henhaobio
Sign in

Genome biologyGenomicsPeer-reviewed journal article

Transcriptome assembly from long-read RNA-seq alignments with StringTie2

Sam KovakaiD, Aleksey V. ZiminiD, Geo M. PerteaiD, Roham RazaghiiD, Steven L. SalzbergiD, Mihaela PerteaiD

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

Abstract

RNA sequencing using the latest single-molecule sequencing instruments produces reads that are thousands of nucleotides long. The ability to assemble these long reads can greatly improve the sensitivity of long-read analyses. Here we present StringTie2, a reference-guided transcriptome assembler that works with both short and long reads. StringTie2 includes new computational methods to handle the high error rate of long-read sequencing technology, which previous assemblers could not tolerate. It also offers the ability to work with full-length super-reads assembled from short reads, which further improves the quality of assemblies. On 33 short-read datasets from humans and two plant species, StringTie2 is 47.3% more precise and 3.9% more sensitive than Scallop. On multiple long read datasets, StringTie2 on average correctly assembles 8.3 and 2.6 times as many transcripts as FLAIR and Traphlor, respectively, with substantially higher precision. StringTie2 is also faster and has a smaller memory footprint than all comparable tools.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (8 Jul 2019) · CC BY 4.0

StringTie2 assembles error-prone long reads, recovering on average 8.3 and 2.6 times as many correct transcripts as FLAIR and Traphlor, and outperforms Scallop on short reads.

Short reads rarely span more than two exons, leaving multi-isoform splice graphs ambiguous, whereas long reads can resolve isoforms but carry error rates of 8-10% or higher that Traphlor handled poorly, and FLAIR only reports annotated transcripts fully covered by reads. StringTie2 puts error-tolerant long-read assembly that can find unannotated isoforms, plus short-read super-reads, into one faster and leaner reference-guided tool.

Key results

  1. ShownStringTie2 was more sensitive and precise than Scallop on simulated human short reads and on real Illumina data from human, Arabidopsis and maize. Its only loss was on precision for one maize run, which a larger allowed alignment gap plus disabled end trimming reversed.

    33 real datasets (23 human, five Arabidopsis, five Z. mays): mean relative gains of 3.9% in sensitivity and 47.3% in precision. On maize ERR986144, Scallop precision was 17.4% vs 16.3%, but StringTie2 had a 24% relative sensitivity gain; with -t -g 200 it led on all maize runs. · Abstract; Transcriptome assembly of short RNA-seq reads; Alignment and assembly parameters

  2. ShownOn eight real human long-read datasets (PacBio full-length and fragment sets, ONT cDNA and direct RNA), StringTie2 correctly assembled far more annotated transcripts than FLAIR and Traphlor, with much higher precision. It still beat FLAIR without any guide annotation.

    Traphlor produced nothing on ONT cDNA; on the other datasets StringTie2 averaged 9564 vs 3708 correct transcripts (2.6 times). With a guide, StringTie2 found 16,000 more correct transcripts than FLAIR on average, at roughly three to six times higher precision; on direct RNA, 29,744 vs 4,442 (6.7 times). · Abstract; Transcriptome assembly of third-generation…

  3. SuggestedOn simulated long reads, guided StringTie2 beat FLAIR on every dataset. When the guide annotation was randomly subsampled, FLAIR accuracy fell sharply while StringTie2 stayed accurate, which indicates it can assemble transcripts missing from the annotation.

    PBSIM PacBio/ONT sets for yeast, fly and human chr19: Traphlor trailed both tools except on S. cerevisiae PacBio, where it beat FLAIR. Human chromosome 19 ONT with 1% to 100% of the annotation: at 20%, FLAIR had 30% sensitivity and 50% precision vs 74% and 80% for StringTie2. · Transcriptome assembly of third-generation…; Simulated Data

  4. ShownBuilding super-reads from short reads (MaSuRCA k-unitig extension with the error-correction count thresholds removed, coverage split by expectation-maximization) improved StringTie2 sensitivity, precision and expression estimates.

    Both metrics rose on all human datasets and on all but three plant datasets (one Z. mays: precision up, sensitivity unchanged; two Arabidopsis: sensitivity up, precision unchanged). On the 150 million-read simulation, super-reads raised sensitivity, precision and Spearman correlation with true expression. · Transcriptome assembly of short RNA-seq reads; Super-Read Construction and Quantification; Tables

  5. ShownStringTie2 used far less memory than StringTie1, which the authors credit to new data structures (collapsing identically aligned reads, a sparse bit-vector splice graph), and on real data it ran faster with a smaller memory footprint than Scallop and FLAIR.

    Memory cut on average by a factor of 40 vs StringTie1 on three shared datasets. On real short reads, 1.8 times faster with 17 times less memory than Scallop. On real long reads vs FLAIR: 68 times faster and 9 times less memory with annotation; 93 times faster and 27 times less memory without. · New data structures in StringTie2 compared to…; Transcriptome assembly of short RNA-seq reads; Transcriptome assembly of third-generation…

Methods, data and reproducibility

StringTie2 extends StringTie1's network-flow assembly with collapsed read storage and sparse bit-vector splice graphs; for error-prone long reads it shifts unsupported splice sites to nearby well-supported ones and prunes oversized splice graphs, and it can weight paths using MaSuRCA-derived super-reads with expectation-maximization coverage estimates.

  • Human Illumina RNA-seq runs previously used in the Scallop evaluation (three also used in the original StringTie study) · human · n = ten datasets · SRR307903, SRR315323, SRR315334, SRR534307, SRR545723, SRR307911, SRR387661, SRR534291, SRR534319, SRR545695
  • Randomly selected GEUVADIS human short-read RNA-seq samples · human · n = 13 samples · GEUVADIS: ERR188021, ERR188022, ERR188041, ERR188060, ERR188063, ERR188073, ERR188107, ERR188311, others
  • Arabidopsis thaliana short-read RNA-seq runs aligned to TAIR10 · Arabidopsis thaliana · n = five datasets · ERR1886195, ERR1886317, ERR1886341, ERR1886384, ERR1886412
  • Zea mays short-read RNA-seq runs aligned to the B73 RefGen assembly · Zea mays · n = five datasets · ERR986085, ERR986103, ERR986106, ERR986108, ERR986114
  • Flux Simulator short-read set built from all protein-coding and lncRNA transcripts on hg38 main chromosomes · human (hg38) · n = 150 million 75-bp paired-end reads
  • PBSIM long reads from Krizanovic et al. (2017): yeast PacBio, fruit fly PacBio and ONT, human chr19 PacBio and ONT · S. cerevisiae S288, D. melanogaster r6, human GRCh38.p7 · n = five simulated datasets
  • PacBio human tissue isoform sequencing data: full-length-enriched (PacBioFL) and fragment-containing (PacBioNFL) sets · human · n = three PacBioFL and three PacBioNFL datasets
  • Oxford Nanopore cDNA and direct RNA-seq runs from the NA12878 RNA sequencing consortium · human (NA12878) · n = one cDNA and one direct RNA dataset
Methods & tools
StringTie2, Scallop, FLAIR, Traphlor, HISAT2, minimap2 (-splice), GMAP, MaSuRCA (with QuorUM error correction), Flux Simulator, PBSIM, gffcompare, RefSeq GRCh38.p8 annotation
Software & environment
StringTie2 in C++, dependency-free, multi-threaded; references hg38 with RefSeq GRCh38.p8, TAIR10, B73 RefGen (MaizeGDB); long reads run with -L; minimap2 -splice; maize reruns -t -g 200 (Scallop --min_bundle_gap 200); no tool versions given.
Reproducibility gaps
No hardware or absolute runtimes in the main text (relative figures only, Supplementary Tables S3, S5); ONT NA12878 consortium data lack accessions; Scallop, FLAIR and Traphlor versions unstated; maize outlier cited as both ERR986144 and ERR986114.
Limitations (6)
  • On real data, the truth set is the union of annotated transcripts that any compared tool assembled correctly, so absolute sensitivity is inflated whenever no tool recovers a transcript; only relative comparisons between tools are considered reliable. · stated by the authors
  • Default settings are not optimal for every dataset: maize samples with coverage gaps (possibly from the repetitive genome) needed the allowed alignment gap raised from 50bp to 200bp and end trimming disabled before StringTie2 overtook Scallop on precision. · stated by the authors
  • Long-read throughput is low enough that only the most highly expressed genes can be quantified, so accurate transcript-level abundance still relies on short-read sequencing. · stated by the authors
  • Long-read errors are mostly insertions and deletions, so alignments disagree around splice sites and small exons; splice-site correction does not remove spurious exons from random insertions, and support-based edge pruning removes only some false edges. · stated by the authors
  • StringTie2 defaults were tuned on the same simulated human short-read set later used to compare it with Scallop, which could favour StringTie2 in that simulation benchmark. · noted in this summary
  • On real data, unannotated long-read assemblies count as false positives and are never independently validated, so the claimed advantage over FLAIR in finding novel isoforms rests mainly on the simulated annotation-subsampling test. · noted in this summary

Read this if you assemble transcripts from PacBio or Oxford Nanopore RNA-seq alignments, or want to compare StringTie2, Scallop, FLAIR and Traphlor on accuracy, speed and memory.

Citations
1,989
FWCI
48.12
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
THISShumate '22Tung '19Zhang '21Nip '23Yan '25Shao '17
82% semantic matchBioinformatics
Alaina Shumate, Brandon Y. Wong, Geo M. Pertea+1 more·PLoS Computational Biology·Jun 1, 2022· 482 citations