CoChat
Sign Up
Cited by 12,700Genomics and Phylogenetic StudiesSingle-cell and spatial transcriptomicsGene expression and cancer classification

Salmon provides fast and bias-aware quantification of transcript expression

Open access
Patro, Rob, Duggal, Geet, Love, Michael I, Irizarry, Rafael A, Kingsford, Carl·2017·Nature Methods
DOI 10.1038/nmeth.4197PMC PMC5600148

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

We introduce Salmon, a lightweight method for quantifying transcript abundance from RNA-seq reads. Salmon combines a new dual-phase parallel inference algorithm and feature-rich bias models with an ultra-fast read mapping procedure. It is the first transcriptome-wide quantifier to correct for fragment GC-content bias, which, as we demonstrate here, substantially improves the accuracy of abundance estimates and the sensitivity of subsequent differential expression analysis.

Integrity signals

Registry

Crossref · 10.1038/nmeth.4197

Cites

36 of 36 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 26 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    Salmon attains the same order-of-magnitude speed gains as kallisto and Sailfish.

    “simultaneously achieves the same order-of-magnitude speed benefits as kallisto and Sailfish”
    Quote in the paperSubstantial claim
  • Findingown data

    Salmon's dual-phase inference and sample-specific bias models improve inter-replicate concordance relative to kallisto and eXpress.

    “Salmon’s dual-phase inference algorithm and sample-specific bias models yield improved inter-replicate concordance (Supplementary Fig. 4) compared to both kallisto and eXpress.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.
    • Rests on an assumption: Sequence-specific and fragment GC biases are assumed to have a conditional dependence.
  • Findingown data

    Salmon generally outperforms kallisto and eXpress in accuracy.

    “We show that Salmon typically outperforms both kallisto and eXpress in terms of accuracy”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.
  • Findingreasoning

    Salmon is more accurate than kallisto and eXpress.

    “Our novel quantification procedure, called Salmon, achieves greater accuracy than tools such as kallisto and eXpress”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.
  • Findingown data

    For differential expression testing, Salmon's quantification estimates give much higher sensitivity at the same false discovery rate, with sensitivity 53% to 250% higher.

    “the quantification estimates produced by Salmon exhibit markedly higher sensitivity at the same false discovery rate (Table 1c) — achieving a sensitivity 53% to 250% higher, at the same FDRs.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.
  • May combine two claims. 14
  • Classified as probably not a substantial claim. 15

  • May combine two claims.

    Findingown data

    Salmon provides accurate and fast quantification of transcript abundance from RNA-seq reads.

    “We introduce Salmon, a method for quantifying transcript abundance from RNA-seq reads that is accurate and fast.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.

  • May combine two claims.

    Findingown data

    Correcting fragment GC content bias markedly improves abundance estimate accuracy and the reliability of downstream differential expression analysis.

    “which we demonstrate substantially improves the accuracy of abundance estimates and the reliability of subsequent differential expression analysis”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.
    • The authors state this as an assumption: Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.
    • The authors state this as an assumption: All fragments are assumed to be generated independently.
    • The authors state this as an assumption: The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.
    • Rests on an assumption: Sequence-specific and fragment GC biases are assumed to have a conditional dependence.

  • May combine two claims.

    Findingown data

    Setting --biasSpeedSamp to 5 cuts additional bias correction time on sample ERR188021 from 119 to 20 seconds, with only a slight rise in false positive calls on GEUVADIS data.

    “setting --biasSpeedSamp to 5 reduces the additional bias correction time on sample ERR188021 mentioned above from 119 to 20 seconds, yet it leads to only a marginal increase in the number of false positive calls on the GEUVADIS data”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Salmon is built to use multiple CPU cores, and its mapping and inference procedures scale well with read count.

    “Salmon is designed take advantage of multiple CPU cores, and the mapping and inference procedures scale well with the number of reads in an experiment.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Salmon can quantify about 600 million 75bp paired-end reads in roughly 23 minutes with 30 threads, close to kallisto's roughly 20 minutes for the same task.

    “Salmon can quantify an experiment of approximately 600 million reads (75bp, paired-end) in ≈ 23 minutes (1384 seconds) using 30 threads— this roughly matches the speed of the recently-introduced kallisto, which takes ≈ 20 minutes (1198 seconds) to complete the same task using the same number of threads on the same machine”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    The first 8 GEUVADIS samples exhibited weak GC content dependence while the last 8 samples exhibited more severe fragment-level GC bias.

    “The first 8 GEUVADIS samples exhibited weak GC content dependence while the last 8 samples exhibited more severe fragment-level GC bias.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    No earlier transcriptome-wide quantification tool corrected for fragment GC content bias; Salmon does.

    “Salmon is the first transcriptome-wide quantifier to correct for fragment GC content bias”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The setup for continuous accuracy tracking overlaps considerably with the setup suggested for continuous analysis by Beaulieu-Jones and Greene.

    “We note that this setup overlaps considerably with the setup suggested for “continuous analysis” by Beaulieu-Jones and Greene [32].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Brett, K.B.-J. & Greene, C.S. Preprint at https://doi.org/10.1101/056473 (2016). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    With enough training data, one can learn a separate bin for each read position.

    “given sufficient training data, it is even possible to learn a separate bin for each position of a read [15].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] Roberts, A. & Pachter, L. Nat. Methods 10, 71–73 (2013). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The bootstrap sampling process is inspired by kallisto.

    “Additionally, inspired by kallisto [10], Salmon also provides the ability to draw bootstrap samples, which is an alternative way to assign confidence to the estimates returned by the main inference algorithm.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [10] Bray, N.L., Pimentel, H., Melsted, P. & Pachter, L. Nat. Biotechnol. 34, 525–527 (2016). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Lack of synchronization in the online phase means the most up-to-date values of α may not be used when estimating xy.

    “This lack of synchronization means that when estimating xy, we cannot be certain that the most up-to-date values of α are being used.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Even a single spatially homogeneous error profile seems to perform reasonably well.

    “even a single, spatially homogeneous error profile appears to work reasonably well [22]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Nariai, N. et al. BMC Genomics 15 (Suppl. 10), S5 (2014). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Salmon's approach is unique in combining experimental data and bias models with an efficient dual-phase inference procedure.

    “Salmon’s approach is unique in how it combines models of experimental data and bias with an efficient dual-phase inference procedure.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Polyester allows simulating experiments with differential transcript expression and biological variability, enabling assessment of accuracy and downstream differential expression performance.

    “In addition to the ability to generate reads, Polyester allows simulating experiments with differential transcript expression and biological variability. Thus, we can assess not only the accuracy of the resulting estimates, but also how these estimates would perform in a typical downstream analysis task like differential expression testing.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [13] Frazee, A.C., Jaffe, A.E., Langmead, B. & Leek, J.T. Bioinformatics 31, 2778–2784 (2015). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Very small non-zero expression values may negatively affect the MARD metric, so truncating TPMs below 0.01 before computing MARD mitigates this.

    “However, eXpress does not perform such a truncation, and very small, non-zero values may have a negative effect on the MARD metric. To mitigate such effects, we first truncate to 0 all TPMs less than 0.01 before computing the MARDs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The SCVB0 inference algorithm is essentially equivalent to variants of the online-EM algorithm with a modified prior.

    “As outlined in [19], the SCVB0 inference algorithm is essentially equivalent to variants of the online-EM [23] algorithm with a modified prior.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [19] Foulds, J., Boyles, L., DuBois, C., Smyth, P. & Welling, M. in Proc. 19th ACM SIGKDD Int. Conf. Knowledge Discov. & Data Mining 446–454 (ACM, 2013). source

      Not checked in the free run.

    • cites [23] Cappé, O. in Mixtures: Estimation and Applications (eds. Mengersen, K.L., Robert, C.P. & Titterington, D.M.) Ch. 2 (John Wiley & Sons, 2011).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Salmon and kallisto truncate very tiny expression values to 0 by default, while eXpress does not.

    “We note that Salmon and kallisto, by default, truncate very tiny expression values to 0. For example, any transcript estimated to produce < 1 × 10−8 reads is assigned an estimated read count of 0 (which, likewise, affects the TPM estimates). However, eXpress does not perform such a truncation”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    A parallel self-hosted continuous integration system runs Salmon on simulated samples on each commit, stores accuracy metrics, and allows tracking accuracy and identifying performance regressions across commits.

    “For this task, a parallel, self-hosted continuous integration system has been created. On each commit to the Salmon repository, a Drone (Drone-CI) server pulls the latest commit, and builds it in a clean CentOS5 environment using Docker [30]. In addition to the functionality tests, Salmon is run on simulated samples (generated using Polyester [13]).”
    • Also stated in: § Validation · plain
      “This allows us to track accuracy over various Salmon commits, and to identify the commit corresponding to any performance regressions.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [30] Merkel, D. Linux J. 2014 (2014).

      Not checked in the free run.

    • cites [13] Frazee, A.C., Jaffe, A.E., Langmead, B. & Leek, J.T. Bioinformatics 31, 2778–2784 (2015). source

      Not checked in the free run.

    • cites [31] Di Tommaso, P., Chatzou, M., Baraja, P.P. & Notredame, C. figshare https://dx.doi.org/10.6084/m9.figshare.1254958.v2 (2014). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The likelihood over rich equivalence classes with aggregate weights is not exactly equivalent to the full fragment likelihood, but the weights allow consideration of conditional probabilities without continuously reconsidering each fragment.

    “Though the likelihood function over equivalence classes that considers these weights (Equation (10)) is no longer exactly equivalent to the likelihood defined over all fragments (Equation (9)), these weights nonetheless allow us to take into consideration the conditional probabilities specified in the full model, without having to continuously reconsider each of the fragments in ℱ.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Because the bias model evaluates bias along reference transcriptome bases, its cost scales with the number of expressed transcripts rather than reads, making the extra cost nearly independent of fragment count.

    “since the bias model works by evaluating the bias along bases of the reference transcriptome, it scales in the number of active (i.e., expressed) transcripts rather than in the number of reads, so the extra cost of bias modeling is almost entirely independent of the number of fragments in an experiment.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Gibbs sampling procedure for posterior sampling adopts the model of Turro et al.

    “To perform this Gibbs sampling, we adopt the model of Turro et al. [27] (details in Supplementary Note 2).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27] Turro, E. et al. Genome Biol. 12, R13 (2011). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The inference procedure is generic over the type of alignments being processed, accepting either regular alignments or quasi-mappings.

    “The inference procedure itself is generic over the type of alignments being processed; they may be either regular alignments (e.g. coming from a bam file), or quasi-mappings computed from the raw reads (e.g. coming from FASTA/Q files).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    Testing during Salmon development is highly automated, with continuous integration that builds each commit and runs functionality tests, but these tests do not assess quantification accuracy.

    “Testing during the Salmon development and release process is highly-automated.”
    • Also stated in: § Validation · plain
      “A publicly-facing continuous integration service (Travis-CI) automatically builds each commit, and runs a small set of functionality tests to ensure that no breaking changes have been committed. However, these tests do not assess or track quantification accuracy.”
    Quote in the paperNot a substantial claim
23 of 23

  • The authors state this as an assumption.

    For a given sequencing experiment, the true transcriptome and its transcript copy numbers are treated as fixed and known.

    “Assume that, for a particular sequencing experiment, the underlying true transcriptome is given as 𝒥 = {(t1, … , tM), (c1, … , cM)}, where each ti is the nucleotide sequence of some transcript (an isoform of some gene) and each ci is the corresponding number of copies of t in the sample.”

  • The authors state this as an assumption.

    Without experimental bias, fragments are assumed to be sampled in proportion to transcript copy number times effective length.

    “We adopt a generative model of the sequencing experiment that dictates that, in the absence of experimental bias, library fragments are sampled proportional to ci · ℓ̃ι.”

  • The authors state this as an assumption.

    All fragments are assumed to be generated independently.

    “if we assume that all fragments are generated independently”

  • The authors state this as an assumption.

    The auxiliary model parameters are assumed not to depend on the current transcript abundance estimates.

    “This probability, in turn, depends on auxiliary models whose parameters do not explicitly depend upon the current estimates of transcript abundances.”

  • The paper relies on this without support.

    Sequence-specific and fragment GC biases are assumed to have a conditional dependence.

    “Additionally, we note that sequence-specific and fragment GC biases do seem to display a conditional dependence.”

  • The authors state this as an assumption.

    Samples from the same human population are expected to have few to no true differences in transcript abundance.

    “Because the samples are from the same human population, it is expected that there would be few to no true differences in transcript abundance produced by this comparison.”

  • The paper relies on this without support.

    RSEM estimates on sample NA12716_7 represent true expression.

    “RSEM was run on sample NA12716_7 from the GEUVADIS RNA-seq data to learn model parameters and estimate true expression, and the learned model was then used to generate 20 different simulated datasets”
7 of 7

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • PSAP variants are associated with lysosomal storage disorders and Parkinson's Disease.

    “Given the screen phenotypes of PSAP, which were unexpected based on its known functions, and the association of PSAP variants not only with lysosomal storage disorders60–62 but also Parkinson’s Disease63, we decided to further investigate the underlying mechanisms of PSAP in neuronal redox regulation.”

    From Genome-wide CRISPRi/a screens in human neurons link lysosomal failure to ferroptosis (2021) · cites this paper as [61]

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.

  • Fragment-origin probability is also defined as zero

    Minor

    The text identifies this conditional probability as the chance of generating a fragment from its source transcript, then says the same probability is uniformly zero. If it were zero for every fragment, the stated likelihood model could not assign positive probability to any observation.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 5

  1. Bishop, C.M. et al. Pattern Recognition and Machine Learning (Springer, 2006).
  2. Cappé, O. in Mixtures: Estimation and Applications (eds. Mengersen, K.L., Robert, C.P. & Titterington, D.M.) Ch. 2 (John Wiley & Sons, 2011).
  3. Hsieh, C.-J., Yu, H.-F. & Dhillon, I.S. ICML 15, 2370–2379 (2015).
  4. Li, X., David, G., Andersen, M.K. & Freedman, M.J. in Proc. Ninth Eur. Conf. Computer Syst. 27 (ACM, 2014).
  5. Jackman, S. & Birol, I. F1000Research 5, 1795 (2016).

Cited by

Sorted by citations

Related

Authors
Patro, R., Duggal, G., Love, M. I., Irizarry, R. A., & Kingsford, C.
Affiliations
  • Stony Brook University (US)
  • DNAnexus (United States) (US)
  • Harvard University (US)
  • Carnegie Mellon University (US)
Corresponding author
  • Patro, Rob
  • Kingsford, Carl
ORCID
Venue
Nature Methods
Publisher
Springer Science and Business Media LLC
Published
2017-03-06
Volume
14
Issue
4
Pages
417-419
ISSN
  • 1548-7091 (ISSN-L)
  • 1548-7105
DOI
10.1038/nmeth.4197
PMC
PMC5600148
OpenAlex
W2592811885
Semantic Scholar
30ca32529ca227263b3a1e5986fb57b22afde2bd
Cited by
12,700
References
36
Citations by year
2026: 1822 · 2025: 2138 · 2024: 2056 · 2023: 1605 · 2022: 1559 · 2021: 1474 · 2020: 978 · 2019: 613 · 2018: 301 · 2017: 107 · 2016: 40 · 2015: 2 · 2012: 1
FWCI
226.51
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Methods (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • Digital Access to Scholarship at Harvard (DASH) (Harvard University) (submittedVersion, cc-by)
Funders
  • National Science Foundation (1564917, 1319998, 1256087)
  • Gordon and Betty Moore Foundation (GBMF4554)
  • Carnegie Mellon University
  • National Institutes of Health
  • Directorate for Biological Sciences
  • Biotechnology and Biological Sciences Research Council
  • National Human Genome Research Institute (R21HG006913, R01HG005220, R01HG007104)
  • National Cancer Institute (T32CA009337)
  • National Institute of General Medical Sciences (R01GM083084)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene expression and cancer classification
Keywords
salmon, RNA-seq, bias correction, differential expression analysis
MeSH terms
Algorithms, Base Composition, Bayes Theorem, Sequence Analysis, RNA, Gene Expression Profiling

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

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