Summary

Single-cell alternative splicing analysis reveals RNA-processing differences that can be hidden when transcriptomic data are reduced to gene-level counts. A 2026 iScience benchmark compared six representative methods—BRIE2, scQuint, Psix, SpliZ, MARVEL, and SCASL—using full-length, short-read single-cell RNA-seq datasets. It evaluated splicing-based cell clustering, differential alternative splicing detection, and computational scalability. The benchmark provides useful guidance, but it does not establish one universal winner for every dataset. Method performance depends on sequencing protocol, splice-junction coverage, event definition, missingness, preprocessing, biological design, and computing environment. Experimental strategy matters just as much as software choice: 3'-biased droplet data, full-length short-read data, and long-read transcript data provide different levels of splicing and isoform information.

Research-use note: The analytical methods and sequencing strategies discussed below are intended for research use. Splicing associations require appropriate biological and experimental validation before translational interpretation.

Key Takeaways

• Alternative splicing adds a transcript-processing layer beyond gene-level expression.
• The 2026 benchmark evaluated six methods using full-length, short-read Smart-seq-type datasets.
• scQuint showed the strongest overall performance within that benchmark, not across every possible dataset.
• Annotation-free methods produced broader and denser splicing representations in the tested datasets, but broader detection does not automatically mean greater biological accuracy.
• Runtime and memory differed substantially among tools, but the reported numbers are benchmark-specific.
• 3'-biased droplet data can contain useful splice-junction information, but comprehensive isoform analysis is more difficult.
• Long-read sequencing can observe transcript structures more directly, while still requiring alignment, QC, isoform calling, filtering, and quantification.
• Biological replication and independent validation remain essential for mechanistic conclusions.

Figure 1. Single-cell alternative splicing analysis adds transcript-processing information beyond gene-level expression.

Why Analyze Alternative Splicing Beyond Gene Expression?

Gene-level expression asks how much RNA is produced from a gene. Alternative splicing asks a different question: Which transcript structures are being produced from that gene? A single gene can generate multiple transcript isoforms through exon skipping, alternative splice-site usage, intron retention, mutually exclusive exons, and other RNA-processing events. Two cells can therefore show similar total gene expression while using different transcript structures. That distinction matters because isoform usage can influence:

• protein-coding sequence;
• regulatory domains;
• untranslated regions;
• transcript stability;
• localization;
• nonsense-mediated decay;
• cell-state-specific RNA processing.

A 2026 review in Trends in Genetics describes isoform-resolved single-cell transcriptomics as a layer that extends beyond conventional gene-count analysis and highlights two major routes: computational reconstruction from full-length short reads and more direct transcript-structure observation with long reads. [2].

Splicing Can Add Cell-State Information

Alternative splicing should not be treated as “better than gene expression.” It is complementary. Gene expression may cleanly separate two cell populations while splicing adds little. In another system, the populations may have similar expression profiles but distinct splice-junction usage or transcript isoforms. The 2026 iScience benchmark found that splicing-derived features could capture biologically meaningful cell-type differences across several full-length single-cell datasets. The same study later used the strongest-performing workflow to examine cell-type-specific splicing programs during differentiation and in tumor-associated immune cells. The practical question is therefore: Does the biological hypothesis depend on RNA-processing differences that gene counts alone cannot resolve? If the answer is no, a dedicated splicing workflow may add unnecessary complexity.

What Did the 2026 Benchmark Actually Test?

The benchmark by Jiang and colleagues was designed to compare computational tools for single-cell alternative splicing analysis, not to compare every single-cell RNA-seq technology. That scope is essential for interpreting the rankings correctly.

Six Methods Were Evaluated

The study compared:

• BRIE2
• scQuint
• Psix
• SpliZ
• MARVEL
• SCASL

Five of the six supported differential alternative splicing analysis in the benchmark; Psix was evaluated for splicing-based cell representation and clustering but not for the same differential-event workflow. The study focused on three practical questions:

1. Can splicing profiles separate biological cell populations?
2. Can the methods recover differential alternative splicing patterns that agree with independent evidence?
3. How much computational time and memory do the methods require?

The Data Were Full-Length Short-Read Single-Cell RNA-Seq

This is one of the most important limitations of the study. The benchmark focused on full-length, plate-based, short-read single-cell RNA-seq datasets generated with Smart-seq-based protocols. It was not a benchmark of ordinary 3'-tag droplet data. For the clustering evaluation, the authors curated five datasets containing a total of 7,167 cells across 46 cell types. These values describe the benchmark design only; they are not required sample sizes for a splicing project. Researchers planning a new full-length experiment can review our Single-cell Transcriptome Smart-seq Services.

Benchmark Results Depend on the Evaluation Framework

Cell clustering was assessed using:

• accuracy;
• adjusted Rand index;
• normalized mutual information;
• silhouette score;
• a composite benchmark clustering index.

Differential splicing performance was evaluated using several evidence layers, including concordance with matched bulk RNA-seq, recovery of experimentally validated events, and functional enrichment. Computational scalability was assessed by subsampling increasing cell numbers and recording runtime and peak memory. This makes the benchmark useful, but still conditional on the authors' preprocessing, parameter search, software versions, event definitions, datasets, and hardware. [1].

Six Methods, Two Analytical Strategies

One useful way to understand the six tools is to divide them by how they define splicing events. The 2026 benchmark broadly grouped the methods into predefined-annotation and annotation-free strategies.

Figure 2. Six splicing methods use different event definitions and annotation strategies, so method choice should follow the data and research question.

Annotation-Dependent Does Not Mean “Inferior”

Annotation-dependent methods are useful when the study has a predefined biological question. For example:

• a known skipped exon;
• a curated set of disease-associated splice events;
• a defined gene panel;
• a targeted validation study;
• an analysis where event-level interpretability matters more than discovery breadth.

Reference-guided analysis can make downstream interpretation simpler because the events already map to defined transcript structures. Its main limitation is that the detectable event space depends on the annotation and the method's event definition.

Annotation-Free Does Not Mean “Automatically Better”

Annotation-free approaches reduce dependence on a predefined transcript catalog and can recover known and previously unannotated splice-site usage. That flexibility can be valuable when:

• novel junctions are expected;
• transcript annotation is incomplete;
• the biological system is poorly characterized;
• broad event discovery is the priority.

But broader discovery also creates additional requirements:

• stronger filtering;
• splice-junction QC;
• interpretation of novel events;
• harmonization across samples;
• independent validation.

A method detecting more events may simply be defining events more broadly.

Event Definitions Are Not Directly Comparable

A skipped exon, an intron cluster, a splice-site usage event, and a gene-level splicing score are not the same mathematical object. This matters when reading benchmarking plots. For example, one method may report 20,000 splice-site events while another reports 5,000 exon-skipping events. The larger number does not prove higher sensitivity because the denominators and event definitions differ. The most useful selection rule is: Choose the event representation that matches the biological question, rather than choosing a tool because it is labeled annotation-free or annotation-dependent.

Benchmark 1 — Can Splicing Profiles Separate Cell Types?

The first benchmark asked whether cells could be grouped using splicing-derived features rather than conventional gene-expression counts. Across five benchmark datasets, scQuint and SCASL showed the highest average performance. The reported mean benchmark clustering index values were:

• scQuint: 0.78
• SCASL: 0.76
• Psix: 0.70
• MARVEL: 0.67
• BRIE2: 0.53
• SpliZ: 0.41

These values should be interpreted exactly as they were generated: They are benchmark-specific results from the tested full-length short-read datasets and parameter-search framework. They are not universal accuracy values for the software.

Why Did scQuint and SCASL Perform Well Here?

In the benchmark, scQuint and SCASL identified relatively large numbers of splicing events while producing lower missing-value proportions than several alternatives. The authors therefore obtained denser cell-by-splicing matrices. That matters because splicing analysis is vulnerable to sparsity. If many cells have no usable junction evidence for an event, clustering may reflect missing data rather than biological differences. The study also reduced one source of unfairness by using a fixed number of top events with low missingness when comparing clustering performance.

Can Alternative Splicing Distinguish Cell Types?

Sometimes. A useful interpretation is: Splicing can provide complementary cell-state information when the biological populations differ in RNA processing and the sequencing protocol provides sufficient junction coverage. It may add less value when:

• junction coverage is weak;
• the cell states differ mainly in total gene expression;
• the relevant splicing events occur outside well-covered transcript regions;
• technical missingness is stronger than the biological effect.

The benchmark therefore supports the potential of splicing-based cellular representation. It does not prove that splicing will separate every cell type better than expression.

Benchmark 2 — Which Methods Detect Differential Splicing Most Reliably?

The second task was differential alternative splicing event, or DASE, detection. This is often the more important question for experimental studies. Researchers may want to compare:

• two cell types;
• two differentiation stages;
• treated versus untreated conditions;
• disease-associated versus reference states;
• multiple biological cohorts.

The Benchmark Used Several Evidence Layers

The authors compared single-cell DASE results with matched bulk RNA-seq from biologically matched purified populations. They also examined recovery of previously validated events and biological enrichment of genes containing detected splicing changes. Matched bulk data are useful as an orthogonal concordance reference, but they should not be treated as a perfect universal ground truth for every cell-level event. Bulk measurements average across cells and can miss rare state-specific events.

scQuint Performed Strongly in the Tested Datasets

The 2026 study reported that scQuint achieved the highest concordance with matched bulk differential-splicing results at both event and gene levels. It also ranked validated splicing events strongly and showed biologically relevant enrichment in the tested differentiation datasets. These findings support the statement: scQuint showed strong differential-splicing performance in this benchmark. They do not support: scQuint will be the most accurate method for every organism, protocol, annotation, or biological comparison.

More Detected Events Do Not Automatically Mean More True Events

This is an important interpretation trap. A method can produce more DASEs because of:

• broader event definitions;
• different imputation strategies;
• different missing-data handling;
• looser filtering;
• different significance models;
• more detected splice sites.

Therefore, DASE count alone is not an accuracy metric.

Cells Are Not Biological Replicates

Another common mistake is to treat thousands of cells from one biological specimen as thousands of independent replicates. They are not. For condition-level inference, donor-, specimen-, or experimental-unit replication remains essential. Cells sampled from the same specimen share biological and technical context. A rigorous differential-splicing design should distinguish: cell-level measurement from biological replication and should preserve sample identity throughout the statistical analysis.

Benchmark 3 — Computational Cost and Scalability

A method can be biologically attractive but impractical at project scale. The 2026 paper therefore compared runtime and memory as the number of cells increased. For 3,000 cells in the reported benchmark:

• scQuint and Psix completed processing within minutes;
• SCASL required more than 3.8 hours;
• BRIE2 required approximately 53 hours;
• SpliZ required approximately 16 days.

Peak memory at 3,000 cells was reported at approximately:

• 9 GB for BRIE2, scQuint, and Psix;
• 30 GB for MARVEL;
• 124 GB for SpliZ;
• 169 GB for SCASL.

These numbers should never be presented as fixed hardware requirements. They were observed under the study's specific:

• computing environment;
• software versions;
• preprocessing;
• data size;
• parameter settings;
• implementation choices.

A different server, software release, alignment strategy, or dataset can change them substantially.

Scalability Is Part of Method Selection

Runtime matters more as projects expand. A pipeline that is acceptable for 500 cells may become difficult when the project requires:

• thousands of cells;
• several donors;
• repeated parameter searches;
• multiple cell-type-specific contrasts;
• bootstrap or sensitivity analyses;
• reanalysis after annotation updates.

The practical question is therefore not simply: Which method performs well? It is also: Can the method be rerun, audited, and scaled within the actual study design?

What the Benchmark Does Not Prove

Benchmark papers are most useful when their limits are as clear as their rankings.

Figure 3. Benchmark rankings depend on sequencing protocol, event definition, biological task, and computational environment.

scQuint Is Not Universally “The Best”

The appropriate statement is: scQuint showed the strongest overall performance across the tested datasets and tasks in the 2026 benchmark. That is narrower—and more defensible—than declaring it universally superior.

Full-Length Short-Read Rankings Do Not Automatically Apply to 3' Droplet Data

The benchmark used Smart-seq-based full-length datasets. 3'-biased droplet libraries have different coverage patterns, different sparsity, and different junction information. A ranking derived from one protocol cannot be copied directly to another.

Clustering Agreement Does Not Equal Biological Truth

A cluster matching known cell labels is useful evidence. But high clustering performance can still be influenced by:

• sequencing depth;
• batch;
• annotation;
• missingness;
• feature selection;
• technical covariates.

Splicing-based separation should therefore be interpreted together with known biology and independent molecular evidence.

Rankings Can Change With Analysis Choices

Performance can depend on:

• RNA-seq aligner;
• splice-junction filters;
• transcript annotation version;
• minimum-read thresholds;
• imputation;
• number of selected events;
• dimensionality reduction;
• clustering parameters;
• software updates.

For this reason, reproducible reporting should document the complete analysis environment rather than only the final tool name.

Can 3' Droplet scRNA-seq Be Used for Alternative Splicing?

Yes—but with important limitations. 3'-tag single-cell RNA-seq was primarily designed for scalable gene-expression counting, not full-transcript isoform reconstruction.

Why 3' Bias Makes Splicing Harder

Reads are concentrated near transcript ends. Many alternative exons and splice junctions occur farther upstream. As a result, 3'-biased data can have:

• limited coverage of internal junctions;
• sparse PSI estimates;
• uneven event detection;
• reduced ability to reconstruct complete isoforms.

This makes comprehensive isoform-level analysis substantially more difficult than with full-length transcript protocols.

Useful Splicing Information Can Still Be Present

“Difficult” is not the same as “impossible.” A 2025 Nature Communications study introduced DOLPHIN, which integrates exon-level reads and splice-junction reads to improve cell representation and alternative-splicing detection under sparse single-cell conditions. The study demonstrates that exon and junction information in conventional scRNA-seq can still contain useful signal when analyzed appropriately. [5]. However, computational recovery cannot create junction coverage that was never sequenced.

When Not to Force Splicing Analysis From 3' Data

A new experiment should be considered when the primary goal is:

• discovery of novel isoform structures;
• coordinated exon usage across an entire transcript;
• transcript start and end characterization;
• full-length fusion transcripts;
• complex isoform switching;
• precise transcript architecture.

If the original protocol does not capture the required information, changing the software will not fully solve the problem. For broader experimental planning, see our Single-cell Sequencing Service.

Practical Method-Selection Framework

The most useful method depends on the data already available and the biological claim you need to make.

Scenario 1 — Full-Length Smart-seq Data and a Cell-Clustering Question

You have full-length short-read single-cell RNA-seq and want to know whether splicing adds information beyond gene expression. A practical path is:

1. confirm sufficient splice-junction coverage;
2. compare splicing matrix missingness;
3. evaluate an annotation-free method such as scQuint or SCASL;
4. compare splicing-based clusters with expression-based cell states;
5. test whether the splicing structure is reproducible across biological samples.

The 2026 benchmark supports scQuint and SCASL as reasonable starting points for this type of dataset, while computational cost may favor scQuint in larger analyses.

Scenario 2 — A Predefined Skipped-Exon Hypothesis

Suppose the project focuses on one known exon in a gene implicated by previous biology. The goal is not broad novel-event discovery. An annotation-dependent event model may be easier to interpret and validate. In this case, a narrower method can be preferable because it reduces the complexity of:

• event discovery;
• novel-junction filtering;
• transcript interpretation.

Broader discovery is only useful when the study needs it.

Scenario 3 — Existing 3' Droplet Data

Before selecting a splicing tool, first inspect whether the relevant splice junctions are actually covered. If not, the best decision may be:

• limit the analysis to well-supported events;
• use aggregation carefully;
• validate selected events independently;
• generate a more suitable full-length dataset for the next experiment.

Scenario 4 — Complete Isoform Structure Is the Main Question

If the project needs to know which exon combinations belong to the same transcript molecule, short-read event analysis may be insufficient. This is where full-length long-read sequencing becomes more valuable. Researchers can review our Single-Cell Full-Length RNA Sequencing Service for Isoform and Splicing Analysis.

Common Failure Modes and Troubleshooting

Single-cell splicing analysis is especially sensitive to coverage and event definition.

Too Few Junction Reads

A common symptom is a PSI matrix dominated by missing values. This can lead to:

• unstable event estimates;
• low overlap across cells;
• clustering driven by missingness;
• apparently extreme PSI values supported by very few reads.

Practical check: quantify junction coverage and missingness before interpreting biological patterns. If the event is not adequately covered, a more sophisticated model cannot fully replace missing information.

Coverage Bias Across the Transcript

3'- or 5'-biased libraries do not sample all splice junctions equally. Even full-length short-read protocols can show uneven coverage. An apparent difference between cell groups can therefore reflect differences in transcript coverage rather than true splice regulation. Practical check: inspect read and junction coverage around high-priority events and compare coverage distributions between groups.

PSI Instability at Low Read Counts

Percent spliced in, or PSI, is often represented from 0 to 1. At low coverage, values can become misleadingly extreme. For example, one inclusion read and no exclusion read can produce a value that appears fully included, even though the evidence is weak. Practical check: interpret PSI together with supporting read counts, missingness, and replicate-level consistency. A change from PSI 0 to PSI 1 should not automatically be called a complete biological isoform switch.

Cell-Type Composition Can Confound Comparisons

Suppose treatment A contains more cell type X than treatment B. A splicing difference in the pooled cells may simply reflect cell composition. Practical check: compare matched cell types or states when the biological claim is within-cell-type regulation. Separate:

• composition change;
• within-state splicing change.

Treating Cells as Independent Biological Replicates

This is one of the most important statistical failure modes. Cells from one donor are nested within that donor. They do not provide independent replication of a donor-level treatment effect. Practical check: preserve sample or donor identity and use a statistical design appropriate to the experimental unit.

Annotation-Version Dependence

Reference-guided methods depend on transcript annotation. Results can change when the reference changes. Practical check: record:

• genome build;
• gene annotation release;
• transcript annotation release;
• software version;
• event-definition rules.

Without these, the analysis is difficult to reproduce.

Imputation Can Create Apparent Structure

Imputation can reduce missingness and stabilize noisy matrices. It can also smooth real heterogeneity or create apparent similarity among cells. Practical check: confirm important findings with raw junction evidence and evaluate whether conclusions survive reasonable changes in imputation or filtering.

Different Tools Are Measuring Different Objects

One tool may report exon-skipping PSI. Another may report relative intron usage. Another may summarize splicing at the gene level. Practical check: harmonize biological interpretation before comparing the number or significance of detected events.

Batch Effects Can Affect Junction Detection

Different sequencing batches may differ in:

• read depth;
• fragment quality;
• alignment quality;
• transcript coverage;
• cell quality.

These technical differences can appear as splicing differences. Practical check: inspect batch-associated junction coverage and include biological replication across batches whenever possible. For broader transcriptomic QC principles, see How to Validate Single-Cell RNA-Seq Data?.

Short-Read, Full-Length Short-Read, or Long-Read?

Software selection comes after experimental design. The first decision is what molecular resolution the study needs.

Figure 4. Sequencing strategy should match the required resolution, from gene expression and junction usage to complete transcript isoform structure.

Long Reads Reduce—But Do Not Eliminate—Inference

A common oversimplification is that long-read sequencing “reads the isoform directly” and therefore needs no computational inference. Long reads do improve transcript structure resolution because individual reads can span much larger portions of an RNA molecule. However, analysis still requires:

• read QC;
• alignment;
• error handling;
• barcode recovery where applicable;
• transcript collapsing;
• splice-junction validation;
• isoform classification;
• filtering;
• quantification.

A 2025 Nature Communications study on Longcell specifically highlighted technical issues such as barcode recovery, read truncation, misalignment, and isoform quantification in long-read single-cell and spatial RNA data. [4]. Long-read QC is also a distinct discipline. The 2025 SQANTI-reads study showed that read coverage, junction-chain distributions, splice-site strength, and multisample consistency all affect interpretation. [6].

When Full-Length Short-Read Is Enough

Full-length short-read data remain a strong choice when the main goals are:

• local splice-junction usage;
• exon-skipping analysis;
• differential event detection;
• high-confidence cell-type-specific splicing;
• integration with deep short-read coverage.

Long-read sequencing adds most value when transcript connectivity itself is central to the question. For a broader comparison, see Short-Read vs Full-Length Transcriptomics in Single-Cell Research.

How to Validate a Single-Cell Splicing Finding

A computationally detected event is a research result, not automatically a mechanism. Single-cell splicing analysis can support:

• cell-type-specific splicing associations;
• differential junction usage;
• candidate isoform switching;
• hypotheses about RNA-processing regulation.

It does not by itself prove:

• a protein-level change;
• functional activity of the predicted isoform;
• causal regulation;
• a disease-driving mechanism;
• a treatment-response mechanism.

Match the Validation to the Claim

Useful validation options include:

• sashimi or junction-level visualization;
• independent biological replicates;
• comparison with full-length bulk RNA-seq;
• targeted RT-PCR;
• amplicon sequencing;
• targeted long-read sequencing;
• full-length long-read transcript confirmation;
• protein-level validation when the isoform changes coding sequence;
• functional perturbation when causality is claimed.

When Is Independent Validation Most Important?

Prioritize validation when:

• the event is novel;
• junction coverage is low;
• the effect is driven by a small subgroup;
• the event changes the main biological conclusion;
• a new isoform is proposed;
• a biomarker hypothesis is being developed;
• a mechanistic or causal claim is being made.

A recent Nature Reviews Genetics review emphasizes that modern splicing research increasingly combines high-resolution transcript discovery with direct functional testing, reinforcing the distinction between observing an isoform and demonstrating its biological effect. [3].

Single-Cell Alternative Splicing Analysis at CD Genomics

A useful analysis starts with the sequencing protocol and the biological output required.

Step 1: Identify the Data Type

Before selecting software, determine whether the input is:

• 3'-biased droplet scRNA-seq;
• full-length short-read single-cell RNA-seq;
• long-read full-length transcript data;
• aligned BAM files from an existing study;
• a new experiment that can still be redesigned.

This determines which splice junctions and transcript structures are actually observable.

Step 2: Define the Biological Output

Different questions require different outputs. Examples include:

• known exon-skipping events;
• novel splice junctions;
• differential alternative splicing;
• splice-site usage;
• cell-type-specific PSI patterns;
• isoform switching;
• full-length transcript structures;
• fusion transcripts;
• candidate events for validation.

Step 3: Build the Analysis Around Observable Evidence

Depending on data type and project scope, analysis outputs may include:

• read and alignment QC;
• splice-junction QC;
• event-level PSI matrices where appropriate;
• missingness assessment;
• cell-state annotation;
• differential splicing analysis;
• splice-site usage analysis;
• transcript or isoform annotation;
• junction and sashimi visualizations;
• functional enrichment;
• candidate validation lists.

Researchers with existing single-cell transcriptomic datasets can review our Single-Cell RNA-Seq Data Analysis Service.

Share the sequencing protocol, read configuration, biological replicate structure, and the splicing question you need to answer. The analysis can then be matched to the information actually present in the data rather than forcing every dataset into the same splicing workflow.

Frequently Asked Questions

Conclusion: Choose the Evidence Level Before Choosing the Tool

The 2026 six-method benchmark provides valuable guidance for single-cell alternative splicing analysis, particularly for full-length short-read datasets. Its most useful message is not that one software package wins every comparison. It is that analysis quality depends on the interaction among:

• sequencing protocol;
• junction coverage;
• splicing-event definition;
• missingness;
• biological replication;
• computational method;
• validation strategy.

When the question concerns local splice usage, full-length short-read data and dedicated splicing tools can be highly informative. When the question concerns complete transcript architecture, long-read sequencing may reduce dependence on reconstructing isoforms from short fragments. The best workflow therefore begins with the biological claim you want to support—and then selects the minimum sequencing and analytical evidence needed to support it.

References

1. Jiang Q, Li G, Xing L, Zhang D, Sun J. Comprehensive assessment of alternative splicing analysis methods for single-cell RNA-seq. iScience. 2026;29(6):116090. DOI: 10.1016/j.isci.2026.116090.
2. Wang F, Xing Y, Lin L. Beyond gene expression: Single-cell transcriptomics at isoform resolution. Trends in Genetics. 2026;42(8):722-737. DOI: 10.1016/j.tig.2026.05.013.
3. Sousa-Luís R, Carmo-Fonseca M. Tools and tactics for studying alternative splicing. Nature Reviews Genetics. 2026;27:642-660. DOI: 10.1038/s41576-026-00952-4.
4. Fu Y, Kim H, Roy S, et al. Single cell and spatial alternative splicing analysis with Nanopore long read sequencing. Nature Communications. 2025;16:6654. DOI: 10.1038/s41467-025-60902-2.
5. Song K, Zheng Y, Zhao B, et al. DOLPHIN advances single-cell transcriptomics beyond gene level by leveraging exon and junction reads. Nature Communications. 2025;16:6202. DOI: 10.1038/s41467-025-61580-w.
6. Keil N, Monzó C, McIntyre L, Conesa A. Quality assessment of long read data in multisample lrRNA-seq experiments using SQANTI-reads. Genome Research. 2025;35(4):987-998. DOI: 10.1101/gr.280021.124.

Research Use and Trust Statement

This article describes sequencing and computational approaches for research use.

• Single-cell splicing results should be interpreted in the context of sequencing protocol, read coverage, event definition, biological replication, and QC.
• The 2026 six-method benchmark used full-length, short-read Smart-seq-type datasets; its rankings should not be generalized automatically to 3' droplet, long-read, spatial, or other data types.
• Benchmark runtime, memory, cell counts, and performance values are study-specific observations rather than universal software or service specifications.
• More detected events do not automatically indicate greater accuracy.
• Cells from the same specimen are not independent biological replicates.
• PSI values supported by very few junction reads can be unstable.
• Long-read sequencing reduces reliance on transcript reconstruction from short fragments but does not eliminate computational analysis or QC.
• Novel, low-coverage, mechanistic, biomarker-related, or translational splicing findings should receive appropriate independent validation.
• Splicing association does not by itself demonstrate functional or causal effects.
• Services and analyses described here are intended for research use only.