henhaobio
Sign in

Nature MethodsBioinformaticsPeer-reviewed journal article

SQANTI3: curation of long-read transcriptomes for accurate identification of known and novel isoforms

Francisco J. Pardo-PalaciosiD, Ángeles Arzalluz-LuqueiD, Liudmyla KondratovaiD, Pedro SalgueroiD, Jorge Mestre‐TomásiD, Rocío AmoríniD, Eva Estevan-MorióiD, Tianyuan LiuiD, Adalena V. NanniiD, Lauren M. McIntyreiD, Elizabeth TsengiD, Ana ConesaiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The emergence of long-read RNA sequencing (lrRNA-seq) has provided an unprecedented opportunity to analyze transcriptomes at isoform resolution. However, the technology is not free from biases, and transcript models inferred from these data require quality control and curation. In this study, we introduce SQANTI3, a tool specifically designed to perform quality analysis on transcriptomes constructed using lrRNA-seq data. SQANTI3 provides an extensive naming framework to describe transcript model diversity in comparison to the reference transcriptome. Additionally, the tool incorporates a wide range of metrics to characterize various structural properties of transcript models, such as transcription start and end sites, splice junctions, and other structural features. These metrics can be utilized to filter out potential artifacts. Moreover, SQANTI3 includes a Rescue module that prevents the loss of known genes and transcripts exhibiting evidence of expression but displaying low-quality features. Lastly, SQANTI3 incorporates IsoAnnotLite, which enables functional annotation at the isoform level and facilitates functional iso-transcriptomics analyses. We demonstrate the versatility of SQANTI3 in analyzing different data types, isoform reconstruction pipelines, and sequencing platforms, and how it provides novel biological insights into isoform biology. The SQANTI3 software is available at https://github.com/ConesaLab/SQANTI3.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (3 Jun 2023) · CC BY 4.0

SQANTI3 extends long-read transcriptome quality control with transcript-end subcategories, a short-read TSS ratio, rules or random-forest filters and a rescue module, raising SIRV-benchmarked F-score stepwise.

Long-read RNA-seq uncovers thousands of unannotated transcripts, yet errors from RNA degradation, library preparation and mapping produce false models, and the original SQANTI mainly targeted splice-junction errors rather than variability at 5' and 3' ends. SQANTI3 adds end-site evidence from CAGE, Quant-seq or short reads, flexible artifact filters, recovery of discarded known genes and isoform-level functional annotation, and it serves as the LRGASP evaluation engine.

Key results

  1. ShownSuccessive SQANTI3 curation steps improved accuracy on spike-in controls. The random-forest filter removed false positives more effectively than the rules filter, and true novel isoforms stayed detectable.

    SIRV ground truth (69 short SIRVs; 39 decoy annotations added, 26 true isoforms hidden): IsoSeq3 found 65 of 69 (94%) at 32% precision; ML filter with sample-specific data (HIS) removed 100% of false positives; F-score rose more with ML (82%) than rules (62%) filtering; rules-HIS kept 15 over-annotated FP and 35 partial TPs; rescue selected only 1 to 3 FP. · 2.5 Validation of the SQANTI3 curation approach; 4.2.1 WTC11 dataset; 4.3.6 SIRV evaluation metrics

  2. ShownA short-read TSS ratio, which compares coverage just downstream and just upstream of each start site, agrees well with CAGE-seq. It gives a cheap way to check 5' ends, and polyA motifs similarly support 3' ends.

    WTC11: CAGE-supported TSSs had higher TSS ratios (Wilcoxon p = 2e-16); a ratio >1.5 confirmed 88.2% of CAGE-supported transcripts, refTSS supported 92.6% of TSSs above 1.5, and such sites were likelier to have external support (Fisher's exact p < 2e-16). PolyA motif presence was significantly associated with Quant-seq 3' validation. · SQANTI3 validates 3’ and 5’ ends with additional…; 2.2 SQANTI3 QC of a WTC11 human cell-line…; Processing of short read data

  3. SuggestedIn the WTC11 IsoSeq3 transcriptome, most full splice matches (FSM) have ends that differ from the reference. Incomplete splice matches (ISM) are mostly 3' fragments, few of which have 5' end support, which fits a mix of 5' degradation and real but unannotated end sites.

    Of 228,379 models, 56,795 (24.87%) were FSM and 67,804 (29.69%) ISM; only 22.5% (12,805) of FSM were Reference Match and 58.8% varied at 3' ends; 3' fragments made up 70.2% of ISM (47,594), with 9.4% known TSS, 16% CAGE support and 39.5% TSS ratio above threshold; 4,591 FSM had unsupported low-ratio TSSs. · 2.2 SQANTI3 QC of a WTC11 human cell-line…

  4. ShownThe Rescue module restores known genes and transcripts that filtering removed entirely. Rescued loci have higher short-read expression and higher APPRIS TRIFID functional scores than loci that stayed excluded.

    Mean recovery 10,389 isoforms and 2,884 genes after ML filtering vs 36,811 transcripts and 2,638 genes after rules filtering. ML-HIS: of rescue candidates (113,320 in total), 59.3% matched a reference and 29.7% a long-read isoform, 11% stayed unassigned; 11,599 unique targets added, 94.1% via automatic rescue. · 2.4 The SQANTI3 rescue module to enhance…; 4.1.3 SQANTI3 Rescue module

  5. SuggestedIsoform-level functional annotation of a curated H1 hESC-to-endoderm transcriptome found splicing-driven domain changes. One example is an ACOX2 isoform switch that adds peroxisomal Acyl-CoA oxidase domains while gene-level expression stays unchanged.

    H1 hESC vs endoderm: 65,255 curated models from 14,541 known genes; 82.7% received at least one functional feature; 4,247 (29.25%) genes and 21,489 transcripts differentially expressed (FDR < 0.05) vs 450 genes with differential isoform usage (FDR < 0.05), enriched for ubiquitination sites, nuclear localization signals and RNA recognition domains. · 2.7 SQANTI3 supports Functional…; 4.2.3 H1-endoderm differentiation dataset; 4.3.5 Running IsoAnnotLite and tappAS

Methods, data and reproducibility

SQANTI3 compares long-read transcript models with a reference, assigns structural (sub)categories and computes junction, end-site and expression descriptors from short reads, CAGE or Quant-seq; JSON rules or a random forest flag artifacts, a minimap2-based rescue restores replacements, and IsoAnnotLite transfers isoform-level functional features.

  • WTC11 human iPSC line, cDNA PacBio lrRNA-seq from LRGASP with Lexogen SIRV-Set 4 spike-ins; transcriptome reconstructed with IsoSeq3 · human · n = 228,379 transcript models (209,220 in 17,467 known genes) · ENCODE ENCSR507JOF
  • WTC11 matched Illumina short-read RNA-seq · human · ENCODE ENCSR673UKZ
  • WTC11 CAGE-seq and Quant-seq peaks, kept if found in at least two replicates · human · n = 46,722 CAGE-Seq and 45,813 Quant-Seq peaks · GEO GSE185917 (CAGE); ENCODE ENCSR322MWL (Quant-seq)
  • Lexogen SIRV-Set 4 spike-in controls with modified annotation (39 decoy models added, 26 true isoforms removed) as ground truth · n = 69 short SIRVs, 15 long SIRVs, 92 ERCC controls
  • K562 direct RNA ONT long-read transcriptome reconstructed with the ENCODE TALON pipeline · human · n = 336,445 isoforms (95% in 51,843 known genes) · ENCODE ENCSR917JIA; ENCFF584GRG (TALON); ENCFF668BLB (quantification)
  • K562 single-end Illumina RNA-seq and CAGE-seq used as orthogonal data · human · n = two Illumina samples; 9,248 CAGE-Seq peaks · ENCODE ENCSR792OIJ (Illumina); ENCSR000CJN (CAGE)
  • H1 hESC and in vitro definitive endoderm cDNA PacBio lrRNA-seq (LRGASP), pooled into one IsoSeq3 transcriptome · human · n = 211,107 transcript models (89% in 17,667 known genes) · ENCODE ENCSR271KEJ (H1); ENCSR127HKN (endoderm)
  • H1 and endoderm Illumina short reads plus sample-specific Quant-seq peaks · human · n = 52,558 Quant-Seq peaks · ENCODE ENCSR588EJX, ENCSR266XAJ (Illumina); ENCSR198UNH (Quant-seq)
Methods & tools
SQANTI3 v5.0 (QC, Filter, Rescue modules), IsoSeq3 v3.4.0, cDNA Cupcake (isoform collapse), minimap2, STAR (two-pass mode), Kallisto, TALON (ENCODE pipeline), IsoAnnotLite, tappAS (functional diversity, DE, DIU, enrichment), GENCODE v39 human annotation, refTSS v3.1 and polyASite v2.0, APPRIS (TRIFID scores)
Software & environment
SQANTI3 v5.0 (Python and R); IsoSeq3 v3.4.0 (ccs --min-rq 0.9); GENCODE v39; refTSS v3.1; polyASite v2.0; STAR --twopassMode; Kallisto --bootstrap-samples 100; minimap2 -x map-hifi.
Reproducibility gaps
Exact IsoSeq3 commands and STAR/Kallisto parameters are given only in Supplementary Methods, which are not part of the text. SQANTI3 QC cannot process single-end short reads, so Kallisto was run separately for K562.
Limitations (6)
  • CAGE-seq and Quant-seq end-site data are costly and hard to obtain routinely. The authors offer the short-read TSS ratio and polyA motif detection as cheaper substitutes. · stated by the authors
  • ML filtering depends heavily on the true and false training sets the user chooses. For the reference-guided K562 TALON transcriptome, poorly supported FSMs and a sparse CAGE set (9,248 peaks) compromised the training set, so the authors used a tailored rules filter and a mapping-only rescue instead. · stated by the authors
  • The rules filter depends on the reference annotation, so it kept over-annotated false SIRVs (15 over-annotated FP and 35 partial TPs passed rules-HIS) and left the over-annotation detection rate nearly flat. · stated by the authors
  • Filtering lowered sensitivity by flagging some true SIRVs as artifacts. Rescue recovered most of them, but not the novel true isoforms missing from the reference. · stated by the authors
  • Accuracy is measured on one spike-in set in a single PacBio cDNA dataset. SQANTI3 curation is not compared head-to-head with other long-read QC or filtering tools, so its relative performance is untested. · noted in this summary
  • The functional conclusions for ACOX2, ETS1 and CDYL rest on short-read quantification and IsoAnnotLite annotation transfer. No protein-level or experimental work confirms the predicted domain or SUMO2-site changes. · noted in this summary

Read this if you build or curate PacBio or ONT long-read transcriptomes and need guidance on validating transcript ends, choosing rules versus machine-learning filters, or rescuing known genes lost to filtering.

Citations
212
FWCI
28.46
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTardáguila '18Lopez '26Paniagua '26Mestre‐Tomá '23Yu '26Keil '25
73% semantic matchBioinformatics
Manuel Tardáguila, Lorena de la Fuente, Cristina Martí+16 more·Genome Research·Mar 1, 2018· 458 citations