretation, with key considerations at each step.

The benchmarking studies reveal that no single method outperforms others across all scenarios and metrics. Method selection should be guided by specific research goals, data characteristics, and computational resources. For general-purpose CNV detection, CaSpER and CopyKAT provide the most balanced performance [2] [13]. For subclone identification where batch effects are minimized, InferCNV and CopyKAT excel [2] [4]. When allelic information is available and computational resources are sufficient, Numbat provides robust performance by integrating expression and allele frequency data [3].

Future method development should address current limitations, including sensitivity to batch effects, computational efficiency for large datasets, and improved detection of focal CNVs. The availability of standardized benchmarking pipelines, such as the Snakemake pipeline provided by Colomé-Tatché et al., will facilitate systematic evaluation of new methods as they emerge [3]. As single-cell genomics continues to advance toward clinical applications, accurate and robust CNV detection will play an increasingly important role in understanding tumor evolution and developing targeted therapies.

In the analysis of single-cell sequencing data, accurately interpreting the output of copy number variation (CNV) detection algorithms is as crucial as the analysis itself. These computational tools generally provide results in one of two forms: discrete CNV calls or continuous CNV scores. Discrete calls represent a definitive classification of genomic regions into states such as "loss," "normal," or "gain," providing a clear, categorical interpretation ideal for downstream analyses like clonal grouping. In contrast, continuous scores offer a quantitative measure of the relative copy number signal, allowing researchers to apply their own thresholds and assess the strength or confidence of the prediction [3].

This distinction is not merely an output formatting difference but reflects fundamental methodological approaches and underlying assumptions. The choice between these output types impacts everything from experimental design to biological interpretation, particularly in complex tumor ecosystems where genetic heterogeneity prevails. Understanding the input requirements that lead to these different outputs, and how to properly interpret them, forms a critical component of benchmarking single-cell CNV detection algorithms [2] [4].

Algorithm Classifications and Methodological Approaches

Conceptual Framework of CNV Detection Methods

Single-cell CNV detection methods can be broadly categorized based on their input data requirements and analytical approaches. Expression-based methods rely on the fundamental assumption that genes in amplified regions show elevated expression compared to diploid regions, while deleted regions show reduced expression. These methods require sophisticated normalization strategies against reference diploid cells to distinguish true CNVs from regulatory variation [3]. Allele-based methods incorporate single nucleotide polymorphism (SNP) information from sequencing reads, using allele-specific signals to infer copy number changes. A third category employs hybrid approaches that integrate both expression and allele frequency information for improved accuracy [3] [2].

The computational frameworks underlying these methods vary considerably. Several algorithms implement Hidden Markov Models (HMMs) to segment the genome into regions with distinct copy number states, while others apply segmentation approaches such as circular binary segmentation or dynamic programming to identify breakpoints. Additional strategies include mixture models for estimating CNV states and signal processing techniques that perform multiscale smoothing of input data [3] [6].

Comprehensive Method Categorization

Table 1: Classification of Single-Cell CNV Detection Methods

Method	Primary Input Data	Computational Approach	Output Type	Cell Grouping
InferCNV	Gene expression	HMM	Both discrete & continuous	Subclones
CopyKAT	Gene expression	Segmentation	Both discrete & continuous	Per cell
SCEVAN	Gene expression	Segmentation	Both discrete & continuous	Subclones
CONICSmat	Gene expression	Mixture Model	Discrete	Per cell
CaSpER	Expression + Allelic information	HMM + Signal processing	Both discrete & continuous	Per cell
Numbat	Expression + Allelic information	HMM	Both discrete & continuous	Subclones
HoneyBADGER	Expression ± Allelic information	HMM + Bayesian	CNV probabilities	Per cell
sciCNV	Gene expression	Expression disparity scoring	Continuous scores	Per cell

Methods also differ in whether they provide results for individual cells or group cells into subclones with similar CNV profiles. This distinction significantly impacts output interpretation, as subclonal groupings provide an immediate biological context but may mask rare cell populations or continuous evolutionary processes. Half of the commonly used methods report results per cell (CONICSmat, CopyKAT, CaSpER), while others like InferCNV, SCEVAN, and Numbat group cells into subclones with the same CNV profile [3].

Input Requirements and Data Processing

Fundamental Input Requirements

All scRNA-seq CNV callers require a set of reference diploid cells for normalization, which serves to distinguish technical artifacts from biological signals. For primary tissue samples, the common assumption is that measured tissues contain a mixture of tumor and normal cells, with the latter providing an internal control. However, for cancer cell lines or highly purified samples, researchers must identify matched external reference datasets from healthy cells of similar types [3]. The choice of reference significantly impacts performance, with studies showing that dataset-specific factors including dataset size, the number and type of CNVs in the sample, and reference selection considerably influence results [3] [2].

The sequencing technology represents another critical input consideration. Methods perform differently on droplet-based versus plate-based technologies, with allelic-information-based approaches generally requiring higher sequencing depths to reliably call SNPs from expression data [3]. The number of cells analyzed also affects performance, with some methods requiring minimum cell numbers for robust statistical analysis, while others are specifically designed for large-scale datasets [2].

Experimental Design Considerations

Table 2: Input Requirements Across Platform Types

Requirement	Droplet-Based Platforms	Plate-Based Platforms
Minimum Cells	100+ recommended	Can work with fewer cells
Sequencing Depth	Variable impact by method	Higher depth beneficial for allele-based methods
Reference Cells	Critical for normalization	Critical for normalization
SNP Information	Required for allele-based methods	Required for allele-based methods
Data Normalization	Essential for all methods	Essential for all methods

Benchmarking studies have revealed that the sensitivity and specificity of CNV inference methods vary considerably depending on sequencing depth, read length, and platform selection [2]. Deeper sequencing generally improves detection accuracy but increases computational costs. The resolution of detected CNVs also varies by method, with some tools like CONICSmat reporting results only per chromosome arm, while others provide gene-level or segment-level resolution [3].

Batch effects represent a particularly challenging input consideration, especially when integrating datasets across different platforms. Studies demonstrate that batch effects significantly impact the performance of subclone identification in most methods, potentially leading to artificial clustering based on technical rather than biological differences [2]. Computational correction methods like ComBat can mitigate these effects, with allele-based approaches generally showing greater resilience to batch-related distortions [4].

Output Interpretation: Discrete Calls versus Continuous Scores

Characteristics of Discrete CNV Calls

Discrete CNV calls represent categorical classifications of genomic regions into distinct states, typically including loss, normal, and gain. Some methods further refine these classifications with more granular states such as deep loss, moderate loss, single copy gain, and high amplification. These discrete calls are generated through thresholding approaches applied to continuous signals, where the thresholds may be determined statistically, through machine learning, or set by users based on biological considerations [3].

The primary advantage of discrete calls lies in their interpretability and direct applicability to downstream analyses. They enable clear clonal assignment, phylogenetic reconstruction, and straightforward visualization of CNV landscapes across cells or subpopulations. However, this categorical simplification comes at the cost of losing information about the strength or confidence of the prediction. Discrete calls may also oversimplify complex or subclonal events where copy number changes are present in only a fraction of cells [3] [2].

Characteristics of Continuous CNV Scores

Continuous CNV scores provide quantitative measures of relative copy number across the genome, typically representing normalized expression deviations from the reference diploid profile. These scores preserve the magnitude and confidence of CNV signals, allowing researchers to apply context-specific thresholds and identify subtle variations that might be lost with binary thresholds. Continuous outputs are particularly valuable for assessing clonal evolutionary relationships and identifying rare subpopulations with intermediate CNV states [3] [4].

The interpretation of continuous scores requires careful consideration of baseline establishment and normalization procedures. Methods differ in how they establish the diploid baseline and normalize for technical variability, making direct comparisons of absolute values across methods challenging. Furthermore, the relationship between continuous scores and actual copy number states may be nonlinear and context-dependent, requiring method-specific interpretation [3] [2].

Performance Characteristics Across Output Types

Table 3: Performance Comparison of Selected CNV Detection Methods

Method	Sensitivity	Specificity	Subclone Identification Accuracy	Runtime Efficiency
CaSpER	High	High	Moderate	Moderate
CopyKAT	High	High	High	Moderate
InferCNV	Moderate	Moderate	High	Variable
sciCNV	Moderate	Moderate	Moderate	Fast
HoneyBADGER	Lower sensitivity	Higher specificity	Lower	Slow
Numbat	High (in large datasets)	High (in large datasets)	High	Higher requirements

Benchmarking studies reveal that methods providing both discrete and continuous outputs generally offer flexibility in analysis and interpretation. CaSpER and CopyKAT have demonstrated consistently balanced performance in CNV inference across multiple benchmarking studies, effectively providing both discrete calls and continuous scores [2] [4]. Methods excelling in subclone identification, such as InferCNV and CopyKAT, typically leverage discrete outputs for clear population segmentation, while allele-based methods like Numbat and CaSpER show robust performance in large droplet-based datasets by integrating continuous allele frequency signals [3].

The performance of these methods varies significantly with sequencing depth, with CopyKAT and CaSpER outperforming other methods at lower sequencing depths, while allelic-information-based methods require sufficient depth for accurate SNP calling [2]. In terms of computational efficiency, methods incorporating allelic information generally require higher runtime and memory resources, creating practical constraints for large-scale studies [3].

Experimental Protocols for Benchmarking

Benchmarking Framework Design

Comprehensive benchmarking of CNV detection methods requires a multi-faceted approach evaluating performance against orthogonal validation data. The benchmark pipeline typically involves applying multiple CNV callers to scRNA-seq datasets with known CNV profiles derived from complementary techniques such as single-cell whole-genome sequencing (scWGS), whole exome sequencing (WES), or array comparative genomic hybridization (aCGH) [3] [2] [6].

The evaluation incorporates both threshold-independent metrics like correlation coefficients and area under the curve (AUC) values, and threshold-dependent metrics including sensitivity, specificity, and F1 scores. For discrete calls, the classification accuracy is directly assessed against ground truth, while for continuous scores, partial AUC values are calculated focusing on biologically meaningful threshold ranges [3]. Additional evaluation dimensions include assessing performance on completely euploid datasets, correctness of inferred clonal structures, and robustness to reference dataset selection.

Benchmarking Workflow

Figure 1: Workflow for benchmarking single-cell CNV detection methods

Table 4: Essential Research Reagents and Resources for CNV Benchmarking

Resource	Function	Examples/Specifications
Reference Datasets	Provide ground truth for validation	scRNA-seq with matched scWGS/WES; Cell lines with aCGH validation
Diploid Reference Cells	Normalization control	PBMCs, matched normal tissues, or external healthy references
Computational Infrastructure	Run CNV calling algorithms	High-memory servers (64+ GB RAM); Multi-core processors
Benchmarking Pipelines	Standardized performance assessment	Reproducible Snakemake workflows; Custom evaluation scripts
Orthogonal Validation Data	Establish ground truth CNV profiles	(sc)WGS, WES, aCGH, SNP arrays

The benchmarking environment requires substantial computational resources, with memory requirements varying from modest (8GB) to extensive (64+ GB) depending on the method and dataset size. Runtime also shows considerable variation, with allelic-information-based methods generally requiring more processing time [3]. Reproducible benchmarking pipelines, such as the Snakemake pipeline provided by Colomé-Tatché et al., enable direct testing of new datasets to determine optimal CNV calling strategies [3].

Implications for Research and Clinical Applications

Method Selection Guidelines

The choice between discrete and continuous outputs, and the selection of specific methods, should be guided by research goals, data characteristics, and analytical requirements. For applications requiring clear cell type classification or phylogenetic reconstruction, methods providing high-quality discrete calls like InferCNV and CopyKAT may be preferable. For studies focusing on clonal evolutionary dynamics or detecting subtle CNV changes, methods offering continuous outputs with allele-specific information like CaSpER and Numbat may be more suitable [3] [2] [4].

Dataset size represents another critical consideration. Methods incorporating allelic information generally perform more robustly for large droplet-based datasets but require higher computational resources [3]. For smaller datasets or studies with limited normal reference cells, expression-based methods like CopyKAT may provide more reliable results. The choice of reference dataset significantly impacts performance, particularly for cancer cell lines where external references must be carefully selected [3] [2].

Emerging Trends and Future Directions

The field of single-cell CNV detection continues to evolve rapidly, with several emerging trends influencing input requirements and output interpretation. Integration of multi-omic approaches, combining scRNA-seq with genotypic information, shows promise for improving detection accuracy, particularly for subclonal events. Computational methods are increasingly addressing the challenges of tumor purity and ploidy heterogeneity, with some tools explicitly modeling these factors to improve CNV calling precision [3] [22].

As single-cell genomics transitions toward clinical applications, the need for standardized output formats and interpretation guidelines becomes increasingly important. Method developers are working toward more intuitive output visualizations that effectively communicate both discrete calls and confidence metrics, enabling clinicians and researchers to make informed interpretations. Future benchmarking efforts will need to address these clinical translation challenges, particularly regarding reproducibility across platforms and validation in diverse patient populations [2] [4].

Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for inferring copy number variations (CNVs) at cellular resolution, enabling the dissection of clonal heterogeneity within tumors. The accuracy of these inferences, however, is profoundly influenced by the choice of experimental platform—primarily categorized as droplet-based or plate-based technologies. These platforms differ fundamentally in their throughput, cell loading principles, and the resulting data quality, all of which can impact the performance of computational CNV callers. Framed within a broader thesis on benchmarking single-cell CNV detection algorithms, this guide provides an objective comparison of these platforms, supported by recent experimental data and detailed methodologies, to inform researchers and drug development professionals.

Platform Technologies: Core Principles and Workflows

Droplet-Based Technologies

Droplet-based microfluidics enables high-throughput single-cell analysis by partitioning individual cells into oil emulsion droplets along with barcoded beads [23]. This approach, commercialized by platforms such as 10x Genomics, leverages barcoded oligonucleotides containing cell barcodes and unique molecular identifiers (UMIs) to tag the nucleic acids of thousands of cells in parallel [23]. A key characteristic is Poisson loading, where the number of cells per droplet follows a Poisson distribution. This results in a significant proportion of empty droplets and a predictable rate of multiple cells being encapsulated together, known as doublets [23]. Common validation strategies include species-mixing experiments (e.g., combining human and mouse cells) to quantify the doublet rate, which typically ranges from 0.4% to 11% [23].

Plate-Based Technologies

Plate-based (or microwell-based) methods involve the individual dispensing of cells into the wells of a microtiter plate for processing. These are generally lower-throughput techniques compared to droplet-based systems and are often more manual or require automated liquid handlers [24] [25]. A key advantage is the avoidance of Poisson loading, which gives greater control over cell isolation and significantly reduces the rate of doublets. Recent innovations, such as the traceable medium-throughput scCNV-seq (msCNVS) method, allow for the early barcoding and pooling of up to 384 cells in a microplate, streamlining processing and circumventing the need for whole-genome preamplification [7].

The following diagram illustrates the key procedural and data output differences between the two platforms:

Performance Benchmarking: Quantitative Comparisons

A comprehensive benchmarking study evaluating six popular scRNA-seq CNV callers across 21 datasets revealed that platform-specific factors significantly influence performance [3]. The following table summarizes the impact of the technological platform on key performance metrics.

Table 1: Impact of Platform Type on scRNA-seq CNV Caller Performance

Performance Metric	Droplet-Based Platforms	Plate-Based Platforms
Typical Dataset Size	Large (thousands of cells) [3]	Smaller (hundreds of cells) [3]
Key Influencing Factor	Dataset size, high doublet rate [3] [23]	Lower multiplexing level, controlled cell isolation [7]
Optimal CNV Caller Type	Methods incorporating allelic information (e.g., CaSpER, Numbat) for robust performance [3]	Various callers effective; plate-based-specific protocols like msCNVS show high correlation with bulk data (R=0.90-0.99) [7]
Doublet Rate	0.4% - 11% (requires experimental & computational control) [23]	Significantly lower [7]
Ground Truth Validation	Pseudobulk comparison to (sc)WGS/WES (cell-by-cell often not possible) [3]	Direct cell-by-cell comparison to scWGS possible in some designs [3]

Experimental Protocols for Benchmarking

To ensure rigorous and reproducible benchmarking of CNV callers, specific experimental and computational protocols are employed.

Wet-Lab Validation Experiments

• Species-Mixing Experiments: This is the gold-standard technique for validating droplet-based assays. Human and mouse cells are mixed at a known ratio (e.g., 50:50) and processed together. After sequencing, heterotypic doublets (droplets containing one human and one mouse cell) are identified by their mixed-species expression profile, often visualized on a "barnyard plot". The observed heterotypic doublet rate is used to estimate the total (including homotypic) doublet rate in the experiment [23].
• Orthogonal Ground Truth Validation: The performance of scRNA-seq CNV callers is assessed by comparing their predictions to a CNV ground truth established by an orthogonal method, such as (sc)WGS or WES. For most droplet-based datasets where the same cells cannot be assayed by multiple techniques, scRNA-seq results are combined into a pseudobulk profile before comparison. For some plate-based datasets where scRNA-seq and scWGS are performed on the same cells, a direct cell-by-cell comparison is feasible [3].

Computational Analysis Workflow

The bioinformatic benchmarking typically follows a standardized pipeline [3]:

• Data Processing: Raw sequencing data is processed using the CNV caller's recommended workflow with default parameters.
• Reference Selection: A set of manually annotated euploid (normal) cells is used as a reference for normalization across all methods to ensure comparability.
• Performance Evaluation: The caller's output (either discrete CNV calls or normalized expression scores) is compared to the ground truth using:
  • Threshold-independent metrics: Correlation and Area Under the Curve (AUC) scores, calculated separately for gains and losses. Partial AUC is used to focus on biologically meaningful threshold ranges [3].
  • Threshold-dependent metrics: Sensitivity, specificity, and F1 score, with optimal gain/loss thresholds determined from the data [3].

The Scientist's Toolkit: Essential Reagents and Materials

Table 2: Key Research Reagent Solutions for scCNV Profiling

Item	Function	Example Use Case
Barcoded Beads	Oligonucleotides containing cell barcodes and UMIs for labeling cellular nucleic acids within droplets.	Droplet-based scRNA-seq (10x Genomics) [23].
Modified Tn5 Transposome	An enzyme used for DNA tagmentation (fragmentation and tagging). Can be pre-loaded with barcodes for early cell labeling.	msCNVS protocol for early barcoding of cells in a microplate [7].
Cell Hashing Oligos	Antibody-derived tags (ADTs) or lipid-oligo conjugates used to label cells from different samples prior to pooling.	Sample multiplexing and doublet detection in droplet-overloading experiments [23].
Commercial Microfluidic Kits	Integrated kits containing reagents and chips for partitioning cells.	10x Genomics Single Cell ATAC kit, adapted for Droplet Hi-C [24] [25].

Implications for CNV Caller Selection and Experimental Design

The choice between platforms has direct consequences for downstream analysis:

• For large-scale droplet datasets: Methods that incorporate allelic information (e.g., CaSpER, Numbat) have been shown to perform more robustly, albeit with higher computational runtime [3]. The choice of a suitable euploid reference dataset for normalization is also critical for performance [3].
• For plate-based datasets: A wider range of CNV callers may be applicable. The msCNVS method, combined with its bespoke two-dimensional fitting (2DFit) algorithm, has demonstrated high accuracy in determining absolute CNV profiles and ploidy, making it suitable for clinical applications with limited cell numbers [7].
• Doublet management is critical in droplet data: The inherent doublet rate in droplet-based platforms necessitates both experimental (e.g., cell hashing) and computational doublet detection and removal strategies to prevent erroneous CNV calls that could misrepresent clonal structure [23].

The following diagram summarizes the decision-making logic for platform and tool selection based on project goals:

Inference of copy number variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data fundamentally relies on comparing gene expression patterns of target cells against a reference set of cells with a known diploid genome [3]. This reference set serves as a baseline for normalizing expression data, enabling the detection of genomic regions that are gained or lost in the target cells. The selection of an appropriate reference is therefore a critical analytical decision that directly impacts the accuracy and reliability of CNV calling. Within the context of benchmarking single-cell CNV detection algorithms, research has demonstrated that performance variability among methods is significantly influenced by dataset-specific factors, with the choice of reference dataset being one of the most prominent [3] [18]. This guide objectively compares how different computational methods perform under various reference selection scenarios, providing researchers with data-driven strategies to optimize their scCNV analyses.

Performance Comparison of Reference Strategies

Benchmarking studies systematically evaluate the impact of reference selection by measuring a method's ability to recover ground truth CNVs—typically established by orthogonal methods like (sc)Whole Genome Sequencing ((sc)WGS) or Whole Exome Sequencing (WES)—across different reference scenarios [3]. Key performance metrics include the Root Mean Square Error (RMSE) for diploid samples, F1 scores for aneuploid samples, and correlation coefficients with the ground truth.

Table 1: Summary of scCNV Method Performance with Different Reference Types

Method	Algorithm Type	Performance with Matched Internal Reference	Performance with External Reference	Automatic Reference Cell Detection
Numbat	Expression + Allelic Frequency (HMM)	Excellent (Lowest RMSE on diploid data) [18]	Superior performance compared to expression-only methods [18]	Yes (High concordance with manual annotation) [18]
CaSpER	Expression + Allelic Frequency (Multiscale Smoothing)	Excellent [18]	Robust performance with external datasets [18]	Not Specified
InferCNV	Expression-only (HMM)	Good with correct internal reference [3]	Performance degrades with reference mismatch [18]	No
CopyKAT	Expression-only (Statistical Model)	Good with correct internal reference [3] [2]	Moderate performance drop with external reference [18]	Yes [18]
SCEVAN	Expression-only (Segmentation)	Good with correct internal reference [3]	Performance varies [3]	Yes [18]
CONICSmat	Expression-only (Mixture Model)	Good with correct internal reference [3]	Performance varies [3]	No

Table 2: Quantitative Performance Impact of Reference Selection on a Diploid PBMC Dataset [18]

Method	RMSE (T-cells from same sample as reference)	RMSE (Monocytes from same sample as reference)	RMSE (T-cells from external dataset as reference)
Numbat	Low	Low	Lowest
CaSpER	Low	Low	Low
InferCNV	Low	Moderate	High
CopyKAT	Low	Moderate	High
SCEVAN	Low	High	High
CONICSmat	Low	High	High

The data reveals a clear hierarchy in reference quality. The optimal scenario uses normal cells from the same sample as the reference, as they share identical technical noise profiles [18]. When this is unavailable, methods that incorporate allelic frequency information (Numbat, CaSpER) demonstrate significantly more robust performance when using external reference datasets or non-ideal cell types from the same sample [18]. This is visually summarized in the relationship between reference quality and method performance below.

Detailed Experimental Protocols from Benchmarking Studies

Protocol 1: Evaluating Reference Impact on a Diploid Sample

This protocol assesses how a method performs when no CNVs are present and quantifies the false positive signal introduced by reference mismatch.

• Objective: To measure a method's tendency to generate false positive CNV calls in a diploid sample using different reference cell types and external datasets [3] [18].
• Dataset: scRNA-seq data from a diploid sample, specifically Peripheral Blood Mononuclear Cells (PBMCs) from a healthy donor [3] [18].
• Cell Annotation: Manual annotation of cell types (e.g., T-cells, monocytes) within the PBMC dataset is performed using known marker genes and Louvain clustering [3].
• Reference Scenarios:
  • Ideal: T-cells from the same PBMC sample are used as the reference.
  • Sub-optimal Internal: Monocytes from the same PBMC sample are used as the reference.
  • External: T-cells from a completely different, external PBMC dataset are used as the reference [18].
• Analysis: Each CNV calling method is run on the diploid T-cells using the different reference sets.
• Evaluation Metric: The Root Mean Square Error (RMSE) is calculated between the method's CNV prediction and a diploid baseline (ground truth of no CNVs). A lower RMSE indicates better performance [18].

Protocol 2: Evaluating Reference Impact on Aneuploid Tumors

This protocol tests a method's accuracy in detecting true CNVs within heterogeneous tumor samples when using different references.

• Objective: To evaluate the impact of reference choice on the sensitivity and specificity of CNV detection in aneuploid samples [3] [18].
• Datasets: scRNA-seq datasets from primary cancer samples (e.g., Multiple Myeloma (MM), gastric cancer line SNU601) with a known ground truth from (sc)WGS or WES [3] [18].
• Reference Scenarios:
  • Matched Normal Cells: Healthy cells from the same sample (e.g., immune cells from the tumor microenvironment) are used.
  • External Healthy Dataset: A healthy dataset from a similar tissue or cell type is used.
  • External Cancer Dataset: A cancer dataset is used as a negative control to test robustness [18].
• Analysis: Methods are run on the tumor cells using each reference scenario.
• Evaluation Metrics:
  • Threshold-independent: Correlation and Area Under the Curve (AUC) scores between the pseudobulk CNV profile and the ground truth [3].
  • Threshold-dependent: Maximum F1 score, sensitivity, and specificity for gains and losses after determining optimal thresholds [3].

Table 3: Key Resources for scCNV Benchmarking and Analysis

Resource Name	Type	Function in Research	Example Use Case / Note
Benchmarking Pipeline [3] [18]	Computational Tool	A Snakemake pipeline for reproducible benchmarking of scCNV callers on new datasets.	Allows researchers to identify the optimal method for their specific data.
Reference Datasets	Data	Provide diploid baseline expression profiles for normalization.	Healthy PBMC or tissue-specific atlases are commonly used.
Orthogonal CNV Data ((sc)WGS/WES)	Data / Method	Serves as the ground truth for validating scRNA-seq-based CNV calls.	Essential for calculating performance metrics like AUC and F1 score.
Cell Type Annotation Tools	Computational Tool	Identify putative normal cells within a mixed sample to use as an internal reference.	e.g., Louvain clustering with marker gene analysis [3].
BAFExtract [18]	Computational Tool	A required tool for the CaSpER method to extract allele frequency information from scRNA-seq reads.	Enables the use of allelic information for more robust CNV calling.

The consensus from independent benchmarking studies is clear: the most robust strategy for scCNV analysis is to use matched normal cells from the same sample as the reference, whenever available [3] [18]. For samples where such an internal reference is not present, such as cancer cell lines or highly purified tumor samples, the choice of computational method becomes paramount. In these scenarios, selecting a method that integrates allelic frequency information, such as Numbat or CaSpER, is strongly recommended due to their demonstrated resilience to reference dataset mismatch [3] [18]. Expression-only methods like InferCNV and CopyKAT can perform well with an ideal internal reference but show greater performance degradation with suboptimal references. Therefore, researchers must align their reference selection and method choice with the biological context and cellular composition of their samples to ensure the accurate detection of copy number variations.

Optimizing CNV Detection: Navigating Technical Pitfalls and Performance Limitations

Batch effects represent one of the most significant technical challenges in single-cell RNA sequencing (scRNA-seq) analysis, particularly in cross-platform studies and multi-center investigations. These non-biological variations arise when samples are processed in different batches—across different times, handling personnel, reagent lots, protocols, or sequencing platforms—introducing systematic technical biases that can confound biological interpretations [26]. In the context of benchmarking single-cell copy number variation (CNV) detection algorithms, batch effects significantly impact the fidelity of CNV inference, as they can obscure true biological variations and introduce artifacts that affect downstream analyses [3] [12]. The presence of batch effects complicates the integration of datasets generated across different platforms and centers, which is essential for robust algorithm benchmarking and validation.

The challenge is particularly pronounced in scRNA-seq data due to its high-dimensionality, sparsity, and complexity [27]. Different scRNA-seq technologies can be broadly categorized into full-length and 3' end counting-based methods, each with distinct characteristics regarding sensitivity, specificity, and incorporation of unique molecular identifiers (UMIs) [28]. These technological differences, combined with variations in experimental execution across centers, create complex batch effects that must be addressed before meaningful biological comparisons can be made, including accurate CNV detection from transcriptomic data.

Impact of Batch Effects on Cross-Platform scRNA-seq Analysis

Empirical Evidence from Multi-Center Studies

A comprehensive multi-center benchmarking study generating 20 scRNA-seq datasets from two biologically distinct cell lines across four sequencing centers and multiple platforms demonstrated that batch effects substantially impact cross-platform analysis [28]. This study utilized a human breast cancer cell line (HCC1395) and a matched B lymphocyte cell line (HCC1395BL) from the same donor, processed both individually and as mixtures using four scRNA-seq platforms: 10x Genomics Chromium, Fluidigm C1, Fluidigm C1 HT, and Takara Bio's ICELL8 system. The findings revealed that technical factors including technology platform, inter-laboratory differences in cell handling, and library construction protocols introduced substantial variability that affected gene detection and cell classification accuracy.

The study design specifically addressed the challenge of distinguishing technical factors from biological variability, which is particularly difficult when only mixtures of cells are used across different platforms [28]. By distributing samples of both cell lines to different centers for independent culture and processing, the researchers evaluated the experimental variability encountered in real-world collaborations. This approach revealed that batch effects were quite large and that the ability to assign cell types correctly across platforms and sites was highly dependent on bioinformatic pipelines, particularly the batch correction algorithms employed.

Consequences for CNV Detection Algorithms

Batch effects present particular challenges for scRNA-seq CNV calling methods, which infer copy number variations based on the assumption that genes in gained regions show higher expression and in lost regions show lower expression compared to genes in diploid regions [3]. Technical variations introduced by batch effects can mimic or obscure these expression patterns, leading to false positives or negatives in CNV detection. A recent benchmarking study of six popular scRNA-seq CNV callers (InferCNV, copyKat, SCEVAN, CONICSmat, CaSpER, and Numbat) across 21 datasets found that batch effects significantly affected the performance of subclone identification in mixed datasets for most methods tested [3] [12].

The impact of batch effects on CNV detection is method-dependent, with algorithms varying in their sensitivity to technical variations. Methods that include allelic information (CaSpER and Numbat) generally perform more robustly for large droplet-based datasets but require higher computational runtime [3]. The selection of appropriate reference datasets for normalization, a critical step in CNV calling, is also complicated by batch effects, as ideal reference cells should be biologically comparable but technically consistent with the test dataset—a balance difficult to achieve with strong batch effects present.

Benchmarking Batch Effect Correction Methods

Multiple computational methods have been developed to address batch effects in scRNA-seq data, each employing distinct strategies and operating on different components of the data structure. These approaches can be broadly categorized based on their correction methodology and the data objects they modify:

Table 1: Batch Effect Correction Methods and Their Characteristics

Method	Input Data	Correction Object	Correction Methodology	Returns
Harmony [29] [30]	Normalized count matrix	Embedding	Soft k-means with linear batch correction within clusters	Corrected embedding
Seurat [29] [30]	Normalized count matrix	Embedding	CCA alignment and mutual nearest neighbors	Corrected count matrix
BBKNN [29]	k-NN graph	k-NN graph	UMAP on merged neighborhood graph	Corrected k-NN graph
ComBat/ComBat-seq [29]	Raw/Normalized count matrix	Count matrix	Empirical Bayes linear correction/Negative binomial regression	Corrected count matrix
MNN/fastMNN [30]	Normalized count matrix	Count matrix	Mutual nearest neighbors with linear correction	Corrected count matrix
LIGER [30]	Normalized count matrix	Embedding	Quantile alignment of factor loadings	Corrected embedding
SCVI [29]	Raw count matrix	Embedding	Variational autoencoder modeling batch effects	Corrected count matrix and embedding

Performance Comparison Across Metrics

Multiple benchmarking studies have evaluated batch effect correction methods using different metrics and scenarios. A comprehensive 2020 benchmark evaluating 14 methods across ten datasets using five evaluation metrics (kBET, LISI, ASW, ARI, and computational runtime) recommended Harmony, LIGER, and Seurat 3 as top performers [30]. Harmony was particularly noted for its significantly shorter runtime, making it a recommended first choice for batch integration.

A more recent 2025 evaluation introduced RBET (Reference-informed Batch Effect Testing), a novel statistical framework designed to evaluate batch correction performance with sensitivity to overcorrection [31]. Using extensive simulations and real data examples, this study demonstrated that RBET provides more biologically meaningful evaluations compared to existing metrics like kBET and LISI, particularly because it remains robust to large batch effect sizes and can detect overcorrection where true biological variation is erroneously removed.

Table 2: Performance Comparison of Batch Effect Correction Methods

Method	Batch Removal Effectiveness	Biological Preservation	Computational Efficiency	Overcorrection Risk
Harmony [29] [30]	High	High	High	Low
Seurat [29] [30]	High	Medium	Medium	Medium-High
BBKNN [29]	Medium	Medium	High	Low
ComBat/ComBat-seq [29]	Medium	Low	Medium	Medium
MNN/fastMNN [29] [30]	Medium-High	Medium	Medium	Medium
LIGER [29] [30]	High	Medium	Medium	Medium
SCVI [29]	Medium	Low	Low (training required)	Medium

Special Considerations for CNV Detection

When batch correction is applied in the context of CNV detection from scRNA-seq data, special considerations must be addressed. A key challenge is that overcorrection can remove genuine biological variations resulting from underlying CNVs, thereby reducing the sensitivity of CNV detection algorithms [31]. Methods that aggressively correct batch effects might erroneously normalize expression differences that actually originate from copy number alterations rather than technical artifacts.

The 2025 benchmarking study of scRNA-seq CNV callers revealed that the choice of reference dataset significantly impacts performance, and this dependency is exacerbated when batch effects are present [3]. For cancer cell lines, where no directly matched reference cells exist, researchers must select external reference datasets with healthy cells from similar cell types, making appropriate batch correction essential for meaningful comparisons.

Furthermore, methods that modify the count matrix directly (such as ComBat, ComBat-seq, MNN, and Seurat) may introduce artifacts that affect CNV calling, which typically relies on relative expression differences across genomic regions [29]. In contrast, methods that only correct embeddings or k-NN graphs (such as Harmony, BBKNN, and LIGER) preserve the original count matrix but may provide less comprehensive batch effect removal for downstream CNV analysis.

Experimental Protocols for Batch Effect Evaluation

Multi-Center Study Design

The benchmark scRNA-seq dataset generation protocol described in [28] provides a robust framework for evaluating batch effects and correction methods. The experimental workflow can be summarized as follows:

Figure 1: Experimental Workflow for Batch Effect Assessment in Multi-Center Studies

Cell Line Preparation and Culture

• Obtain characterized reference cell lines (e.g., HCC1395 breast cancer cells and HCC1395BL B lymphocytes from ATCC)
• Culture cells according to manufacturer specifications using appropriate media
• HCC1395: RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS)
• HCC1395BL: Iscove's Modified Dulbecco's Medium supplemented with 20% FBS

Multi-Center Processing

• Distribute cell lines to multiple sequencing centers while maintaining consistent cell viability standards
• Process cells using different scRNA-seq platforms at each center:
  • Full-length protocols: Fluidigm C1 system (LLU Center) with SMART-Seq v4 Ultra Low Input RNA kit
  • 3' end counting-based protocols: 10x Genomics Chromium (LLU, NCI), Fluidigm C1 HT (FDA), ICELL8 (Takara Bio)
• Generate both individual samples and predefined mixtures of both cell lines

Library Preparation and Sequencing

• Prepare libraries according to platform-specific manufacturer protocols
• Sequence libraries using Illumina platforms (HiSeq 2500/4000) with appropriate read lengths:
  • 150×2 bp paired-end for full-length protocols
  • 75×2 bp paired-end for 3' end counting-based protocols

RBET Evaluation Framework

The Reference-informed Batch Effect Testing (RBET) framework provides a novel approach for evaluating batch correction performance with sensitivity to overcorrection [31]. The methodology consists of two main steps:

Figure 2: RBET Evaluation Framework for Batch Effect Correction

Reference Gene Selection

• Strategy A (Literature-Based): Collect experimentally validated tissue-specific housekeeping genes from published literature as reference genes (RGs)
• Strategy B (Data-Driven): Select genes stably expressed both within and across phenotypically different clusters directly from the dataset
• Validate that candidate RGs are not differentially expressed across batches in the uncorrected data

Batch Effect Detection

• Map the dataset into a two-dimensional space using UMAP
• Apply maximum adjusted chi-squared (MAC) statistics for batch effect detection between sample distributions
• Calculate RBET score where smaller values indicate better batch correction performance
• Compare with traditional metrics (kBET, LISI) to evaluate sensitivity to overcorrection

Computational Tools and Algorithms

Table 3: Essential Computational Tools for Batch Effect Correction

Tool/Algorithm	Primary Function	Implementation	Key Applications
Harmony [29] [30]	Batch effect correction	R/Python	Multi-dataset integration, cross-platform analysis
Seurat [29] [30]	Data integration and correction	R	Multi-modal single-cell data integration
BBKNN [29]	Batch-balanced k-nearest neighbors	Python	Large dataset integration, graph-based correction
InferCNV [3]	CNV inference from scRNA-seq	R	Tumor heterogeneity analysis, subclone identification
copyKat [3] [12]	CNV inference and subclone detection	R	Cancer evolution studies, aneuploidy detection
CaSpER [3]	CNV inference with allelic information	R/Python	Comprehensive CNV profiling, allele-specific analysis
Numbat [3]	CNV inference with haplotype phasing	R	High-resolution CNV detection, phylogenetic analysis

Reference Datasets and Biological Materials

• Reference Cell Lines: HCC1395 and HCC1395BL (ATCC) provide well-characterized reference samples with matched normal and tumor profiles [28]
• Benchmarking Datasets: The 20-dataset benchmark from the multi-center study enables standardized method evaluation [28]
• Housekeeping Gene Panels: Tissue-specific reference genes with stable expression patterns serve as internal controls for batch effect assessment [31]
• PBMC Controls: Peripheral blood mononuclear cells from healthy donors provide euploid reference data for CNV caller evaluation [3]

Batch effects present substantial challenges for cross-platform scRNA-seq analysis and significantly impact the performance of CNV detection algorithms. Based on comprehensive benchmarking studies, Harmony emerges as a strongly recommended batch correction method due to its effective batch removal, biological preservation, computational efficiency, and lower risk of overcorrection [29] [30]. However, method selection should be guided by specific experimental contexts and downstream applications.

For CNV detection studies, careful consideration must be given to the potential for overcorrection, which might remove genuine biological signals resulting from copy number alterations. The RBET framework provides a valuable approach for evaluating correction performance with sensitivity to this risk [31]. Additionally, the choice of appropriate reference datasets remains critical for both batch correction and subsequent CNV inference, particularly for cancer cell lines where matched normal references are unavailable [3].

Future developments in batch effect correction should focus on improving method calibration to minimize artifacts while preserving biological integrity, particularly in the context of genetic alteration detection from transcriptomic data. As single-cell technologies continue to evolve and datasets grow in scale and complexity, robust batch effect management will remain essential for valid biological interpretations and reliable CNV detection in cross-platform analyses.

The reliable detection of copy number variations (CNVs) is fundamentally dependent on the quality of the input sequencing data, with sequencing depth representing one of the most critical determinants of success. In the context of single-cell CNV detection, this relationship becomes even more pronounced due to the inherent technical challenges of single-cell data, including amplification bias, allelic dropout, and sparse coverage. The broader thesis of benchmarking single-cell CNV detection algorithms must therefore include a rigorous examination of how these data quality parameters influence diagnostic accuracy. As computational methods evolve to extract increasingly subtle signals from sequencing data, establishing clear minimum requirements for reliable detection enables researchers to design cost-effective experiments while minimizing false discoveries. This guide synthesizes recent benchmarking evidence to provide actionable recommendations for data generation and quality control in single-cell CNV studies, with particular emphasis on the interplay between sequencing depth, data quality, and algorithmic performance across diverse experimental contexts.

Experimental Designs for Benchmarking Depth Requirements

Systematic Evaluation Frameworks

Comprehensive benchmarking studies employ carefully designed experimental frameworks to quantify the relationship between sequencing depth and CNV detection performance. These typically involve either simulated datasets with known ground truth CNVs or real datasets validated through orthogonal methods such as single-cell whole-genome sequencing ((sc)WGS) or whole exome sequencing (WES) [3] [32]. For simulation-based approaches, tools like SInC V2.0 generate synthetic sequencing data with predefined CNV characteristics across different coverage depths, tumor purities, and variant types [9]. This enables precise calculation of performance metrics including precision, recall, and F1-score at each depth level.

For real data validation, studies utilize reference cell lines with well-characterized CNV profiles or clinical samples with matched orthogonal validation data [11]. The benchmarking process typically involves downsampling sequencing data to various depth levels followed by CNV calling to assess how sensitivity and specificity degrade with reduced coverage [9]. Performance is evaluated against ground truth using metrics such as correlation coefficients, area under the curve (AUC) values, and F1 scores for detecting gains versus losses separately [3] [32]. This systematic approach allows researchers to establish evidence-based thresholds for minimum sequencing requirements across different biological contexts and computational methods.

Key Performance Metrics and Evaluation Criteria

The evaluation of depth requirements incorporates multiple performance dimensions to provide a comprehensive assessment of detection reliability. Threshold-independent metrics like correlation and AUC scores offer insights into overall signal quality, while threshold-dependent metrics like sensitivity, specificity, and F1-score measure classification accuracy at biologically meaningful cutoffs [3] [32]. For single-cell methods, additional considerations include the ability to correctly identify euploid cells, resolve subclonal structures, and accurately determine breakpoints [3] [33]. The partial AUC metric is particularly valuable as it focuses on the biologically meaningful range of thresholds for gains and losses rather than the complete threshold spectrum [32]. Together, these metrics provide a multi-faceted view of how sequencing depth impacts practical utility across diverse research applications.

Minimum Depth Requirements Across Sequencing Modalities

Single-Cell RNA Sequencing for CNV Detection

For scRNA-seq CNV detection, the sequencing depth requirements are intrinsically linked to both the computational method employed and the specific biological question being addressed. Benchmarking of six popular scRNA-seq CNV callers reveals distinct performance characteristics across different data quality tiers [3] [32]. Methods utilizing only gene expression values (InferCNV, copyKat, SCEVAN, CONICSmat) generally require sufficient depth to robustly detect expression differences across genomic regions, while methods incorporating allelic information (CaSpER, Numbat) have additional depth requirements for reliable SNP calling [3]. Although exact minimum depth thresholds are method-dependent, the benchmarking demonstrates that droplet-based technologies typically require deeper sequencing to compensate for sparser coverage compared to plate-based methods [32].

The performance characteristics of these methods further inform depth requirements. Methods incorporating allelic information (CaSpER and Numbat) demonstrate more robust performance for large droplet-based datasets but require higher computational resources [3] [32]. The expression-based methods show greater variability in their ability to correctly identify ground truth CNVs, euploid cells, and subclonal structures across the 21 tested datasets [3]. This suggests that for applications requiring high confidence in subclonal resolution, investing in deeper sequencing to leverage allele-aware methods may be warranted despite increased computational costs.

Whole-Genome Sequencing Platforms

For WGS-based CNV detection, depth requirements vary significantly based on the specific technology and application context. Low-pass whole-genome sequencing has emerged as a cost-effective alternative to microarrays for detecting clinically significant CNVs, with typical requirements of 1-10x coverage depending on the desired resolution [34]. For detecting larger CNVs such as aneuploidies, 1-2x coverage may be sufficient, while comprehensive detection of deletions, duplications, and loss of heterozygosity typically requires ≥5x coverage [34].

Standard WGS for germline CNV detection typically operates at ~30x coverage, though specialized clinical applications may adjust these requirements based on specific diagnostic needs [11]. Benchmarking studies reveal that CNV callers generally perform better for deletions (up to 88% sensitivity) than duplications (up to 47% sensitivity), with particularly poor detection of duplications under 5 kb [11]. This performance asymmetry suggests that applications focused on duplication detection may benefit from increased sequencing depth.

Single-cell WGS presents unique challenges for depth determination due to amplification biases and uneven coverage. The recently introduced HiScanner method leverages high-coverage scWGS data (>20x) to identify CNAs with high resolution by combining read depth, B-allele frequency, and haplotype phasing [33]. For low-coverage scWGS applications typical in cancer studies (~0.5x average coverage), methods like CHISEL and Alleloscope attempt to leverage phasing and BAF information, but struggle with detecting small CNAs (<5 Mb) [33].

Specialized Sequencing Applications

Whole genome bisulfite sequencing (WGBS) presents unique challenges for CNV detection due to bisulfite conversion. A comprehensive benchmark of 35 strategies for CNV detection from WGBS data identified optimal aligner-caller combinations: bwameth-DELLY and bwameth-BreakDancer for deletions, and walt-CNVnator and bismarkbt2-CNVnator for duplications [35]. While specific depth thresholds were not provided, the study emphasized that accurate CNV detection from WGBS data requires specialized computational approaches optimized for bisulfite-converted sequences.

Long-read sequencing technologies offer advantages for detecting complex structural variations but have distinct depth considerations. Clinical validation of a long-read sequencing pipeline demonstrated high analytical sensitivity (98.87%) and specificity (>99.99%) for detecting diverse variant types, though specific depth requirements depend on the application and technology platform [36].

Table 1: Minimum Recommended Sequencing Depth by Technology and Application

Sequencing Technology	Application Context	Minimum Depth	Key Considerations
scRNA-seq	Droplet-based CNV calling	Method-dependent	Allele-aware methods require depth for reliable SNP calling [3]
scRNA-seq	Plate-based CNV calling	Method-dependent	Generally less depth required than droplet-based [32]
Low-pass WGS	Aneuploidy detection	1-2x	Suitable for large CNVs only [34]
Low-pass WGS	Comprehensive CNV detection	≥5x	Detects deletions, duplications, and LOH [34]
Standard WGS	Germline CNV detection	~30x	Better detection of deletions than duplications [11]
scWGS	High-resolution CNA detection	>20x	Required for methods like HiScanner [33]
scWGS	Clonal pattern detection	~0.5x	Limited to large, chromosomal arm-sized CNAs [33]

Impact of Data Quality Beyond Sequencing Depth

Critical Factors Influencing Detection Reliability

While sequencing depth receives significant attention, multiple additional factors profoundly impact CNV detection reliability. Tumor purity substantially influences detection accuracy, with low-purity samples causing signal confounding that mimics copy-neutral events [9]. Systematic evaluations demonstrate that CNV detection performance degrades significantly at tumor purities below 60%, with some tools struggling to maintain acceptable sensitivity even at 40% purity [9]. The type and size of CNVs also dramatically affect detectability, with shorter variants (<100 kb) frequently overlooked and longer variants more readily detected across all methods [9].

The choice of reference dataset for normalization emerges as a particularly critical factor in scRNA-seq CNV detection, with benchmarking revealing substantial performance variations depending on reference selection [3] [32]. For cancer cell lines where matched normal references are unavailable, the selection of appropriate external reference datasets becomes essential for reliable detection [32]. Additionally, technical factors including read length, GC bias, and the uniformity of coverage distribution significantly influence detection accuracy, sometimes outweighing the impact of raw sequencing depth [34] [33].

Method-Specific Considerations and Limitations

Different computational approaches exhibit distinct strengths and limitations that interact with data quality parameters. Methods employing read-depth approaches generally detect CNVs based on coverage depth correlations with copy number but struggle with small variants (<100 kb) [34] [9]. Split-read methods identify breakpoints at base-pair resolution but are limited in detecting large-scale variants (≥1 Mb) [34]. Assembly-based approaches theoretically detect all variation types but face prohibitive computational demands [34].

For single-cell methods, technical artifacts including allelic dropout, amplification bias, and phase switch errors introduce unique challenges that necessitate specialized computational strategies [33]. The HiScanner method addresses these challenges by inferring optimal bin size based on allelic dropout patterns and performing joint segmentation across cells to amplify signals of clonal CNAs [33]. These methodological innovations enable higher-resolution detection but impose specific data quality requirements that must be considered during experimental design.

Table 2: Performance of CNV Detection Tools Across Data Quality Dimensions

Tool Category	Strength	Limitations	Data Quality Requirements
Expression-based scRNA-seq callers (InferCNV, copyKat)	No requirement for SNP information; lower computational demands [3]	Performance varies with reference selection; limited resolution for small CNVs [32]	Dependent on expression coverage and reference quality [3]
Allele-aware scRNA-seq callers (CaSpER, Numbat)	More robust for large droplet-based datasets; leverage haplotype information [3]	Higher runtime and memory requirements; need sufficient SNP coverage [32]	Require depth for reliable genotype calling in addition to expression [3]
Read-depth WGS callers	Detect CNV dosage; work for large-sized CNVs [34]	Struggle with small variants (<100 kb) [9]	Uniform coverage distribution critical [34]
Split-read WGS callers	Base-pair breakpoint resolution [34]	Limited for large variants (≥1 Mb) [34]	Longer read lengths beneficial [34]
BAF-aware scWGS callers (HiScanner, CHISEL)	Allele-specific copy number state inference [33]	Sensitive to phase switch errors and allelic dropout [33]	Require heterozygous SNPs and accurate phasing [33]

Table 3: Essential Research Reagents and Computational Solutions for CNV Detection Studies

Item	Function	Example Applications
Reference cell lines (HG002, Coriell Institute catalog)	Provide ground truth for benchmarking and validation [11]	Establishing performance baselines; validating novel methods [11]
Orthogonal validation technologies ((sc)WGS, WES)	Generate verification data for scRNA-seq CNV calls [3] [32]	Confirming CNVs detected in primary modality [3]
Specialized alignment tools (bwameth, WALT)	Optimized mapping for specific sequencing applications [35]	Processing bisulfite-converted sequences for CNV detection [35]
Benchmarking pipelines (Snakemake workflow)	Enable reproducible method comparisons [3] [32]	Standardized evaluation of new CNV callers [3]
Visualization platforms (NxClinical, ViScanner)	Facilitate interpretation of complex CNV data [34] [33]	Integrative analysis of CNVs, SNVs, and AOH regions [34]

Experimental Workflow for Determining Depth Requirements

Determining Sequencing Depth Requirements Workflow

The reliable detection of CNVs from sequencing data requires careful consideration of multiple interacting factors beyond simple depth metrics. Based on current benchmarking evidence, the following recommendations emerge:

First, align sequencing depth with specific biological questions. For detecting chromosomal-scale alterations in cancer samples, low-coverage approaches (0.5-5x) may suffice, while studies aiming to identify small CNAs (<5 Mb) in heterogeneous samples require significantly deeper coverage (>20x) [34] [33]. Second, match computational methods to data characteristics. Expression-based scRNA-seq callers offer practical solutions for preliminary analyses, while allele-aware methods provide more robust detection for well-powered studies with sufficient sequencing depth for reliable genotype calling [3]. Third, implement comprehensive quality control measures including reference selection optimization, tumor purity assessment, and technical artifact mitigation [3] [9] [33].

As single-cell technologies continue to evolve, the relationship between sequencing depth and detection reliability will undoubtedly shift. Emerging methods that more efficiently extract biological signals from sparse data may gradually reduce depth requirements, while increasingly sophisticated algorithms may leverage additional information to improve resolution. Through continued benchmarking efforts and method development, the field will establish more refined guidelines that balance practical constraints with scientific rigor in the pursuit of reliable CNV detection across diverse biological contexts.

Tumor purity, defined as the proportion of cancerous cells within a heterogeneous tissue sample, represents a fundamental challenge in genomic analysis. In the context of single-cell copy number variation (CNV) detection, low tumor purity can obscure genetic signals, leading to inaccurate variant calling and misinterpretation of tumor heterogeneity [37] [38]. The contamination of normal cells in tumor tissues dilutes the tumor-derived genomic signal, potentially obscuring true copy number alterations and reducing the sensitivity of detection algorithms [39]. This technical limitation has profound implications for cancer research and clinical practice, where accurate CNV profiling is essential for understanding tumor evolution, identifying therapeutic targets, and tracking treatment response.

Benchmarking studies have systematically revealed that tumor purity significantly impacts the fidelity of CNV detection across multiple computational methods [37] [39] [38]. When tumor purity falls below critical thresholds, the accuracy of mutation detection decreases substantially, with false-negative rates increasing dramatically [37]. This comprehensive guide examines how different scRNA-seq CNV detection methods perform under varying tumor purity conditions, providing researchers with evidence-based recommendations for selecting appropriate tools and implementing protocols that mitigate purity-related challenges.

Performance Comparison of CNV Detection Methods

Quantitative Benchmarking Across Tumor Purity Conditions

Systematic evaluations of CNV detection tools have revealed significant performance variations under different tumor purity conditions. A comprehensive benchmarking study assessed six popular scRNA-seq CNV callers across 21 datasets, examining their ability to correctly identify ground truth CNVs established through orthogonal validation methods like single-cell or bulk whole-genome sequencing [(sc)WGS] or whole-exome sequencing (WES) [3]. The methods evaluated included InferCNV, CopyKat, SCEVAN, CONICSmat, CaSpER, and Numbat, representing both expression-based and allele-integrated approaches [3].

Table 1: Performance Metrics of scRNA-seq CNV Callers on Heterogeneous Samples

Method	Data Type Used	Performance at High Purity	Performance at Low Purity	Tumor Cell Classification Accuracy	Reference Dependency
Numbat	Expression + Allelic	High	Moderate-High	Best (high concordance with manual)	Low (robust to reference choice)
CaSpER	Expression + Allelic	High	Moderate	Moderate	Low (robust to reference choice)
CopyKAT	Expression only	High	Moderate	High	Moderate
InferCNV	Expression only	High	Moderate	Moderate-High	High
SCEVAN	Expression only	Moderate-High	Moderate	High (automatic annotation)	High
CONICSmat	Expression only	Moderate	Low	Poor	High

A separate benchmarking study focusing on five commonly used scCNV inference methods (HoneyBADGER, inferCNV, sciCNV, CaSpER, and CopyKAT) found that CaSpER and CopyKAT generally outperformed other methods in balanced CNV inference, though effectiveness varied with sequencing depth and platform type [2] [4]. For subclone identification in mixed tumor populations, inferCNV and CopyKAT demonstrated superior performance, particularly when analyzing data from a single platform [2].

For low-coverage whole-genome sequencing (lcWGS) data, ichorCNA demonstrated superior performance in precision and runtime at high tumor purity (≥50%), making it the optimal choice for lcWGS-based workflows in high-purity samples [39]. However, at lower purity levels, methods incorporating allelic information like Numbat and CaSpER showed more robust performance, particularly when reference datasets were suboptimal [3].

Impact of Tumor Purity Thresholds on Detection Fidelity

Research has established that tumor purity significantly affects mutation detection accuracy, with a critical threshold identified at approximately 30% [37]. Below this purity level, the number of false-negative mutations increases substantially, and variant allele frequencies (VAFs) become significantly underestimated [37]. Experimental evidence demonstrates that when tumor purity drops from 100% to 20-30%, the number of detected mutations can decrease by nearly half, and tumor mutational burden (TMB) values become substantially underestimated [37].

Table 2: Impact of Tumor Purity on Somatic Mutation Detection in Colorectal Cancer

Tumor Purity Level	False Negative Rate	False Positive Rate	Mutation Detection F-score	VAF Reduction	CNV Detection Accuracy
>50%	Low (<5%)	Very Low (<2%)	>0.95	Minimal	High
30-50%	Moderate (5-15%)	Low (2-5%)	0.85-0.95	Moderate	Moderate-High
<30%	High (15-40%)	Low (2-5%)	0.60-0.85	Substantial	Low-Moderate

The degradation in detection sensitivity at low purity affects different mutation types unevenly. Copy number variations with low amplification levels or heterozygous deletions are particularly susceptible to being missed in impure samples [37]. Additionally, subclonal populations with lower VAFs become increasingly difficult to detect as overall tumor purity decreases, potentially leading to incomplete reconstruction of tumor evolutionary history [37] [2].

Experimental Protocols for Assessing Method Performance

Benchmarking Framework and Validation Standards

Well-designed benchmarking studies follow rigorous experimental protocols to assess how CNV detection methods perform under varying tumor purity conditions. The general workflow involves multiple stages from dataset selection and ground truth establishment to method evaluation and statistical analysis [3] [2].

Figure 1: Workflow for benchmarking CNV detection methods against tumor purity challenges.

The benchmark evaluation typically employs multiple metrics to comprehensively assess method performance [3] [2]:

• Sensitivity and Specificity: Calculated by comparing detected CNVs against orthogonal ground truth measurements from (sc)WGS or WES data.
• F1 Scores: Harmonic mean of precision and recall, providing balanced assessment of accuracy.
• Area Under Curve (AUC): Evaluation of method performance across all possible classification thresholds.
• Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI): Metrics for quantifying accuracy in subclone identification.
• Runtime and Memory Requirements: Practical considerations for implementation scalability.

Tumor Purity Simulation and Dilution Experiments

To systematically evaluate purity-dependent performance, researchers employ both experimental and computational approaches to control tumor purity [37]:

• Physical Sample Mixing: Precisely controlled mixtures of cancer cell lines and normal cells are created to simulate different purity levels (e.g., 10%, 30%, 50%, 70%, 90% tumor content).

• Computational Dilution: In silico downsampling of sequencing reads from pure tumor samples mixed with normal cell data to simulate varying purity conditions.

• Precision Microdissection: Physical isolation of tumor-rich regions from tissue sections to create high-purity samples from otherwise heterogeneous tissues.

For the physical mixture approach, studies often use well-characterized cancer cell lines (e.g., HCC1395, MCF7, A375) mixed with normal lymphoblastoid cells or peripheral blood mononuclear cells (PBMCs) in predetermined ratios [3] [2]. These mixtures are then processed through scRNA-seq protocols, with the resulting data serving as ground-truth validated test sets for CNV detection algorithms.

The Scientist's Toolkit: Essential Research Reagent Solutions

Table 3: Key Research Reagents and Computational Tools for CNV Detection Studies

Resource Category	Specific Examples	Application Context	Performance Considerations
Reference Datasets	Healthy PBMCs, matched normal tissues	Normalization baseline for CNV callers	Same-sample references optimal; cross-tissue references reduce performance
CNV Calling Algorithms	InferCNV, CopyKAT, CaSpER, Numbat, SCEVAN, CONICSmat	scRNA-seq CNV detection	Allele-integrated methods more robust to reference choice
Validation Technologies	(sc)WGS, WES, chromosomal microarray, karyotyping	Ground truth establishment	(sc)WGS considered gold standard but not always available
Cell Line Models	HCC1395, COLO320, HCT116, MCF7, A375, PBMCs	Method benchmarking and validation	Well-characterized CNV profiles enable accuracy assessment
Single-Cell Platforms	10x Genomics, Fluidigm C1, ICELL8, SMART-seq2	scRNA-seq data generation	Platform choice affects sensitivity; full-length better for fusion detection

The selection of appropriate reference datasets emerges as particularly critical for accurate CNV detection in low-purity samples [3]. When available, normal cells from the same sample provide the optimal reference. If external references must be used, cell-type matching becomes essential, with methods like Numbat and CaSpER demonstrating greater robustness to reference choice due to their incorporation of allelic information [3].

For bulk sequencing analyses, the AITAC algorithm provides an alternative approach that utilizes regions with copy number deletions to model the non-linear relationship between tumor purity and observed read depths, without requiring predetermined mutation genotypes [40]. This method can infer tumor purity and absolute copy numbers simultaneously, offering a different strategy for addressing tumor heterogeneity challenges.

The comprehensive benchmarking of CNV detection methods reveals that addressing tumor purity challenges requires method selection tailored to specific experimental conditions and sample characteristics. Based on current evidence, the following recommendations emerge for researchers working with heterogeneous tumor samples:

• For samples with expected low tumor purity (<30%), prioritize methods that incorporate allelic information (Numbat, CaSpER) as they demonstrate greater robustness to purity challenges and reference selection issues [3].

• When working with suboptimal reference datasets, allele-aware methods again outperform expression-only approaches, maintaining better accuracy when matched normal references are unavailable [3].

• For subclone identification in moderately pure samples, inferCNV and CopyKAT provide the most accurate population discrimination, particularly when analyzing data from a single platform [2].

• Always aim for tumor purity >30% through precision sampling or enrichment techniques, as below this threshold all methods exhibit significantly degraded performance [37].

• Validate critical CNV findings with orthogonal methods when working with low-purity samples, especially for potential clinical applications [37] [39].

As single-cell genomics continues to advance, the development of more robust algorithms that specifically address the tumor purity challenge remains crucial. Future method development should focus on integrating multi-omic signals, improving reference-free normalization approaches, and leveraging machine learning techniques to distinguish true biological signals from purity-related artifacts. Until then, the careful application of current methods with awareness of their limitations under different purity conditions will provide the most reliable path forward for CNV detection in heterogeneous samples.

In the field of single-cell genomics, the computational demand for copy number variation (CNV) detection is a critical practical consideration. Researchers and clinicians must balance the need for precise, reliable calls against the constraints of available computing infrastructure and project timelines. This guide provides an objective comparison of the runtime and memory performance of leading single-cell CNV detection algorithms, based on recent, comprehensive benchmarking studies. Understanding these computational profiles enables scientists to select the optimal tool for their specific data type and experimental goals, ensuring efficient resource allocation without compromising analytical integrity [3] [2].

Performance Comparison of scCNV Detection Tools

Quantitative Runtime and Memory Profiles

The computational cost of scCNV callers varies significantly based on their underlying algorithms. Methods that incorporate allelic information generally provide robust performance for large droplet-based datasets but require substantially higher runtime and memory [3]. The following table summarizes the quantitative performance metrics for the most widely used tools.

Table 1: Computational Performance of scRNA-seq CNV Callers

Method	Computational Approach	Runtime Profile	Memory Requirements	Key Computational Notes
InferCNV	HMM (expression-based)	Moderate	Moderate	---
CopyKAT	Statistical segmentation	Moderate	Moderate	---
SCEVAN	Segmentation	---	---	---
CONICSmat	Mixture Model	---	---	Per-chromosome-arm resolution [3]
CaSpER	HMM (expression + allelic frequency)	Higher	Higher	Integrates multi-scale smoothing [3] [2]
Numbat	HMM (expression + allelic frequency)	Higher	Higher	Includes iterative phylogeny reconstruction [3] [41]

Performance in Context: Key Findings from Benchmarking Studies

A large-scale benchmarking study evaluating six popular methods on 21 scRNA-seq datasets confirmed that the integration of allelic information, as performed by CaSpER and Numbat, comes with a tangible computational cost. Despite this, their performance is often more robust for large, complex droplet-based datasets [3]. A separate independent study published in Precision Clinical Medicine in June 2025 also identified CaSpER and CopyKAT as top performers for overall CNV inference accuracy, though their effectiveness can be influenced by sequencing depth and platform [2] [4].

Experimental Protocols for Benchmarking

Benchmarking Framework and Methodology

To ensure fair and reproducible comparisons, recent benchmarking studies have adopted rigorous experimental protocols. The core methodology involves executing each CNV calling tool on a common set of well-characterized datasets and evaluating the results against orthogonal ground truth data.

Table 2: Essential Research Reagents and Computational Solutions

Item Name	Function/Description	Example/Application in Benchmarking
Reference scRNA-seq Datasets	Provide standardized input for comparing algorithm performance.	21 datasets including cancer cell lines (gastric, COLO320, MCF7) and primary tumors (ALL, BCC, MM) [3].
Ground Truth CNV Data	Orthogonal measurements to validate computational predictions.	(sc)WGS or WES data from the same sample or cell line [3] [2].
Diploid Reference Cells	Required for expression normalization by all methods.	Healthy cells from the same sample; external datasets (e.g., PBMCs) for cell lines [3].
Benchmarking Pipeline	A reproducible workflow to run and evaluate multiple tools.	A publicly available Snakemake pipeline for standardized testing [3].
High-Performance Computing (HPC) Cluster	Infrastructure to execute computationally intensive tasks.	Necessary for running methods like CaSpER and Numbat, especially on large datasets [3].

The general workflow can be summarized in the following diagram:

Key Experimental Variables

Benchmarking studies systematically test performance under different conditions to provide comprehensive guidance [3] [2]:

• Dataset Size and Technology: Tools are evaluated on both droplet-based (e.g., 10x Genomics) and plate-based (e.g., SMART-seq2) platforms with varying numbers of cells.
• Reference Dataset Choice: The selection of diploid reference cells for normalization is a critical step that significantly impacts the results and computational efficiency.
• Sequencing Depth: Performance is assessed across a range of sequencing depths to determine the trade-off between data quality and computational cost.

Analysis of Computational Characteristics

The computational differences between tools are not arbitrary but stem from their fundamental algorithmic designs. The following diagram illustrates the relationship between methodological approaches and their resulting computational profiles.

Strategic Selection Guide

Choosing the appropriate tool requires aligning methodological strengths with experimental goals and constraints [3] [2] [4]:

• For large droplet-based datasets (e.g., 10x Genomics) where accuracy is paramount and computational resources are available, allelic-frequency methods (CaSpER, Numbat) are preferable despite their higher resource demands.
• For standard-resolution subclone identification without extreme resource requirements, expression-based HMM and segmentation methods (InferCNV, CopyKAT) offer a balanced compromise.
• When working with limited computational resources or when analyzing euploid samples, simpler approaches may be more practical, though users should verify performance on diploid controls.

Runtime and memory considerations are central to the effective application of single-cell CNV detection tools in modern research. While methods integrating allelic information like CaSpER and Numbat often demonstrate superior performance, this comes with a substantial computational cost. Conversely, expression-based tools such as InferCNV and CopyKAT provide a more computationally efficient option while still delivering robust results for many applications. There is no universally superior tool; the optimal choice depends on the specific experimental context, including dataset size, available computational resources, and the primary biological question. Researchers are encouraged to use public benchmarking pipelines to validate tool performance on their own data types before committing to large-scale analyses.

The accurate detection of copy number variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data has emerged as a powerful approach for deciphering tumor heterogeneity, detecting rare subclones, and reconstructing cancer evolutionary lineages [3] [2]. However, the performance of scCNV detection tools varies dramatically depending on data type, experimental design, and most critically, the strategies employed for parameter tuning and biological validation [3] [4]. Establishing biological ground truth represents the fundamental challenge in benchmarking these computational methods, as it directly determines the reliability of performance metrics and practical recommendations for researchers.

Numerous scCNV calling methods have been developed, employing diverse computational strategies ranging from hidden Markov models and segmentation approaches to more recent deep learning and reinforcement learning frameworks [3] [42] [43]. These tools can be broadly categorized into expression-based methods (InferCNV, CopyKAT, SCEVAN, CONICSmat) that utilize only gene expression information, and integrated methods (CaSpER, Numbat, HoneyBADGER) that combine expression data with allelic information from single nucleotide polymorphisms (SNPs) [3] [2]. A third category of DNA-based methods (SCYN, HiScanner, CNRein) utilizes single-cell DNA sequencing data specifically designed for CNV detection [6] [44] [43]. Each category presents distinct advantages and limitations for parameter optimization and validation against biological ground truth.

This guide provides a comprehensive comparison of experimental strategies for validating scCNV detection algorithms, synthesizing insights from major benchmarking studies to establish best practices for the field. We focus specifically on experimental designs for establishing ground truth, parameter tuning strategies across methodological approaches, and quantitative performance assessments to guide researchers in selecting and optimizing CNV detection tools for their specific biological applications.

Establishing Biological Ground Truth: Experimental Strategies and Protocols

Orthogonal Validation Technologies

Establishing reliable biological ground truth requires validation against orthogonal technologies that directly measure genomic alterations. The table below summarizes the primary experimental approaches used in major benchmarking studies for validating scCNV calls.

Table 1: Orthogonal Technologies for Validating scCNV Calls

Validation Method	Resolution	Key Applications	Advantages	Limitations
Single-cell Whole Genome Sequencing (scWGS)	High (varies with coverage)	Gold standard for cell-by-cell comparison [3]	Direct measurement of CNVs at single-cell level	Not measured in same cell as scRNA-seq
Bulk Whole Genome Sequencing (WGS)	High	Providing pseudobulk ground truth [3] [2]	Comprehensive genomic coverage	Masks cellular heterogeneity
Whole Exome Sequencing (WES)	Moderate (exonic regions)	Validation of coding region CNVs [3]	Cost-effective for focused validation	Limited to exonic regions
Chromosomal Microarray Analysis	Moderate	Clinical validation standard [7] [6]	Well-established clinical standard	Limited resolution for small CNVs
Array Comparative Genomic Hybridization (aCGH)	Moderate	Historical gold standard [6]	Quantitative, reproducible	Being replaced by sequencing methods
G-banding Karyotyping	Low (chromosomal)	Detection of large-scale alterations [7]	Visualizes entire genome	Very low resolution

Reference Datasets with Established Ground Truth

Benchmarking studies have established several reference datasets with orthogonal validation that serve as community resources for method development and evaluation:

• Cancer cell lines: Well-characterized cancer cell lines (e.g., gastric, colorectal, breast cancer lines) with known CNV profiles validated through multiple technologies [3]. These provide controlled systems with defined expectations for CNV patterns.

• Mixed cell line experiments: Artificially mixed samples of multiple cancer cell lines (e.g., 3-5 lung adenocarcinoma lines) that mimic tumor subpopulations with known proportions [2] [4]. These enable precise evaluation of subclone identification accuracy.

• Patient-derived samples with matched normal tissue: Clinical samples with adjacent normal tissue that provides reference diploid cells [2] [44]. These represent real-world complexity but with built-in controls.

• Synthetic datasets: Computationally generated CNV profiles with introduced biological and technical noise [6] [43]. These enable systematic evaluation of specific performance parameters with complete knowledge of ground truth.

The following diagram illustrates a comprehensive validation workflow integrating multiple orthogonal approaches:

Figure 1. Comprehensive workflow for validating scCNV detection algorithms through orthogonal technologies. scRNA-seq data undergoes computational CNV calling, with results validated against multiple orthogonal technologies to generate comprehensive performance metrics.

Performance Benchmarking: Quantitative Comparisons Across Methods

Method Categories and Computational Approaches

scCNV detection methods employ diverse computational frameworks, each with distinct strengths and limitations for parameter tuning and optimization:

Table 2: scCNV Method Categories and Characteristics

Method Category	Representative Tools	Core Algorithm	Data Requirements	Parameter Sensitivity
Expression-based	InferCNV [3], CopyKAT [2], SCEVAN [3], CONICSmat [3]	HMM, segmentation, mixture models	scRNA-seq + reference cells	High - reference selection critical
Integrated (Expression + Allelic)	CaSpER [3], Numbat [3], HoneyBADGER [2]	HMM with B-allele frequency	scRNA-seq with SNP information	Medium - multiple data types
DNA-based	SCYN [6], HiScanner [44], CNRein [43]	Dynamic programming, joint segmentation, reinforcement learning	scDNA-seq	Low-medium - optimized for DNA
Deep Learning	RCANE [42]	Graph neural networks + LSTM	Bulk RNA-seq	Low - learned representations

Quantitative Performance Metrics

Benchmarking studies employ multiple quantitative metrics to evaluate different aspects of CNV detection performance. The most comprehensive evaluations assess: (1) CNV prediction accuracy against ground truth, (2) ability to identify euploid cells and samples, (3) subclone identification accuracy, and (4) computational efficiency [3].

Table 3: Key Performance Metrics for scCNV Method Evaluation

Performance Category	Specific Metrics	Interpretation	Optimal Range
CNV Prediction Accuracy	AUC/Partial AUC [3], Sensitivity/Specificity [2], F1-score [3]	Ability to correctly identify gained/lost regions	>0.8 (excellent)
Subclone Identification	Adjusted Rand Index (ARI) [2], Fowlkes-Mallows index (FM) [2], Normalized Mutual Information (NMI) [2]	Concordance with known cell lineages	0-1 (higher better)
Euploid Detection	Mean square error deviation [3]	Deviation from diploid baseline	Lower values better
Computational Efficiency	Runtime, Memory usage [3]	Practical scalability	Application-dependent

Comparative Performance Across Methods

Recent large-scale benchmarking studies provide comprehensive performance comparisons across scCNV detection methods. The table below synthesizes findings from evaluations across multiple datasets and experimental conditions:

Table 4: Comparative Performance of scCNV Detection Methods

Method	CNV Prediction Accuracy	Subclone Identification	Euploid Detection	Computational Efficiency	Key Strengths
CaSpER	High [2] [4]	Medium [2]	Moderate [3]	Medium runtime [3]	Robust for large datasets, integrates allelic information [3]
CopyKAT	High [2] [4]	High [2] [4]	Moderate [3]	Fast [3]	Excellent for subclone identification, works well on single-platform data [2]
InferCNV	Medium [3]	High [2] [4]	Poor on euploid samples [3]	High memory requirements [3]	Sensitive for rare populations, effective with sufficient cells [4]
SCEVAN	Medium [3]	Medium [3]	Not reported	Medium runtime [3]	Segmentation-based approach
Numbat	Medium [3]	High [3]	Not reported	High runtime [3]	Groups cells into subclones, uses allelic information [3]
HoneyBADGER	Low-Medium [2]	Low [2]	Not reported	Medium runtime	Allele-based version resilient to batch effects [4]
sciCNV	Low-Medium [2]	Medium [2]	Not reported	Not reported	Performance affected by batch effects [2]

Parameter Tuning Strategies: Method-Specific Considerations

Critical Parameters Across Method Categories

Each category of scCNV detection methods requires optimization of distinct parameter sets. The following diagram illustrates key parameter tuning considerations across the major methodological approaches:

Figure 2. Parameter tuning considerations across major scCNV method categories. Each methodological approach requires optimization of distinct parameters, with all approaches ultimately requiring validation against biological ground truth.

Reference Selection Strategies

For expression-based methods, reference selection represents the most critical parameter influencing performance [3]. Multiple strategies exist for selecting appropriate diploid reference cells:

• Annotation-based references: Using manually annotated normal cell types from the same sample as reference [3]. This approach provides the most biologically accurate normalization but requires high-quality cell type annotations.

• Automatic reference detection: Leveraging built-in functionality in methods like CopyKAT and SCEVAN to automatically identify normal cells [3]. Performance varies significantly across methods and datasets.

• External reference datasets: Employing matched external reference datasets with healthy cells from similar tissue types when no normal cells are available in the sample [3]. This approach is common for cancer cell line studies but introduces additional technical variability.

• Reference-free approaches: Utilizing methods like RCANE that don't require explicit reference samples [42]. These approaches are particularly valuable when no suitable reference exists.

Benchmarking studies have demonstrated that dataset-specific factors—including dataset size, the number and type of CNVs present, and reference dataset choice—significantly influence optimal parameter settings [3]. Methods incorporating allelic information (e.g., CaSpER, Numbat) generally perform more robustly for large droplet-based datasets but require higher computational runtime [3].

Addressing Batch Effects and Technical Variability

Batch effects significantly impact the performance of most scCNV inference methods, particularly for subclone identification across multiple platforms [2]. Benchmarking studies reveal that:

• Expression-based CNV inference methods (InferCNV, CaSpER, sciCNV, CopyKAT) are highly affected by batch effects when estimating tumor subpopulations using datasets derived from multiple platforms [2].

• The allele-based version of HoneyBADGER, although less sensitive overall, proves more resilient to batch-related distortions [4].

• Batch correction tools like ComBat can mitigate these effects when analyzing data across platforms [2].

Computational efficiency varies substantially across methods, with runtime differences of multiple orders of magnitude observed in benchmarking studies [3] [6]. This practical consideration becomes critical when analyzing large datasets with thousands of cells.

Table 5: Essential Research Resources for scCNV Validation Studies

Resource Category	Specific Resources	Application	Key Features
Benchmarking Pipelines	Snakemake pipeline [3]	Reproducible method comparison	Standardized evaluation metrics, 21 benchmark datasets
Reference Datasets	TCGA [42], DepMap [42]	Training and validation	Pan-cancer molecular data with clinical annotations
Visualization Tools	ViScanner [44], HiGlass [44]	Exploration of CNV profiles	Interactive genome browsing, multiple resolution levels
Simulation Frameworks	CNAsim [43], SCSsim [6]	Controlled performance evaluation	Ground truth knowledge, customizable noise parameters
Batch Correction Tools	ComBat [2]	Multi-platform integration	Addresses technical variability across datasets
Phasing Tools	SHAPE-IT 4 [43]	Haplotype reconstruction	Enables allele-specific analysis

Establishing biological ground truth through rigorous parameter tuning and validation remains challenging yet essential for advancing scCNV detection algorithms. Benchmarking studies consistently demonstrate that method performance is highly context-dependent, with optimal tool selection influenced by dataset size, sequencing technology, CNV characteristics, and available reference data [3] [2].

The most robust validation strategies employ multiple orthogonal technologies, with scWGS emerging as the gold standard for cell-level validation [3] [44]. For expression-based methods, reference selection represents the most critical parameter, while DNA-based methods require careful optimization of evolutionary constraints and joint segmentation parameters [3] [43].

Future methodological developments should focus on improved batch effect correction, more efficient utilization of allelic information, and evolution-aware approaches that incorporate biological constraints to reduce spurious CNA calls [2] [43]. As single-cell genomics continues its transition from basic research to clinical applications, rigorous validation against biological ground truth will remain paramount for deriving biologically and clinically meaningful insights from scCNV data.

Head-to-Head Performance Benchmark: Sensitivity, Specificity, and Clinical Validation

The accurate detection of Copy Number Variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data is crucial for understanding genetic heterogeneity in cancer and other diseases. As tumors are composed of genetically distinct cellular subpopulations, single-cell technologies offer the unique ability to capture within-sample heterogeneity of CNVs and identify subclones relevant for tumor progression and treatment outcome [3]. Several computational tools have been developed to infer CNVs from scRNA-seq data, creating a critical need for standardized benchmarking frameworks to evaluate their performance [3] [2]. This guide provides a comprehensive overview of the evaluation metrics, ground truth establishment methods, and experimental protocols used in benchmarking single-cell CNV detection algorithms, offering researchers a structured approach for comparative tool assessment.

Core Evaluation Metrics for scCNV Detection

Benchmarking studies employ multiple metrics to comprehensively evaluate different aspects of scCNV detection tool performance, ranging from overall accuracy to clonal identification capability.

CNV Calling Accuracy Metrics

The accuracy of CNV detection is typically evaluated using threshold-independent and threshold-dependent metrics that compare computational predictions against established ground truth data.

• Threshold-Independent Metrics: These include correlation analysis between predicted and expected CNV signals, and Area Under the Curve (AUC) values derived from Receiver Operating Characteristic (ROC) analysis. AUC scores are often calculated separately for gains versus all and losses versus all [3]. Some studies employ partial AUC, which focuses on biologically meaningful threshold ranges, where values below 0.5 indicate that most thresholds fall outside the meaningful value range [3].

• Threshold-Dependent Metrics: After establishing optimal gain and loss thresholds based on multi-class F1 scores, researchers calculate sensitivity (recall), specificity, precision, and F1 scores (the harmonic mean of precision and recall) [3] [9]. These metrics provide practical insights into tool performance under real-world usage scenarios where specific thresholds must be applied.

Subclone Identification Metrics

For methods that group cells into subclones with similar CNV profiles, clustering performance is evaluated using metrics that compare computational groupings to known cell line identities:

• Adjusted Rand Index (ARI): Measures the similarity between two data clusterings, adjusted for chance [2].
• Fowlkes-Mallows Index (FM): Defined as the geometric mean between precision and recall for pair-wise assignments [2].
• Normalized Mutual Information (NMI): Measures the mutual information between two clusterings, normalized by the entropy of each [2].
• V-Measure: An entropy-based metric that evaluates both homogeneity (how pure clusters are) and completeness (how well all similar cells are grouped together) [2].

Table 1: Summary of Key Evaluation Metrics for scCNV Detection

Metric Category	Specific Metric	Evaluation Focus	Interpretation
CNV Calling Accuracy	Correlation	Agreement between predicted and ground truth CNV signals	Higher values indicate better concordance
	AUC/Partial AUC	Overall discriminative ability for gains and losses	Values >0.5 indicate predictive power better than chance
	Sensitivity/Recall	Ability to correctly identify true CNV events	Higher values indicate fewer false negatives
	Specificity	Ability to correctly identify non-CNV regions	Higher values indicate fewer false positives
	F1 Score	Balance between precision and recall	Higher values indicate better overall accuracy
Subclone Identification	Adjusted Rand Index (ARI)	Similarity between computational and known groupings	Values range from 0 (random) to 1 (perfect match)
	Normalized Mutual Information (NMI)	Agreement between two clusterings	Higher values indicate better alignment
	V-Measure	Balance between homogeneity and completeness	Higher values indicate better clustering quality
Technical Performance	Runtime	Computational efficiency	Lower values indicate faster performance
	Memory Usage	Resource requirements	Lower values indicate better scalability

Establishing Ground Truth for Validation

A critical challenge in benchmarking scCNV callers is establishing reliable ground truth data for validation. Different approaches have been developed depending on the sample type and available orthogonal data.

Cell Line-Based Ground Truth

Cell lines with well-characterized genomic profiles provide a controlled environment for method validation:

• Paired Tumor-Normal Systems: The HCC1395 breast cancer cell line and matched "normal" B lymphocyte cell line (HCC1395BL) from the same donor provides a well-controlled system for evaluating sensitivity and specificity [2]. This approach allows researchers to use the normal B lymphocyte cells as reference for CNV calling algorithms.

• Artificial Mixtures: Mixed samples consisting of three or five human lung adenocarcinoma cell lines (e.g., A549, H1650, H1975, H2228, and HCC827) mimic tumor subpopulations with known proportions, enabling accurate assessment of subclone identification performance [2].

Orthogonal Genomic Validation

For primary tissue samples where no reference cell lines exist, ground truth is established through orthogonal genomic measurements:

• Whole Genome Sequencing (WGS): Both bulk and single-cell WGS provide high-confidence CNV calls that serve as validation standards. In one benchmarking study, bulk WGS was performed on fresh frozen tissues with 30x coverage, while single-cell whole exome sequencing (scWES) was conducted on individual cells [2].

• Whole Exome Sequencing (WES): When WGS is not available, WES data can provide validation for coding regions, though with limited coverage of non-coding regions [3].

• Third-Generation Sequencing: Long-read technologies like PacBio sequencing can generate a comprehensive set of CNVs, though these often require filtering and consolidation for comparison with scRNA-seq derived calls [39].

Validation Data Processing

To enable comparison between scRNA-seq predictions and ground truth, several processing steps are typically required:

• Pseudobulk Profiles: When ground truth is not measured in the same cells as scRNA-seq, per-cell results from scRNA-seq methods are combined into an average CNV profile before comparison [3].

• Reciprocal Overlap Criteria: A common approach requires ≥50% reciprocal overlap between predicted CNVs and ground truth events for validation, though less stringent criteria may be applied to rescue partially overlapping calls [45].

• Region Limitation: Since scRNA-seq methods only predict CNV status for genomic regions containing genes, comparisons are limited to gene regions even when WGS ground truth covers nearly the complete genome [3].

Experimental Protocols for Benchmarking scCNV Callers

A comprehensive benchmarking study involves systematic evaluation across multiple datasets and experimental conditions to assess tool robustness and limitations.

Core Benchmarking Workflow

The following diagram illustrates the standard workflow for benchmarking single-cell CNV detection methods:

Key Experimental Considerations

• Reference Selection: All scRNA-seq CNV methods require a set of euploid reference cells for normalization. For primary tissues, manually annotated healthy cells should be used consistently across methods. For cancer cell lines without matched reference, external datasets with healthy cells from similar cell types must be carefully selected [3].

• Parameter Settings: Tools should be run as recommended in their respective tutorials or using default parameters to ensure fair comparison. Some benchmarking studies also evaluate the impact of parameter modifications on performance [3] [46].

• Platform Variability: Evaluation should include data from multiple scRNA-seq platforms (e.g., 10x Genomics, Fluidigm C1, ICELL8, CEL-seq2, Drop-seq) with varying sequencing depths and read lengths to assess platform-specific performance [2].

• Batch Effect Assessment: Combining datasets across different platforms introduces batch effects that significantly impact subclone identification accuracy. Methods should be tested with and without batch effect correction tools like ComBat [2].

The Scientist's Toolkit: Essential Research Reagents

Table 2: Key Research Reagents and Computational Tools for scCNV Benchmarking

Category	Item	Function in Benchmarking
Reference Cell Lines	HCC1395/HCC1395BL pair	Paired tumor-normal system for sensitivity/specificity testing
	Lung adenocarcinoma lines (A549, H1650, H1975, H2228, HCC827)	Artificial tumor mixtures for subclone identification assessment
	NA12878 cell line	Well-characterized genome for validation against gold standard sets
Computational Tools	InferCNV	Expression-based CNV caller using HMM; groups cells into subclones
	CopyKAT	Statistical model using segmentation approach; reports per-cell results
	CaSpER	Integrates expression with allele frequency; uses multiscale smoothing and HMM
	HoneyBADGER	Bayesian framework integrating HMM; uses expression and allele information
	SCEVAN	Segmentation-based approach; groups cells into subclones
	Numbat	Combines expression with allele frequency; uses HMM; groups cells
Benchmarking Resources	Biodiscovery Nexus Software	Platform-agnostic CNV analysis for cross-platform comparisons
	CNVbenchmarker2	Custom framework for performing tool evaluations
	Reproducible Snakemake Pipeline	Automated benchmarking workflow for consistent tool assessment

Robust benchmarking of single-cell CNV detection methods requires a multifaceted approach incorporating diverse evaluation metrics, carefully established ground truth, and systematic experimental protocols. The field has moved beyond simple accuracy measurements to comprehensive assessments that consider biological context, technical variability, and practical usability. Recent studies have revealed that method performance varies significantly based on dataset-specific factors including dataset size, the number and type of CNVs in the sample, sequencing depth, and the choice of reference dataset [3] [2]. No single method outperforms others across all scenarios, highlighting the importance of context-specific tool selection. As the field evolves, standardized benchmarking frameworks will continue to play a crucial role in guiding researchers toward optimal method selection and driving improvements in future algorithm development.

Copy number variations (CNVs) are crucial genomic alterations in cancer, driving tumor evolution and heterogeneity. The advent of single-cell RNA sequencing (scRNA-seq) has enabled the inference of CNVs from transcriptomic data, allowing researchers to explore genetic diversity at cellular resolution. Several computational methods have been developed for this purpose, but their performance varies significantly across different experimental conditions. This guide objectively compares leading scRNA-seq CNV detection tools, with a specific focus on their performance across cancer cell lines, where CaSpER and CopyKAT have demonstrated superior balanced accuracy according to multiple independent benchmarking studies [10] [2] [4].

Method Comparison at a Glance

Table 1: Overview of scRNA-seq CNV Detection Methods

Method	Underlying Approach	Data Requirements	Key Strengths	Performance in Cancer Cell Lines
CaSpER	Multiscale signal processing	Expression + Allelic frequency	Balanced CNV inference, integrates dual signals	Top performer in balanced accuracy [10] [2]
CopyKAT	Bayesian segmentation	Expression matrix	Tumor subpopulation identification, fast runtime	Excellent overall performance, particularly in subclone identification [2] [17]
Numbat	Haplotype-aware HMM	Expression + Allelic + Haplotype	Best tumor-normal classification, cnLOH detection	Excels in imbalanced tumor-normal ratios [17]
InferCNV	Hidden Markov Model	Expression matrix	Sensitive subclone detection, widely adopted	Strong sensitivity with sufficient cells sequenced [10] [2]
SCEVAN	Multi-channel segmentation	Expression matrix	Clonal breakpoint detection	Performance improves with TME cells included [17]
HoneyBADGER	HMM + Bayesian approach	Expression ± Allelic	Resilient to batch effects	Less sensitive overall [2]

Quantitative Performance Metrics

Table 2: Performance Metrics Across Cancer Cell Line Studies

Method	Sensitivity	Specificity	Tumor-Normal Classification Accuracy	Subclone Identification	Platform Robustness
CaSpER	High	High	Moderate	Moderate	High across multiple platforms [2]
CopyKAT	High	High	High (F1 score: 0.81-0.99 in most samples) [17]	Excellent (with inferCNV) [2]	Moderate (affected by batch effects) [2]
Numbat	High	High	Best overall (superior F1 scores) [17]	Good	Varies with sequencing depth [17]
InferCNV	High with sufficient cells [10]	Moderate	Variable (improves with TME cells) [17]	Excellent (with CopyKAT) [2]	Low (highly affected by batch effects) [2]
SCEVAN	Moderate	Moderate	Poor with low tumor purity [17]	Good	Moderate

Experimental Protocols in Benchmarking Studies

Dataset Composition and Ground Truth Establishment

Benchmarking studies utilized scRNA-seq datasets with orthogonal CNV validation to establish reliable performance metrics [3] [2]. The experimental workflow typically involved:

• Cell Line Selection: Studies employed various cancer cell lines including:

  • Breast cancer cell lines (HCC1395, HCC1954) [47] [2]
  • Gastric cancer cell lines (9 different lines) [3]
  • Colorectal adenocarcinoma lines (COLO320, HCT116) [3]
  • Lung adenocarcinoma cell lines in mixed samples [2]

• Reference Datasets: Matched normal B-lymphocyte cell lines (HCC1395BL) or healthy donor cells provided diploid references for normalization [2].

• Platform Diversity: Data spanned multiple scRNA-seq technologies including 10x Genomics, Fluidigm C1, ICELL8, and SMART-seq2 to assess platform-specific performance [3] [2].

• Ground Truth Establishment: Orthogonal measurements from single-cell or bulk whole-genome sequencing (scWGS/WGS) and whole-exome sequencing (WES) provided validation data [3] [17].

Performance Evaluation Metrics

Comprehensive benchmarking studies employed multiple quantitative metrics to assess method performance [3] [2] [17]:

• CNV Detection Accuracy: Measured via sensitivity, specificity, and area under the curve (AUC) values comparing predictions to ground truth CNVs.

• Tumor-Normal Classification: F1 scores, precision, and recall for distinguishing tumor cells from normal cells.

• Subclone Identification: Metrics including Adjusted Rand Index (ARI), Fowlkes-Mallows index (FM), Normalized Mutual Information (NMI), and V-Measure to quantify clustering accuracy.

• Computational Efficiency: Runtime and memory requirements under standardized conditions.

Factors Influencing Method Performance

Impact of Technical Variables

Table 3: Impact of Experimental Conditions on Method Performance

Factor	Impact on Performance	Best-Performing Methods
Sequencing Depth	Decreased performance with lower depth (<3k UMIs/cell) [17]	CopyKAT, CaSpER most robust
Reference Selection	Critical for normalization; internal references optimal [3]	Numbat, CaSpER less dependent on reference quality
Tumor Purity	High tumor purity challenges expression-based methods [17]	Numbat consistently outperforms across purity levels
Batch Effects	Significant impact on cross-platform analyses [2]	HoneyBADGER most resilient
Platform Type	Performance varies between full-length and 3'-end protocols [2]	CaSpER, CopyKAT most platform-agnostic

Table 4: Key Research Reagent Solutions for scCNV Analysis

Reagent/Resource	Function	Application Notes
Reference Cell Lines	Diploid baseline for normalization	Matched tissue type improves performance [3]
10x Genomics Chromium	Single-cell partitioning	High-throughput CNV profiling [2]
SMART-seq2 Reagents	Full-length transcriptome	Enhanced detection of allele-specific CNVs [2]
BAFExtract Tool	Allele frequency extraction	Required for CaSpER analysis [18]
Population Haplotypes	Phasing information	Essential for Numbat and allele-aware methods [17]
CellSnp-lite	SNP genotyping from scRNA-seq	Preprocessing for allele-aware methods [48]

Discussion and Research Implications

The consistent outperformance of CaSpER and CopyKAT across cancer cell lines highlights their robustness in controlled experimental settings. CaSpER's integration of gene expression with allelic frequency information provides more comprehensive CNV profiling, while CopyKAT's efficient Bayesian approach offers reliable results with computational efficiency [10] [2].

For research applications focusing on cancer cell lines, the choice between these methods depends on specific experimental needs:

• CaSpER is preferable when allelic information is available and comprehensive CNV characterization is required.
• CopyKAT provides the best balance of accuracy and efficiency for standard expression-based CNV detection.
• Numbat should be considered when working with complex samples containing mixed tumor-normal populations or when detecting copy-number neutral LOH events [17].

These benchmarking results provide solid guidance for researchers selecting computational tools for scCNV analysis in cancer cell lines, though method performance should be validated for specific experimental systems and biological questions.

Copy number variations (CNVs)—genomic alterations involving the gain or loss of DNA segments—are fundamental drivers of tumor evolution and therapeutic resistance [4]. The accurate identification of cellular subpopulations defined by distinct CNV profiles, a process known as subclone identification, is crucial for understanding cancer progression and developing targeted treatments. Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful technology enabling researchers to infer CNVs indirectly from transcriptomic data, thereby linking genetic alterations with cellular phenotypes at unprecedented resolution [19] [3]. However, the computational inference of CNVs from scRNA-seq data presents significant challenges due to technical noise, sparse data, and complex normalization requirements.

Within the landscape of computational tools developed for this purpose, benchmarking studies have consistently identified inferCNV and CopyKAT as top performers in discriminating tumor subpopulations [2] [4] [18]. This guide provides an objective comparison of these and other prominent scRNA-seq CNV detection methods, focusing specifically on their subclone identification capabilities based on recent rigorous benchmarking evidence. We present quantitative performance data, detailed experimental methodologies, and practical guidance to assist researchers in selecting and implementing the optimal tool for their specific research context.

Performance Comparison of scRNA-seq CNV Callers

Quantitative Performance Metrics Across Methodologies

Comprehensive benchmarking studies evaluating six popular computational methods on 21 scRNA-seq datasets have revealed dramatic differences in performance, particularly for subclone identification tasks [3] [2] [18]. The table below summarizes the key performance metrics for each method, with a special emphasis on subclone identification capabilities:

Table 1: Comprehensive Performance Comparison of scRNA-seq CNV Callers

Method	Subclone Identification Performance	Technology Approach	Reference Dependency	Runtime & Resources	Best Use Cases
inferCNV	Excellent (Top performer with CopyKAT) [2]	Expression-based HMM [3]	High [18]	High runtime requirements [18]	Single-platform studies; well-annotated samples [4]
CopyKAT	Excellent (Top performer with inferCNV) [2] [4]	Statistical model with segmentation [3]	Moderate	Fast runtime, low memory [18]	Large droplet-based datasets; subclone discrimination [2]
Numbat	Good [3] [18]	Combines expression + allele frequency [3]	Low (robust to reference choice) [18]	Long runtime requirements [18]	Datasets with SNP information; reference-limited scenarios [3]
SCEVAN	Good (automatic tumor cell detection) [18]	Segmentation approach [19] [3]	Moderate	Moderate requirements [18]	Automated analysis pipelines; large sample sizes [19]
CaSpER	Limited (poor subclone identification) [2] [18]	Expression + allele frequency with HMM [3]	Moderate	Long runtime requirements [18]	CNV inference rather than subcloning [2]
CONICSmat	Limited (poor subclone identification) [18]	Mixture model [3]	High	Fast runtime, low memory [18]	Chromosome-arm level analysis [3]

When evaluating overall CNV detection accuracy rather than subclone identification specifically, CaSpER and CopyKAT demonstrate superior balanced performance in sensitivity and specificity [2] [4]. However, for the specific task of discriminating between cellular subpopulations—a critical requirement for understanding tumor heterogeneity—inferCNV and CopyKAT consistently outperform alternative approaches across multiple benchmarking studies [2] [4].

Specialized Metrics for Subclone Identification

In dedicated evaluations using mixed cell line datasets where ground truth subpopulation identities are known, inferCNV and CopyKAT achieve superior performance according to multiple cluster validation metrics:

Table 2: Subclone Identification Accuracy Metrics on Mixed Cell Line Data

Method	Adjusted Rand Index (ARI)	Normalized Mutual Information (NMI)	Fowlkes-Mallows Index (FM)	V-Measure
inferCNV	0.82 [2]	0.85 [2]	0.84 [2]	0.83 [2]
CopyKAT	0.81 [2]	0.83 [2]	0.82 [2]	0.82 [2]
CaSpER	0.45 [2]	0.52 [2]	0.48 [2]	0.51 [2]
sciCNV	0.78 [2]	0.80 [2]	0.79 [2]	0.79 [2]
HoneyBADGER	0.62 [2]	0.65 [2]	0.63 [2]	0.64 [2]

Notably, the performance of all methods is significantly affected by batch effects when analyzing data derived from multiple scRNA-seq platforms [2]. For multi-platform studies, batch effect correction tools such as ComBat must be implemented prior to CNV analysis to maintain subclone identification accuracy [2].

Experimental Protocols for Benchmarking Studies

Standardized Evaluation Workflow

The superior subclone identification capabilities of inferCNV and CopyKAT were established through rigorous benchmarking pipelines that employed multiple validation strategies across different dataset types [3] [2] [18]. The following diagram illustrates the integrated experimental approach used in these comprehensive evaluations:

Diagram 1: scCNV Benchmarking Workflow (87 characters)

Key Experimental Designs

Mixed Cell Line Experiments

To quantitatively evaluate subclone identification accuracy, benchmarking studies employed artificially mixed samples of known human lung adenocarcinoma cell lines (3-5 distinct lines) [2] [4]. These controlled mixtures simulated tumor subpopulations with known ground truth identities, enabling precise calculation of clustering accuracy metrics including Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Fowlkes-Mallows Index (FM) [2]. The datasets were generated using multiple scRNA-seq technologies (10x Genomics, CEL-seq2, and Drop-seq) to assess platform-specific performance [2] [4]. Each CNV calling method was applied to these mixtures, and their cellular partitioning predictions were compared against the known cell line identities to calculate accuracy metrics [2].

Clinical Sample Validation

For real-world validation, researchers generated matched scRNA-seq, single-cell whole exome sequencing (scWES), and bulk whole genome sequencing (WGS) data from a small cell lung cancer patient [2] [4]. This comprehensive dataset included 92 primary tumor cells and 39 relapse tumor cells, enabling orthogonal validation of CNV predictions and subclone identification across disease progression [2]. Methods were evaluated on their ability to consistently group cells from the same pathological stage while discriminating genetically distinct subpopulations, with results validated against scWES and bulk WGS data [2].

Diploid Sample Analysis

To control for false positive CNV calls, all methods were tested on diploid peripheral blood mononuclear cells (PBMCs) from healthy donors [3] [18]. This critical evaluation measured each method's tendency to incorrectly infer CNVs in genetically normal cells, with performance assessed using root mean squared error (RMSE) between CNV predictions and a diploid baseline [18]. This analysis particularly highlighted the robustness of methods incorporating allelic information (Numbat, CaSpER) when appropriate reference datasets were limited [18].

Table 3: Key Research Reagents and Computational Resources for scCNV Analysis

Resource Category	Specific Examples	Function in scCNV Analysis
scRNA-seq Platforms	10x Genomics Chromium, Fluidigm C1, ICELL8, SMART-seq2 [2]	Generate single-cell transcriptome data for CNV inference
Reference Datasets	PBMCs, matched normal tissues, cell type-specific annotations [3] [18]	Provide diploid expression baselines for CNV detection normalization
Bioinformatics Tools	Seurat [19], CellRanger [19], ComBat [2]	Data preprocessing, normalization, and batch effect correction
Benchmarking Pipelines	Snakemake workflow [3] [18], custom R scripts [2]	Standardized method evaluation and performance comparison
Validation Technologies	scWES [2], bulk WGS [2], (sc)WGS [3]	Provide orthogonal ground truth for CNV and subclone identification

Technical Implementation Considerations

Impact of Experimental Conditions on Performance

The subclone identification performance of all CNV callers, including inferCNV and CopyKAT, is significantly influenced by specific dataset characteristics and analytical choices [3]. The following diagram illustrates the key factors affecting method performance:

Diagram 2: Performance Factor Network (78 characters)

Reference Selection Strategies

The choice of reference diploid cells for expression normalization substantially impacts CNV calling accuracy and subsequent subclone identification [3] [18]. When available, normal cells from the same sample (e.g., stromal, immune, or adjacent normal cells) provide the optimal reference [18]. For cancer cell lines or purified tumor samples where internal references are unavailable, carefully matched external datasets (same tissue type, processing protocol, and sequencing platform) are essential [3]. Methods incorporating allelic information (Numbat, CaSpER) demonstrate greater robustness to suboptimal reference selection, particularly for diploid samples [18].

Practical Implementation Guidelines

For optimal subclone identification, researchers should prioritize CopyKAT for large droplet-based datasets (≥ 1,000 cells) due to its computational efficiency and integrated tumor/normal classification [2] [18]. inferCNV is preferable for studies requiring detailed HMM-based CNV state characterization or when analyzing data from full-length transcript protocols [2] [4]. When SNP information is available and reference cells are limited, Numbat provides a robust alternative despite higher computational demands [3] [18]. For all methods, batch effect correction is essential when integrating datasets across multiple platforms or processing batches [2].

Benchmarking evidence consistently identifies inferCNV and CopyKAT as superior tools for discriminating cellular subpopulations based on CNV profiles derived from scRNA-seq data. Their performance advantages are most pronounced in single-platform studies with adequate sequencing depth and appropriate reference cells. As single-cell technologies continue to evolve toward clinical applications, accurate subclone identification will become increasingly critical for understanding tumor evolution, tracking treatment resistance, and identifying novel therapeutic targets. Researchers should select tools based on their specific experimental systems, analytical priorities, and computational resources, while remaining cognizant of the fundamental limitations inherent in inferring DNA-level alterations from transcriptomic data.

In the benchmarking of single-cell copy number variation (CNV) detection algorithms, a critical performance metric is a tool's ability to correctly identify diploid samples without generating false positive calls. False positives—where normal genomic regions are incorrectly flagged as having copy number alterations—can severely impact downstream analyses, including tumor subclonal reconstruction and the identification of driver CNV events. This challenge is particularly acute for scRNA-seq-based CNV callers, which must infer genomic alterations from transcriptomic data, where expression levels are influenced by complex regulatory mechanisms beyond mere copy number [3]. This guide objectively compares the performance of leading single-cell CNV detection methods on diploid samples, providing researchers with experimental data and methodologies essential for selecting appropriate tools.

Performance Comparison of scCNV Detection Tools on Diploid Samples

Quantitative Performance Metrics

The following table summarizes the performance of various single-cell CNV detection tools when applied to diploid samples, specifically peripheral blood mononuclear cells (PBMCs) from a healthy donor, as benchmarked in independent studies.

Table 1: Performance of scCNV Detection Tools on Diploid Samples

Tool	Primary Strategy	Performance on Diploid PBMCs	Reported False Positive Tendency	Key Study
InferCNV	Expression-based HMM	Not evaluated for diploid-specific performance in head-to-head benchmarks	High (in endometrial cancer study)	[5]
CopyKAT	Expression-based segmentation	Moderate sensitivity, overestimation of tumor cells	High (in endometrial cancer study)	[4] [5]
SCEVAN	Expression-based joint segmentation	Moderate sensitivity, overestimation of tumor cells	High (in endometrial cancer study)	[5]
CaSpER	Expression + Allelic Imbalance HMM	Balanced CNV inference, top performer in clinical samples	Lower (compared to expression-only methods)	[3] [4]
Numbat	Expression + Allelic Frequency HMM	Requires evaluation on large datasets; robust performance	Information Not Available	[3]
SCICNV	Expression-based	Does not directly predict tumor cells; CNV score distribution not clearly distinct	Information Not Available	[5]
HoneyBADGER	Allele Frequency & Expression	Less sensitive overall; more resilient to batch effects	Information Not Available	[4]

Key Findings from Benchmarking Studies

• Methodology Matters: Tools that incorporate allelic information (e.g., CaSpER, Numbat) alongside gene expression generally perform more robustly on large, droplet-based datasets, though they often require higher computational runtime [3].
• Overestimation of Aneuploidy: Multiple studies have reported that popular expression-based tools like SCEVAN and CopyKAT exhibit moderate sensitivity but significantly overestimate the number of tumor cells when analyzing real-world data, a key indicator of false positives in mixed samples [5].
• Reference Dataset Dependency: The performance of all methods, including their false positive rates, is highly influenced by dataset-specific factors. The choice of a high-quality diploid reference dataset for normalization is paramount for accurate results [3].
• Platform and Depth Variability: The effectiveness of CNV callers varies with sequencing depth and platform type. For instance, inferCNV and sciCNV excel in distinguishing subclones on a single platform, but their performance can degrade when integrating data across platforms due to batch effects [4].

Experimental Protocols for Assessing False Positives

Core Benchmarking Workflow

A standardized experimental protocol is crucial for the fair assessment of false positive rates in diploid samples. The following diagram illustrates the key steps in this benchmarking workflow.

Detailed Methodological Components

• Ground Truth and Dataset Selection:

  • Diploid scRNA-seq Data: The foundation of this assessment is a diploid scRNA-seq dataset. A common choice is PBMCs from a healthy donor [3] [5]. This dataset should be generated using the same technology (e.g., droplet-based or plate-based) as the intended application.
  • Orthogonal Validation: Where possible, the PBMC dataset should be accompanied by an orthogonal ground truth for CNVs, such as (sc)WGS or WES data from the same sample, to confirm the absence of CNVs [3].

• Tool Execution and Configuration:

  • Reference Cells: All scRNA-seq CNV methods require a set of known diploid cells for normalization. For a pure diploid sample, a subset of the cells can be designated as the reference. It is critical to use the same set of healthy reference cells across all tools to ensure a reproducible and fair comparison [3].
  • Parameters: Tools should be run with their default parameters as recommended in their official tutorials or documentation to simulate a standard user scenario [3].

• Generation of Pseudobulk Profiles:

  • Since the ground truth is often a bulk measurement (e.g., WGS), the per-cell CNV predictions from each tool are combined into an average CNV profile (pseudobulk) before comparison. This step facilitates a direct and consistent evaluation against the known genomic state [3].

• Performance Metrics and Analysis:

  • Mean Square Error Deviation: For diploid samples, the mean square error (MSE) deviation from a diploid reference genome is calculated. A lower MSE indicates that the tool's inferred CNV profile is closer to the expected diploid state, reflecting a lower false positive rate [3].
  • Cell Classification Accuracy: For tools that automatically classify cells as "normal" or "tumor," the accuracy of this classification on the diploid sample is assessed. A high rate of cells incorrectly classified as "tumor" indicates a high false positive rate [5].
  • Visual Inspection: The inferred CNV profiles should be visually inspected for large-scale, spurious gains or losses that would not be present in a true diploid sample.

Table 2: Key Reagents and Computational Resources for scCNV Benchmarking

Item Name	Function/Description	Example/Note
Diploid scRNA-seq Dataset	Essential positive control for measuring baseline false positive rates.	Peripheral Blood Mononuclear Cells (PBMCs) from a healthy donor [3].
Reference Genomes	Baseline for read alignment and ploidy status comparison.	GRCh37 (hg19) or GRCh38. Ensure consistency across all data and tools [49].
High-Performance Computing (HPC) Cluster	Executes computationally intensive CNV calling algorithms.	Required for tools with high runtime and memory demands (e.g., Numbat, CaSpER) [3].
Benchmarking Pipeline	Standardized, reproducible framework for executing and comparing multiple tools.	Snakemake pipeline from "benchmarkscrnaseqcnv_callers" GitHub repository [3].
Orthogonal Validation Data	Gold-standard data to define ground truth CNV status.	(sc)WGS, WES, or array-CGH data from the same sample [3] [50].
Cell Type Annotations	Curated labels to define diploid reference cells for tool normalization.	Manual annotation based on clustering and marker genes, or published annotations [3].

Accurately assessing the false positive rate of single-cell CNV callers on diploid samples is a non-negotiable step in rigorous algorithm benchmarking. Current evidence indicates that no single tool is universally superior, and performance is highly context-dependent. Methods incorporating allelic frequency information (e.g., CaSpER, Numbat) show promise for more robust performance, while popular expression-only tools (e.g., SCEVAN, CopyKAT) tend to over-call CNVs in complex biological samples. Researchers are advised to use the provided experimental framework to validate their chosen tool's performance on diploid controls specific to their study system, prioritize the selection of an optimal reference dataset, and maintain a critical perspective on automated cell type predictions based solely on inferred CNVs.

Within the broader objective of benchmarking single-cell copy number variation (CNV) detection algorithms, clinical validation stands as the ultimate test for any computational method. As single-cell RNA sequencing (scRNA-seq) is increasingly used to study tumor heterogeneity, accurately inferring CNVs from this transcriptomic data becomes critical for clinical and drug development applications [2]. This guide provides an objective comparison of the performance of leading scCNV inference methods, validated against orthogonal single-cell Whole Exome Sequencing (scWES) and bulk Whole Genome Sequencing (WGS) data from a clinical small cell lung cancer (SCLC) sample [2]. We summarize quantitative performance data and detail the experimental protocols to offer researchers a clear framework for method selection and validation.

The following tables summarize the performance of various scCNV detection methods based on a comprehensive clinical validation study [2]. The data is derived from the analysis of a clinical SCLC dataset, with results benchmarked against orthogonal scWES and bulk WGS.

Table 1: Overall Performance in Clinical Dataset (SCLC) Validation

Method	CNV Inference Sensitivity	CNV Inference Specificity	Subpopulation Identification Accuracy	Key Strengths
CaSpER	High	High	Moderate	Balanced sensitivity and specificity; integrates expression and allele frequency [2].
CopyKAT	High	High	High	Excellent for subclone identification; robust statistical model [2] [4].
inferCNV	Moderate	Moderate	High	Superior for identifying tumor subpopulations [2] [4].
sciCNV	Moderate	Moderate	Moderate	Performs well in subclone identification on single-platform data [2].
HoneyBADGER	Lower	Lower	Lower	Allele-based version shows resilience to batch effects [2] [4].

Table 2: Impact of Technical Factors on Performance

Factor	Impact on CNV Calling	Best-Performing Method(s)
Sequencing Depth	Sensitivity and specificity vary significantly with sequencing depth [2].	CopyKAT, CaSpER [2]
Reference Data Selection	Performance is highly dependent on the selection of appropriate reference diploid cells [2] [3].	All methods are affected
Batch Effects (Multi-platform data)	Significantly impairs subclone identification in most methods [2].	HoneyBADGER (Allele-based) with batch correction [4]
Detection of Rare Populations	Varies by method; requires sufficient sequencing coverage [4].	inferCNV (with enough cells sequenced) [4]

Detailed Experimental Protocols for Clinical Validation

The core clinical dataset used for this validation was generated from a 59-year-old male patient with stage 4 small cell lung cancer [2]. The study design received IRB approval from the Hangzhou Cancer Hospital.

Sample Collection and Processing

• Tissue Acquisition: Multiple cancer tissue samples (0.2×1.5 cm each) were obtained via CT-guided lung biopsy during initial diagnosis and again upon relapse post-chemotherapy. Adjacent normal tissues were also collected and pathologically confirmed as peritumoral normal [2].
• Single-Cell Isolation: Fresh tissues were digested with collagenase IV at 37°C for 30 minutes. The resulting cell suspension was washed and diluted to a concentration of 1000 cells/ml. Single cells were isolated using a micromanipulation system [2].
• Blood Sample Collection: A 10 ml blood sample was drawn from the patient to serve as a germline control [2].

Orthogonal Genomic Profiling

The validation of scRNA-seq-based CNV calls relied on two orthogonal genomic techniques performed on the same patient samples.

• Bulk Cell Whole Genome Sequencing (Bulk WGS):

  • DNA Extraction: Genomic DNA (gDNA) was extracted from fresh-frozen tissues using the Qiagen QIAamp DNA mini kit.
  • Library Preparation & Sequencing: WGS libraries were constructed with the Illumina Tru-seq Nano DNA HT Sample Prep kit using 0.5 µg of gDNA. Sequencing was performed on an Illumina X10 platform to a coverage of 30x with 150 bp paired-end reads [2].

• Single-Cell Whole Exome Sequencing (scWES):

  • DNA Source: Genomic DNA was obtained from 69 primary tumor cells and 76 relapse tumor cells.
  • Sequencing: The exomes of these individual cells were sequenced to generate a ground truth for genetic alterations at single-cell resolution [2].

scRNA-Seq Data Generation for CNV Inference

• cDNA Amplification: cDNA was amplified from 92 primary tumor cells and 39 relapse tumor cells using the SMART-seq2 kit [2].
• Library Preparation & Sequencing: scRNA-seq libraries were constructed with 1 ng of amplified cDNA using the Nextera XT DNA Library Preparation Kit (Illumina). Libraries were sequenced to a depth of 20 million reads per cell on an Illumina X10 with 150 bp paired-end reads [2].

Experimental Workflow Diagram

The following diagram illustrates the integrated workflow for clinical dataset generation and multi-modal validation of CNV calls.

The Scientist's Toolkit: Research Reagent Solutions

The following table details key reagents and kits used in the featured clinical validation study [2], which are essential for replicating this benchmarking workflow.

Table 3: Essential Research Reagents and Kits

Item	Specific Product/Kit	Function in the Experimental Protocol
DNA Extraction Kit	Qiagen QIAamp DNA mini kit	Extraction of high-quality genomic DNA from fresh-frozen tissues for bulk WGS [2].
Bulk WGS Library Prep	Illumina Tru-seq Nano DNA HT Sample Prep kit	Preparation of PCR-free sequencing libraries from 0.5 µg of gDNA for whole-genome sequencing [2].
Single-Cell cDNA Kit	SMART-seq2 Kit	Amplification of cDNA from single cells for full-length transcriptome analysis prior to scRNA-seq library prep [2].
scRNA-seq Library Prep	Nextera XT DNA Library Preparation Kit (Illumina)	Construction of sequencing libraries from 1 ng of amplified single-cell cDNA [2].
Enzymatic Digestion	Collagenase IV (Sigma Aldrich)	Digestion of solid tumor tissue to dissociate it into a single-cell suspension for downstream analysis [2].

This comparison guide demonstrates that validation against orthogonal scWES and bulk WGS is critical for assessing the real-world performance of scCNV callers. The clinical SCLC dataset revealed that CopyKAT and CaSpER provide the most reliable CNV inference, whereas inferCNV and CopyKAT excel at identifying tumor subpopulations [2] [4]. For researchers and drug development professionals, the choice of algorithm must be guided by the specific biological question—whether the priority is overall genetic landscape fidelity or subclonal architecture resolution. Furthermore, performance is heavily influenced by experimental design, including sequencing depth, reference selection, and the mitigation of batch effects.

The detection of copy number variations (CNVs) from single-cell RNA sequencing (scRNA-seq) data has become an indispensable tool for exploring tumor heterogeneity, tracing cancer evolution, and understanding the functional impact of somatic alterations in complex tissues. The fundamental premise underlying these computational methods is that genes within amplified genomic regions demonstrate elevated expression levels, whereas genes in deleted regions show reduced expression compared to diploid regions. However, inferring CNVs from transcriptomic data is inherently challenging due to the complex regulatory mechanisms governing gene expression. This has spurred the development of numerous computational tools, each employing distinct normalization strategies and inference algorithms to enhance accuracy and reliability. Given the expanding application of these methods in both basic research and clinical contexts, a critical need exists for a systematic, evidence-based comparison of their performance under varied experimental conditions. This guide synthesizes findings from recent, comprehensive benchmarking studies to equip researchers with the knowledge needed to select the optimal CNV detection tool for their specific biological questions and technical settings.

Performance Comparison of scCNV Detection Tools

Recent independent benchmarking efforts have evaluated popular scCNV callers on multiple datasets with established ground truths, often derived from (sc)WGS or WES data. The table below summarizes the core characteristics and overall performance profile of the most widely used tools.

Table 1: Overview and Performance Profile of scCNV Detection Tools

Tool	Core Methodology	Data Input	Output Resolution	Key Strength	Key Weakness
InferCNV [3]	Hidden Markov Model (HMM)	Gene Expression	Per gene or segment; Groups cells into subclones	Excels in subclone identification, especially on a single platform [4].	Performance is highly dependent on the choice of reference dataset [3].
CopyKAT [3]	Segmentation	Gene Expression	Per gene or segment	Top performer for overall CNV inference and subclone identification; good sensitivity [12] [4].	Struggles with small dataset sizes [3].
CaSpER [3]	Hidden Markov Model (HMM)	Gene Expression + Allelic Information	Per cell	Robust performance for large, droplet-based datasets; balanced CNV inference [3] [4].	Requires higher computational runtime [3].
Numbat [3]	Hidden Markov Model (HMM)	Gene Expression + Allelic Information	Per segment; Groups cells into subclones	Robust for large, droplet-based datasets by leveraging allelic information [3].	Requires higher computational runtime [3].
SCEVAN [3]	Segmentation	Gene Expression	Per segment; Groups cells into subclones	Effective segmentation-based approach [3].	Performance can be dataset-specific [3].
HoneyBADGER [12]	Bayesian Model	Gene Expression (Allelic version available)	N/A	Allelic version is more resilient to batch effects [4].	Lower sensitivity in detecting rare tumor populations [4].
sciCNV [12]	N/A	Gene Expression	N/A	Good performance in subclone identification from single-platform data [4].	Falls short in detecting rare tumor populations [4].

Quantitative benchmarking reveals how these tools balance sensitivity and specificity. The following table compiles key performance metrics from head-to-head evaluations.

Table 2: Quantitative Performance Metrics Across Benchmarking Studies

Tool	CNV Inference Performance	Subclone Identification	Sensitivity to Factors
InferCNV	Variable, depends on reference and dataset [3].	Excellent, particularly for data from a single platform [12] [4].	Highly sensitive to reference choice and batch effects across platforms [3] [12].
CopyKAT	Consistently high, one of the top overall performers [12] [4].	Excellent [12].	Performance drops with small dataset sizes [3].
CaSpER	Consistently high and robust, especially in droplet-based data [3] [4].	Good [3].	More robust to noise in large datasets; requires higher runtime [3].
Numbat	Robust in large, droplet-based datasets [3].	Good (infers subclones) [3].	Requires higher runtime [3].
SCEVAN	Good, but dataset-specific [3].	Good [3].	Performance varies across datasets [3].
HoneyBADGER	Lower sensitivity compared to top performers [4].	N/A	Allelic information can provide resilience to batch effects [4].
sciCNV	Good for subclones, but poor for rare populations [4].	Excellent on single-platform data [4].	Poor sensitivity for rare tumor populations [4].

Experimental Protocols for Benchmarking

To ensure fair and reproducible comparisons, benchmarking studies follow rigorous experimental protocols. The workflow below outlines the standard procedure for evaluating scCNV tools.

Dataset Curation and Ground Truth Establishment

Benchmarking relies on diverse scRNA-seq datasets from various platforms (e.g., 10x Genomics, SMART-seq2) and sample types, including cancer cell lines, primary tumors, and diploid control cells [3] [12]. A critical component is the establishment of a ground truth for CNVs. This is typically obtained from orthogonal genomic measurements performed on the same sample or cell line, such as:

• Single-cell or Bulk Whole-Genome Sequencing ((sc)WGS): Considered the gold standard for per-cell CNV profiling [3].
• Whole Exome Sequencing (WES): Provides a validated set of CNVs for comparison [3].
• Chromosomal Microarray Analysis (CMA) or Karyotyping: Used in clinical validations, especially for methods like msCNVS [7].

Tool Execution and Profile Generation

Each CNV detection tool is run with its recommended parameters and reference genome. A key challenge is that the ground truth is not always measured in the same cells as the scRNA-seq data. To enable a direct comparison, the per-cell CNV predictions from scRNA-seq tools are often combined into a pseudobulk profile—an average CNV profile across all cells [3]. For datasets where scRNA-seq and scWGS are measured in the same individual cells, a direct cell-by-cell comparison is possible and provides the highest resolution validation [3].

Performance Metrics and Statistical Evaluation

The accuracy of tool predictions is assessed using multiple threshold-independent and threshold-dependent metrics against the ground truth [3]:

• Threshold-independent Metrics:
  • Correlation: Measures the linear relationship between inferred and true CNV signals.
  • Area Under the Curve (AUC): Evaluates the ability to distinguish gains from diploid regions and losses from diploid regions. Partial AUC is often used to focus on biologically meaningful threshold ranges [3].
• Threshold-dependent Metrics:
  • Sensitivity & Specificity: Calculated for gain and loss events after determining optimal thresholds via an F1 score [3].
  • F1 Score: The harmonic mean of precision and recall, used to find the optimal balance for calling gains and losses [3].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table details key resources and computational tools referenced in the featured benchmarking experiments.

Table 3: Key Research Reagents and Computational Tools for scCNV Analysis

Item Name	Function / Application	Relevant Context
Reference Genomes (GRCh38/hg38)	Standard genome for read alignment and CNV calling.	Used as the baseline for mapping and variant identification in modern studies [38].
Diploid Reference Cells	A set of normal cells used for expression normalization by all scCNV tools.	Critical for accurate CNV calling; can be from the same sample (e.g., healthy cells) or an external matched dataset [3].
Benchmarking Pipeline (Snakemake)	A reproducible workflow for running and comparing multiple CNV callers.	The benchmarking pipeline from [1] is publicly available to test new datasets and compare methods [3].
Ginkgo	An open-source web platform for interactive analysis of single-cell CNVs from sequencing data.	Provides an automated pipeline from read binning to visualization and phylogenetic tree construction [51].
Genome in a Bottle (GIAB) Consortium Data	A high-confidence set of variant calls, including deletions, for the NA12878 genome.	Serves as a gold-standard benchmark for validating CNV calls from WGS data [52] [39].

Scenario-Based Recommendations and Decision Framework

Selecting the optimal tool requires matching the tool's strengths to the specific research scenario. The following diagram maps common research objectives to the recommended tools.

Based on the aggregated benchmarking results, the following scenario-based recommendations are proposed:

• For Large-Scale Droplet-Based scRNA-seq Datasets: Tools that incorporate allelic information, such as CaSpER and Numbat, demonstrate more robust performance for large, droplet-based datasets. However, they require a higher computational runtime. CopyKAT is also a strong performer in this context [3].
• For Precise Identification of Tumor Subclones: When the research goal is to resolve the clonal architecture of a tumor, InferCNV and CopyKAT have been shown to outperform other methods [12] [4]. This is particularly true for data generated on a single platform.
• For Datasets with Inherent Batch Effects or Multiple Platforms: Batch effects can significantly compromise the performance of most methods. If integrating datasets from different platforms, applying batch correction tools (e.g., ComBat) is essential. In such scenarios, the allelic version of HoneyBADGER has also shown greater resilience to technical distortions [4].
• For Studies with Limited Cell Numbers or Low Sequencing Depth: The performance of several tools, including CopyKAT, can degrade with small dataset sizes [3]. It is crucial to consult the original publications of each tool for guidance on minimum cell numbers and sequencing depth requirements. For WGS-based CNV detection from low-input samples, specialized wet-lab protocols like msCNVS have been developed to improve accuracy [7].
• For General-Purpose Use and a Balance of Performance: When a single, versatile tool is desired, CopyKAT and CaSpER have consistently ranked as top all-around performers in independent benchmarks, offering a good balance of sensitivity, specificity, and robustness [12] [4].

The rigorous benchmarking of scCNV detection tools reveals a nuanced landscape where no single method is universally superior. The performance of each tool is contingent upon a interplay of factors, including dataset size, sequencing technology, the choice of reference cells, and the specific biological question. Currently, CopyKAT and CaSpER emerge as leading choices for robust overall CNV inference, while InferCNV remains a powerful option for dissecting subclonal populations, provided that technical confounders like batch effects are carefully managed. As the field progresses, the development of more automated and resilient algorithms, along with continued independent benchmarking, will be paramount for unlocking the full potential of scRNA-seq data to illuminate the genetic underpinnings of cancer and other complex diseases. Researchers are encouraged to use publicly available benchmarking pipelines to validate their tool of choice on their own specific data types whenever possible.

Conclusion

Recent benchmarking studies provide crucial guidance for navigating the complex landscape of single-cell CNV detection algorithms. The evidence clearly demonstrates that no single tool excels in all scenarios; rather, method selection must be tailored to specific research objectives. CaSpER and CopyKAT generally provide the most balanced CNV inference, while inferCNV and CopyKAT show superior performance in identifying tumor subpopulations. Critical factors influencing success include appropriate reference selection, sequencing depth, platform compatibility, and awareness of batch effects. Future directions should focus on developing more robust algorithms that better integrate allelic information, improve resistance to technical artifacts, and enhance detection of rare subclones. As single-cell technologies transition toward clinical applications, establishing standardized benchmarking pipelines and validation frameworks will be essential for advancing precision oncology and unlocking the full potential of CNV analysis in understanding tumor evolution and therapeutic resistance.

References

• 1. Tool Comparison for Detecting Tumour Cells in Endometrial Cancer... [https://pubmed.ncbi.nlm.nih.gov/41160707/]
• 2. A benchmarking study of copy number variation inference ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12204187/]
• 3. Benchmarking scRNA-seq copy number variation callers [https://www.nature.com/articles/s41467-025-62359-9]
• 4. Which tool tells the truth? A head-to-head benchmarking of ... [https://www.rna-seqblog.com/which-tool-tells-the-truth-a-head-to-head-benchmarking-of-scrna-seq-cnv-methods/]
• 5. Tool Comparison for Detecting Tumour Cells in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12571190/]
• 6. single cell CNV profiling method using dynamic programming [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-021-07941-3]
• 7. A medium throughput approach for single cell copy number ... [https://www.sciencedirect.com/science/article/pii/S2090123225008811]
• 8. Sensitivity to copy number variation analysis in single cell ... [https://www.sciencedirect.com/science/article/abs/pii/S0378111921005904]
• 9. Comparative study of tools for copy number variation ... [https://www.nature.com/articles/s41598-025-06527-3]
• 10. A head-to-head benchmarking of scRNA-seq CNV methods [https://www.cbs42.com/business/press-releases/ein-presswire/832141205/a-head-to-head-benchmarking-of-scrna-seq-cnv-methods]
• 11. Benchmarking of germline copy number variant callers ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12005901/]
• 12. A benchmarking study of copy number variation inference ... [https://pubmed.ncbi.nlm.nih.gov/40584741/]
• 13. A head-to-head benchmarking of scRNA-seq CNV methods [https://www.einpresswire.com/article/832141205/a-head-to-head-benchmarking-of-scrna-seq-cnv-methods]
• 14. Normalizing single-cell RNA sequencing data - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5549838/]
• 15. Data normalization for addressing the challenges in the ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-024-10364-5]
• 16. SCOPE: A Normalization and Copy-Number Estimation ... [https://www.sciencedirect.com/science/article/pii/S2405471220301113]
• 17. Benchmarking copy number aberrations inference tools ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11879432/]
• 18. Benchmarking scRNA-seq copy number variant callers [https://github.com/colomemaria/benchmark_scrnaseq_cnv_callers]
• 19. A Comparison of Tools That Identify Tumor Cells by ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11351975/]
• 20. CopyMix: Mixture model based single-cell clustering and ... [https://www.sciencedirect.com/science/article/pii/S1476927124002457]
• 21. single cell CNV profiling method using dynamic programming [https://pmc.ncbi.nlm.nih.gov/articles/PMC8596905/]
• 22. Methods for copy number aberration detection from single-cell ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-020-02119-8]
• 23. Concepts and new developments in droplet-based single-cell ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11568944/]
• 24. Hi-C for Fast and Scalable Profiling of Chromatin Architecture... Droplet [https://pmc.ncbi.nlm.nih.gov/articles/PMC11071305/]
• 25. Droplet Hi-C enables scalable, single-cell profiling ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12520981/]
• 26. Batch Effect Correction [https://www.10xgenomics.com/analysis-guides/introduction-batch-effect-correction]
• 27. Deep learning applications in single-cell genomics and ... [https://www.sciencedirect.com/science/article/pii/S0753332223008685]
• 28. A multi-center cross-platform single-cell RNA sequencing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7854649/]
• 29. Batch correction methods used in single-cell RNA sequencing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12315870/]
• 30. A benchmark of batch-effect correction methods for single-cell ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-019-1850-9]
• 31. Reference-informed evaluation of batch correction for ... [https://www.nature.com/articles/s42003-025-07947-7]
• 32. Benchmarking scRNA-seq copy number variation callers - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC12491403/]
• 33. High-resolution detection of copy number alterations in ... [https://www.nature.com/articles/s41467-025-60446-5]
• 34. Shallow Sequencing (Low-Pass) CNV Analysis: A Brief ... [https://bionano.com/blogs/shallow-sequencing-low-pass-cnv-analysis-guide/]
• 35. Benchmarking strategies for CNV calling from whole ... [https://www.sciencedirect.com/science/article/pii/S2001037025000674]
• 36. Validation of a comprehensive long-read sequencing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12082127/]
• 37. Tumour purity as an underlying key factor in tumour mutation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10154864/]
• 38. Comparative study of tools for copy number variation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12218993/]
• 39. Benchmarking copy number variation detection with low- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12481693/]
• 40. Accurate Inference of Tumor Purity and Absolute Copy ... [https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2020.00458/full]
• 41. Numbat-multiome: inferring copy number variations by ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12531993/]
• 42. RCANE: a deep learning algorithm for whole-genome pan- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12460768/]
• 43. CNRein: an evolution-aware deep reinforcement learning ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03553-2]
• 44. High-resolution detection of copy number alterations in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12214996/]
• 45. Comprehensive performance comparison of high-resolution ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-017-3658-x]
• 46. Detection of germline CNVs from gene panel data [https://pmc.ncbi.nlm.nih.gov/articles/PMC11637760/]
• 47. GitHub - mhazarabad/scDNA-Analysis: Single - Cell ... CNV Benchmark [https://github.com/mhazarabad/scDNA-Analysis]
• 48. Robust analysis of allele-specific copy number alterations ... [https://www.nature.com/articles/s41467-024-51026-0]
• 49. Benchmarking germline CNV calling tools from exome ... [https://www.nature.com/articles/s41598-021-93878-2]
• 50. Accurate and reliable high-throughput detection of copy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC1665640/]
• 51. Interactive analysis and assessment of single-cell copy- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4775251/]
• 52. Evaluation of the performance of copy number variant ... [https://www.sciencedirect.com/science/article/pii/S1532046419300929]