Salmon

Salmon employs a two-stage statistical procedure to maximize quantification accuracy. The online phase rapidly processes reads using lightweight mapping and initial abundance estimates. The offline phase refines these estimates using an advanced variational Bayesian inference model, correcting for fragment-level biases and multimapping reads that traditional alignment-free methods often mishandle.

Instead of performing computationally expensive full base-to-base alignment, Salmon uses quasi-mapping to rapidly identify the transcript of origin for each read. This process builds a lightweight index of k-mers and their suffix array intervals, allowing the algorithm to locate a read's potential mapping loci in a fraction of the time required by traditional aligners like STAR or HISAT2.

Salmon explicitly learns and corrects for complex, sample-specific biases that distort abundance estimates:

• Sequence-specific bias: Corrects for non-uniformity caused by nucleotide composition at fragment ends.
• GC-content bias: Accounts for amplification and sequencing efficiency variations correlated with guanine-cytosine content.
• Positional bias: Models the non-uniform distribution of read start positions across transcripts. These models are estimated directly from the input data without requiring external controls.

Beyond raw count estimates, Salmon provides a statistically rigorous output suite. It reports Transcripts Per Million (TPM) for cross-sample comparability and inferential replicates—multiple bootstrap or Gibbs sampling draws from the posterior distribution of abundance estimates. These replicates enable downstream tools like swish in the DESeq2 framework to perform differential expression analysis with proper variance estimates, even for experiments with few biological replicates.

To mitigate the quantification noise introduced by spurious mappings, Salmon implements selective alignment. This scoring mechanism evaluates the quality of a quasi-mapping hit against a probabilistic model of a true alignment. Reads that fail to meet a sensitivity-tuned score threshold are discarded as mapping artifacts, dramatically reducing false positive quantification of intergenic or repetitive regions without sacrificing sensitivity to genuine lowly-expressed transcripts.

Salmon's index can incorporate decoy sequences—genomic regions such as the whole genome scaffold that are known to be unannotated. During quasi-mapping, reads that map better to these decoys than to the transcriptome are flagged and removed from quantification. This decoy-aware strategy prevents the algorithm from forcibly assigning reads from unannotated loci or background DNA contamination to real transcripts, preserving the accuracy of abundance estimates.

Transcripts Per Million is the standard unit for reporting transcript abundance. TPM corrects for both gene length and sequencing depth, ensuring that the sum of all TPM values in a sample equals one million. This allows direct comparison of transcript proportions across samples. Salmon outputs TPM values natively, along with estimated read counts, making it compatible with downstream differential expression tools like DESeq2 and edgeR.

A computational shortcut that determines transcript compatibility by checking whether a read's k-mers match a transcript's sequence, without computing a full nucleotide-level alignment. This eliminates the most computationally expensive step in traditional RNA-seq analysis. Salmon's quasi-mapping approach extends pseudoalignment by also identifying the position and orientation of matches, enabling more accurate likelihood estimation.

Systematic non-biological variation introduced by differences in sample processing, reagent lots, or sequencing platforms. These technical artifacts can confound quantification tools if not modeled. Salmon's sample-specific bias models explicitly account for positional biases, GC-content biases, and sequence-specific biases that vary between batches, reducing the need for post-hoc correction methods like ComBat-Seq.

An iterative statistical algorithm used to estimate transcript abundance when reads map to multiple isoforms. The E-step estimates the probability that each read originated from each compatible transcript, while the M-step updates abundance estimates based on these probabilities. Salmon accelerates this process with an online stochastic EM variant that processes reads in mini-batches, converging faster than traditional batch EM.

High-throughput sequencing of cDNA to profile the complete transcriptome of a biological sample. RNA-seq reveals both gene expression levels and transcript structure, including alternative splicing events. Salmon processes raw RNA-seq reads directly, bypassing the alignment step that typically consumes 70-90% of analysis time. It supports both single-end and paired-end reads from Illumina and other platforms.