CoChat

Comprehensive assessment of alternative splicing analysis methods for single-cell RNA-seq

Open access

The DOI matches its Crossref record; 69 of the 69 references it cites were found in a registry or an index; none of the 20 sources its claims cite is checked for support yet.

Abstract

Alternative splicing expands transcriptomic and proteomic diversity and contributes to cell identity and function. Advances in single-cell RNA sequencing enable transcriptome profiling at cellular resolution, yet their performance across biologically relevant tasks has not been systematically evaluated. Here, we benchmark six representative alternative splicing analysis methods across multiple full-length, short-read scRNA-seq datasets, focusing on biologically relevant tasks including cell clustering and differential splicing detection. Among these methods, scQuint achieves robust performance across datasets and tasks. Application of scQuint to stem cell differentiation and tumor-associated immune cell datasets identifies cell type-specific splicing programs with biological significance. These results provide guidance for method selection and highlight the value of single-cell splicing analysis for uncovering regulatory mechanisms beyond gene expression.

Overview

CoChat Summary

TL;DR

This paper benchmarks six alternative splicing analysis methods for single-cell RNA-seq across multiple Smart-seq datasets and biologically relevant tasks. scQuint consistently performed best for splicing-based cell clustering and differential splicing detection, and its application revealed cell-type-specific splicing programs in stem cell differentiation and tumor-associated immune cells.

Key findings

  • Across five clustering datasets totaling 7,167 cells and 46 cell types, scQuint and SCASL had the highest average accuracy (0.89/0.88), ARI (0.82/0.81), NMI (0.85/0.83), and silhouette score (0.55/0.51); their average BCI scores were 0.78 and 0.76, versus MARVEL 0.67, BRIE2 0.53, Psix 0.70, and SpliZ 0.41.
  • For differential alternative splicing event (DASE) detection, scQuint achieved significantly higher AUCCs at both event and gene levels than other methods; it recovered all nine experimentally validated events within its top 112 ranked DASEs, while SCASL did so within its top 291, and BRIE2, SpliZ, and MARVEL recovered only 5, 7, and 7 within their top 500.
  • scQuint and Psix processed 3,000 cells within minutes, whereas SCASL required over 3.8 h, BRIE2 took 53 h, and SpliZ required approximately 16 days; peak memory at 3,000 cells was around 9 GB for BRIE2, scQuint, and Psix, about 30 GB for MARVEL, 124 GB for SpliZ, and 169 GB for SCASL.
  • Applying scQuint to hepatocyte differentiation and leukocyte datasets identified cell-type-specific splicing programs, including ESRP1 alternative intron PSI changing from ∼40% before differentiation to ∼70% at the end, ITGB1BP1 introns 5–6 mean PSI decreasing from ∼40% in iPSCs to 0, and PTPRC intron 3–5 PSI increasing from naive to activated T cells.

Methods

  • The authors benchmarked six methods—BRIE2, scQuint, Psix, SpliZ, MARVEL, and SCASL—on public full-length Smart-seq2 scRNA-seq datasets; five datasets were used for clustering evaluation and five datasets from three studies, all with matched bulk RNA-seq, were used for DASE evaluation.
  • Clustering performance was assessed with accuracy, ARI, NMI, silhouette score, and a composite BCI score; DASE performance was assessed by AUCC concordance with matched bulk RNA-seq, recovery of nine experimentally validated splicing events, and GO enrichment, with parameter sweeps over variable events, PCA components, neighbors, and clustering resolution.
  • Raw reads were adapter-trimmed with Trimmomatic and aligned with STAR or STARsolo; bulk AS analysis used rMATS-turbo and LeafCutter, and scalability was measured on a specified CPU cluster.

Limitations

  • The paper states that benchmarking focuses on full-length scRNA-seq datasets generated with Smart-seq-based protocols, so performance on droplet-based platforms such as 10× Genomics is not evaluated.
  • The paper states that clustering was assessed with the BCI score, a concordance-based metric emphasizing agreement with expression-derived cell-type annotations, which may underestimate cases where AS captures biologically meaningful populations not well represented by those annotations; scQuint also cannot distinguish splicing event types such as SE, A3SS, and alternative 5′ splice sites, and groups events sharing the same 3′ end, sometimes requiring Sashimi plots for detailed interpretation.
  • The paper notes that the Tasic et al. dataset used a randomly selected subset of cell types, which may affect separability, and that sex information was not consistently available across datasets, so sex-specific effects were not evaluated; the text also makes evident that the benchmark emphasizes real-data biological utility rather than controlled simulation settings.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.isci.2026.116090

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Graphical abstract
Graphical abstract CC BY · PubMed Central
Figure 1 Overview of the benchmarking framework and datasets (A) Schematic of the benchmarking workflow for six computational methods designed to analyze alternative splicing in single-cell transcriptomes. Methods were evaluated based on their performance in cell clustering and in detecting differentially regulated alternative splicing events. (B) A brief overview of the methods included for benchmarking. (C) Overview of the eight single-cell RNA-seq datasets used for evaluation.
Figure 1 Overview of the benchmarking framework and datasets (A) Schematic of the benchmarking workflow for six computational methods designed to analyze alternative splicing in single-cell transcriptomes. Methods were evaluated based on their performance in cell clustering and in detecting differentially regulated alternative splicing events. (B) A brief overview of the methods included for benchmarking. (C) Overview of the eight single-cell RNA-seq datasets used for evaluation.CC BY · PubMed Central
Figure 2 Performance of splicing-based clustering across AS methods (A) Mean accuracy and ARI scores for six methods on five scRNA-seq datasets. Dashed lines mark the median performance across methods. Values are mean ±0.5 × SD; error bars show SD across datasets. (B) Same as A, but evaluated using average NMI and silhouette scores. (C) Aggregate performance quantified by BCI scores (mean ±95% CI, N = 5). Each dot represents one dataset. (D) UMAP embeddings generated from AS profiles of the Camp et al., 2017 dataset illustrate the discriminative ability of each method. Clustering based on gene expression is shown as a reference. The score displayed next to each method indicates its corresponding BCI score. Cell clusters were labeled and colored by cell type annotations: iPSCs (induced pluripotent stem cells), DE (definitive endoderm), HE (hepatic endoderm), IH (immature hepatoblast-like cells), MH (mature hepatocyte-like cells), EC (endothelial cells), MC (mesenchymal cells), LD (3D liver bud).
Figure 2 Performance of splicing-based clustering across AS methods (A) Mean accuracy and ARI scores for six methods on five scRNA-seq datasets. Dashed lines mark the median performance across methods. Values are mean ±0.5 × SD; error bars show SD across datasets. (B) Same as A, but evaluated using average NMI and silhouette scores. (C) Aggregate performance quantified by BCI scores (mean ±95% CI, N = 5). Each dot represents one dataset. (D) UMAP embeddings generated from AS profiles of the Camp et al., 2017 dataset illustrate the discriminative ability of each method. Clustering based on gene expression is shown as a reference. The score displayed next to each method indicates its corresponding BCI score. Cell clusters were labeled and colored by cell type annotations: iPSCs (induced pluripotent stem cells), DE (definitive endoderm), HE (hepatic endoderm), IH (immature hepatoblast-like cells), MH (mature hepatocyte-like cells), EC (endothelial cells), MC (mesenchymal cells), LD (3D liver bud).CC BY · PubMed Central
Figure 3 Performance of DASE detection (A) Bar plots showing the number of significant DASEs identified by each method, with adjusted p value <0.05. (B) Upset plot showing overlaps among the top 200 ranked splicing events for each method (Song et al., 2017 dataset). Ranking was based on statistical significance without applying adjusted p value thresholds. (C and D) AUCC for the top 200 ranked splicing events per method, evaluated at the event level (C) and gene level (D) across matched bulk and single-cell datasets. Boxplots show median (center line), interquartile range (box), and 1.5× IQR (whiskers). Points represent individual datasets. (E) Cumulative recall curves showing the recovery of nine experimentally validated differential splicing events by each method. The top 500 differential splicing events per method were ranked by statistical significance (x axis: rank 1–500; y axis: cumulative true positives). Methods that recover all validated events with fewer top-ranked predictions demonstrate stronger early recall. Validated events were sourced from Song et al., 2017. (F) GO enrichment analysis based on the top 200 ranked splicing events per method. Each bubble represents a GO term enrichment result, with the size indicating enrichment significance (−log10p value) and color intensity representing the gene ratio. Analysis was based on the 72-h serum-induced myoblast differentiation from the Trapnell et al., 2014 dataset. (G–I) Sashimi plots showing example alternative splicing events for DCTN1 (G), DGUOK (H), and NOP56 (I) from Song et al. 2017. For each gene, splicing patterns are visualized in MNs and iPSCs conditions. Bulk RNA-seq data (top) and single-cell RNA-seq data (bottom) are shown for comparison. Sashimi plots were generated directly from BAM files to visualize splicing patterns, and the displayed events were detected by both rMATS and LeafCutter.
Figure 3 Performance of DASE detection (A) Bar plots showing the number of significant DASEs identified by each method, with adjusted p value <0.05. (B) Upset plot showing overlaps among the top 200 ranked splicing events for each method (Song et al., 2017 dataset). Ranking was based on statistical significance without applying adjusted p value thresholds. (C and D) AUCC for the top 200 ranked splicing events per method, evaluated at the event level (C) and gene level (D) across matched bulk and single-cell datasets. Boxplots show median (center line), interquartile range (box), and 1.5× IQR (whiskers). Points represent individual datasets. (E) Cumulative recall curves showing the recovery of nine experimentally validated differential splicing events by each method. The top 500 differential splicing events per method were ranked by statistical significance (x axis: rank 1–500; y axis: cumulative true positives). Methods that recover all validated events with fewer top-ranked predictions demonstrate stronger early recall. Validated events were sourced from Song et al., 2017. (F) GO enrichment analysis based on the top 200 ranked splicing events per method. Each bubble represents a GO term enrichment result, with the size indicating enrichment significance (−log10p value) and color intensity representing the gene ratio. Analysis was based on the 72-h serum-induced myoblast differentiation from the Trapnell et al., 2014 dataset. (G–I) Sashimi plots showing example alternative splicing events for DCTN1 (G), DGUOK (H), and NOP56 (I) from Song et al. 2017. For each gene, splicing patterns are visualized in MNs and iPSCs conditions. Bulk RNA-seq data (top) and single-cell RNA-seq data (bottom) are shown for comparison. Sashimi plots were generated directly from BAM files to visualize splicing patterns, and the displayed events were detected by both rMATS and LeafCutter.CC BY · PubMed Central
Figure 4 Single-cell resolution analysis of DASEs reveals cell-type-specific splicing programs with functional relevance (A) Heatmap of the top 100 cell-type-specific DASEs (adjusted p < 0.05, Δintron PSI ≥0.15) from the Camp et al., 2017 dataset, with enriched GO terms displayed on the right. (B) PSI values of selected DASEs projected onto a force-directed embedding of the iPSC differentiation trajectory (Camp et al., 2017). Left: cell states. Top-right and bottom-right: PSI values and corresponding gene expression levels. (C) Heatmap showing the top 100 cell-type-specific DASEs (adjusted p < 0.05, Δintron PSI ≥0.15), with enriched GO terms shown on the right. Cell types with fewer than 100 cells were excluded. Analysis is based on the Zhang et al. 2019 dataset. (D) PSI values of selected DASEs projected onto a UMAP embedding of cell clusters (Zhang et al., 2019). Left: cell identity. Top-right and bottom-right: PSI values and corresponding gene expression levels. (E) Force-directed layout of CD4+ and CD8+ T cell subtypes from the Zhang et al., 2019 dataset, including all donors. (F) PSI values of the PTPRC intron 3–5 across CD4+ and CD8+ T cell subtypes in the Zhang et al., 2019 dataset, including all donors. (G) Mean PSI values of PTPRC intron 3–5 across T cell subtypes, ordered from naive to activated states.
Figure 4 Single-cell resolution analysis of DASEs reveals cell-type-specific splicing programs with functional relevance (A) Heatmap of the top 100 cell-type-specific DASEs (adjusted p < 0.05, Δintron PSI ≥0.15) from the Camp et al., 2017 dataset, with enriched GO terms displayed on the right. (B) PSI values of selected DASEs projected onto a force-directed embedding of the iPSC differentiation trajectory (Camp et al., 2017). Left: cell states. Top-right and bottom-right: PSI values and corresponding gene expression levels. (C) Heatmap showing the top 100 cell-type-specific DASEs (adjusted p < 0.05, Δintron PSI ≥0.15), with enriched GO terms shown on the right. Cell types with fewer than 100 cells were excluded. Analysis is based on the Zhang et al. 2019 dataset. (D) PSI values of selected DASEs projected onto a UMAP embedding of cell clusters (Zhang et al., 2019). Left: cell identity. Top-right and bottom-right: PSI values and corresponding gene expression levels. (E) Force-directed layout of CD4+ and CD8+ T cell subtypes from the Zhang et al., 2019 dataset, including all donors. (F) PSI values of the PTPRC intron 3–5 across CD4+ and CD8+ T cell subtypes in the Zhang et al., 2019 dataset, including all donors. (G) Mean PSI values of PTPRC intron 3–5 across T cell subtypes, ordered from naive to activated states.CC BY · PubMed Central
Figure 5 Computational resources consumed by the six methods (A and B) Computational time (A) and memory usage (B) of six methods for generating processed alternative splicing profiles across datasets with varying cell numbers.
Figure 5 Computational resources consumed by the six methods (A and B) Computational time (A) and memory usage (B) of six methods for generating processed alternative splicing profiles across datasets with varying cell numbers.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 1 Not valid · 14 Not checked · 109 Own results and interpretations

64 more claims

Create a free account to see all 124 claims and to verify the 20 references still open.

See all claims — free account
  • It may state two separate claims as one. 35
  • Probably not a substantial claim. 30

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 69 references

Cited by

See all 1

Metadata

Authors
Jiang, Q., Li, G., Xing, L., Zhang, D., & Sun, J.
Affiliations
  • Southeast University (CN)
  • Nantong University (CN)
  • Second Affiliated Hospital of Nantong University (CN)
Corresponding author
  • Jiang, Quanlong
  • Zhang, Dongmei
  • Sun, Junjie
Venue
iScience
Publisher
Elsevier BV
Published
2026-06
Volume
29
Issue
6
Pages
116090
ISSN
2589-0042 (ISSN-L)
OpenAlex
W7162082526
Cited by
1
References
69
Citations by year
2026: 1
FWCI
1.13
Type
journal article
Language
en
Open access
Licence
Locations
  • iScience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • Natural Science Foundation of Nantong Municipality
  • Science and Technology Project of Nantong City (JC2024024, JC2023114)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › RNA Research and Splicing
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
Keywords
alternative splicing, single-cell RNA sequencing, cell clustering, full-length transcriptome, stem cell differentiation, tumor-infiltrating immune cells

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-22, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-22 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Comprehensive assessment of alternative splicing analysis methods for single-cell RNA-seq. https://app.cochat.ai/papers/10.1016/j.isci.2026.116090
The paper
Jiang, Q., Li, G., Xing, L., Zhang, D., & Sun, J. (2026). Comprehensive assessment of alternative splicing analysis methods for single-cell RNA-seq. iScience. https://doi.org/10.1016/j.isci.2026.116090

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.