Salmon is a tool for wicked-fast transcript quantification from RNA-seq data. It requires a set of target transcripts (either from a reference or de-novo assembly) to quantify. All you need to run Salmon is a FASTA file containing your reference transcripts and a (set of) FASTA/FASTQ file(s) containing your reads. Optionally, Salmon can make use of pre-computed alignments (in the form of a SAM/BAM file) to the transcripts rather than the raw reads.

The quasi-mapping-based mode of Salmon runs in two phases; indexing and quantification. The indexing step is independent of the reads, and only need to be run one for a particular set of reference transcripts. The quantification step, obviously, is specific to the set of RNA-seq reads and is thus run more frequently. For a more complete description of all available options in Salmon, see below.

The alignment-based mode of Salmon does not require indexing. Rather, you can simply provide Salmon with a FASTA file of the transcripts and a SAM/BAM file containing the alignments you wish to use for quantification.

Using Salmon

As mentioned above, there are two “modes” of operation for Salmon. The first, requires you to build an index for the transcriptome, but then subsequently processes reads directly. The second mode simply requires you to provide a FASTA file of the transcriptome and a .sam or .bam file containing a set of alignments.


Read / alignment order

Salmon, like eXpress [1], uses a streaming inference method to perform transcript-level quantification. One of the fundamental assumptions of such inference methods is that observations (i.e. reads or alignments) are made “at random”. This means, for example, that alignments should not be sorted by target or position. If your reads or alignments do not appear in a random order with respect to the target transcripts, please randomize / shuffle them before performing quantification with Salmon.


Number of Threads

The number of threads that Salmon can effectively make use of depends upon the mode in which it is being run. In alignment-based mode, the main bottleneck is in parsing and decompressing the input BAM file. We make use of the Staden IO library for SAM/BAM/CRAM I/O (CRAM is, in theory, supported, but has not been thorougly tested). This means that multiple threads can be effectively used to aid in BAM decompression. However, we find that throwing more than a few threads at file decompression does not result in increased processing speed. Thus, alignment-based Salmon will only ever allocate up to 4 threads to file decompression, with the rest being allocated to quantification. If these threads are starved, they will sleep (the quantification threads do not busy wait), but there is a point beyond which allocating more threads will not speed up alignment-based quantification. We find that allocating 8 — 12 threads results in the maximum speed, threads allocated above this limit will likely spend most of their time idle / sleeping.

For quasi-mapping-based Salmon, the story is somewhat different. Generally, performance continues to improve as more threads are made available. This is because the determiniation of the potential mapping locations of each read is, generally, the slowest step in quasi-mapping-based quantification. Since this process is trivially parallelizable (and well-parallelized within Salmon), more threads generally equates to faster quantification. However, there may still be a limit to the return on invested threads, when Salmon can begin to process fragments more quickly than they can be provided via the parser.

Providing multiple read files to Salmon

Often, a single library may be split into multiple FASTA/Q files. Also, sometimes one may wish to quantify multiple replicates or samples together, treating them as if they are one library. Salmon allows the user to provide a space-separated list of read files to all of it’s options that expect input files (i.e. -r, -1, -2). When the input is paired-end reads, the order of the files in the left and right lists must be the same. There are a number of ways to provide salmon with multiple read files, and treat these as a single library. For the examples below, assume we have two replicates lib_1 and lib_2. The left and right reads for lib_1 are lib_1_1.fq and lib_1_2.fq, respectively. The left and right reads for lib_2 are lib_2_1.fq and lib_2_2.fq, respectively. The following are both valid ways to input these reads to Salmon:

> salmon quant -i index -l IU -1 lib_1_1.fq lib_2_1.fq -2 lib_1_2.fq lib_2_2.fq -o out

> salmon quant -i index -l IU -1 <(cat lib_1_1.fq lib_2_1.fq) -2 <(cat lib_1_2.fq lib_2_2.fq) -o out

Similarly, both of these approaches can be adopted if the files are gzipped as well:

> salmon quant -i index -l IU -1 lib_1_1.fq.gz lib_2_1.fq.gz -2 lib_1_2.fq.gz lib_2_2.fq.gz -o out

> salmon quant -i index -l IU -1 <(gunzip -c lib_1_1.fq.gz lib_2_1.fq.gz) -2 <(gunzip -c lib_1_2.fq.gz lib_2_2.fq.gz) -o out

In each pair of commands, the first command lets Salmon natively parse the files, while the latter command creates, on-the-fly, an input stream that consists of the concatenation of both files. Both methods work, and are acceptable ways to merge the files. The latter method (i.e. process substitution) allows more complex processing to be done to the reads in the substituted process before they are passed to Salmon as input, and thus, in some situations, is more versatile.

Quasi-mapping-based mode (including lightweight alignment)

One of the novel and innovative features of Salmon is its ability to accurately quantify transcripts using quasi-mappings. Quasi-mappings are mappings of reads to transcript positions that are computed without performing a base-to-base alignment of the read to the transcript. Quasi-mapping is typically much faster to compute than traditional (or full) alignments, and can sometimes provide superior accuracy by being more robust to errors in the read or genomic variation from the reference sequence. More details about quasi-mappings, and how they are computed, can be found here.

Salmon currently supports two different methods for mapping reads to transcriptomes; (SMEM-based) lightweight-alignment and quasi-mapping. SMEM-based mapping is the original lightweight-alignment method used by Salmon, and quasi-mapping is a newer and considerably faster alternative. Both methods are currently exposed via the same quant command, but the methods require different indices so that SMEM-based mapping cannot be used with a quasi-mapping index and vice-versa.

If you want to use Salmon in quasi-mapping-based mode, then you first have to build an Salmon index for your transcriptome. Assume that transcripts.fa contains the set of transcripts you wish to quantify. First, you run the Salmon indexer:

> ./bin/salmon index -t transcripts.fa -i transcripts_index --type quasi -k 31

This will build the quasi-mapping-based index, using an auxiliary k-mer hash over k-mers of length 31. While quasi-mapping will make used of arbitrarily long matches between the query and reference, the k size selected here will act as the minimum acceptable length for a valid match. Thus, a smaller value of k may slightly improve sensitivty. We find that a k of 31 seems to work well for reads of 75bp or longer, but you might consider a smaller k if you plan to deal with shorter reads. Note that there is also a k parameter that can be passed to the quant command. However, this has no effect if one is using a quasi-mapping index, as the k value provided during the index building phase overrides any k provided during quantification in this case. Since quasi-mapping is the default index type in Salmon, you can actually leave off the --type quasi parameter when building the index. To build a lightweight-alignment (FMD-based) index instead, one would use the following command:

> ./bin/salmon index -t transcripts.fa -i transcripts_index --type fmd

Note that no value of k is given here. However, the SMEM-based mapping index makes use of a parameter k that is passed in during the quant phase (the default value is 19).

Then, you can quantify any set of reads (say, paired-end reads in files reads1.fq and reads2.fq) directly against this index using the Salmon quant command as follows:

> ./bin/salmon quant -i transcripts_index -l <LIBTYPE> -1 reads1.fq -2 reads2.fq -o transcripts_quant

If you are using single-end reads, then you pass them to Salmon with the -r flag like:

> ./bin/salmon quant -i transcripts_index -l <LIBTYPE> -r reads.fq -o transcripts_quant

This same quant command will work with either index (quasi-mapping or SMEM-based), and Salmon will automatically determine the type of index being read and perform the appropriate lightweight mapping accordingly.


Order of command-line parameters

The library type -l should be specified on the command line before the read files (i.e. the parameters to -1 and -2, or -r). This is because the contents of the library type flag is used to determine how the reads should be interpreted.

You can, of course, pass a number of options to control things such as the number of threads used or the different cutoffs used for counting reads. Just as with the alignment-based mode, after Salmon has finished running, there will be a directory called salmon_quant, that contains a file called quant.sf containing the quantification results.

Alignment-based mode

Say that you’ve prepared your alignments using your favorite aligner and the results are in the file aln.bam, and assume that the sequence of the transcriptome you want to quantify is in the file transcripts.fa. You would run Salmon as follows:

> ./bin/salmon quant -t transcripts.fa -l <LIBTYPE> -a aln.bam -o salmon_quant

The <LIBTYPE> parameter is described below and is shared between both modes of Salmon. After Salmon has finished running, there will be a directory called salmon_quant, that contains a file called quant.sf. This contains the quantification results for the run, and the columns it contains are similar to those of Sailfish (and self-explanatory where they differ).

For the full set of options that can be passed to Salmon in its alignment-based mode, and a description of each, run salmon quant --help-alignment.


Genomic vs. Transcriptomic alignments

Salmon expects that the alignment files provided are with respect to the transcripts given in the corresponding fasta file. That is, Salmon expects that the reads have been aligned directly to the transcriptome (like RSEM, eXpress, etc.) rather than to the genome (as does, e.g. Cufflinks). If you have reads that have already been aligned to the genome, there are currently 3 options for converting them for use with Salmon. First, you could convert the SAM/BAM file to a FAST{A/Q} file and then use the lightweight-alignment-based mode of Salmon described below. Second, given the converted FASTA{A/Q} file, you could re-align these converted reads directly to the transcripts with your favorite aligner and run Salmon in alignment-based mode as described above. Third, you could use a tool like sam-xlate to try and convert the genome-coordinate BAM files directly into transcript coordinates. This avoids the necessity of having to re-map the reads. However, we have very limited experience with this tool so far.

Multiple alignment files

If your alignments for the sample you want to quantify appear in multiple .bam/.sam files, then you can simply provide the Salmon -a parameter with a (space-separated) list of these files. Salmon will automatically read through these one after the other quantifying transcripts using the alignments contained therein. However, it is currently the case that these separate files must (1) all be of the same library type and (2) all be aligned with respect to the same reference (i.e. the @SQ records in the header sections must be identical).

Description of important options

Salmon exposes a number of useful optional command-line parameters to the user. The particularly important ones are explained here, but you can always run salmon quant -h to see them all.

-p / --numThreads

The number of threads that will be used for quasi-mapping, quantification, and bootstrapping / posterior sampling (if enabled). Salmon is designed to work well with many threads, so, if you have a sufficient number of processors, larger values here can speed up the run substantially.


Default number of threads

The default behavior is for Salmon to probe the number of available hardware threads and

to use this number. Thus, if you want to use fewer threads (e.g., if you are running multiple instances of Salmon simultaneously), you will likely want to set this option explicitly in accordance with the desired per-process resource usage.


If Salmon is passed the --dumpEq option, it will write a file in the auxiliary directory, called eq_classes.txt that contains the equivalence classes and corresponding counts that were computed during quasi-mapping. The file has a format described in Equivalence class file.


This parameter governs the a priori probability that a fragment mapping or aligning to the reference in a manner incompatible with the prescribed library type is nonetheless the correct mapping. Note that Salmon sets this value, by default, to a small but non-zero probability. This means that if an incompatible mapping is the only mapping for a fragment, Salmon will still assign this fragment to the transcript. This default behavior is different than programs like RSEM, which assign incompatible fragments a 0 probability (i.e., incompatible mappings will be discarded). If you wish to obtain this behavior, so that only compatible mappings will be considered, you can set --incompatPrior 0.0. This will cause Salmon to only consider mappings (or alignments) that are compatible with the prescribed or inferred library type.


Note : This option is only important when running Salmon with single-end reads.

Since the empirical fragment length distribution cannot be estimated from the mappings of single-end reads, the --fldMean allows the user to set the expected mean fragment lenth of the sequencing library. This value will affect the effective length correction, and hence the estimated effective lengths of the transcripts and the TPMs. The value passed to --fldMean will be used as the mean of the assumed fragment length distribution (which is modeled as a truncated Gaussian with a standard deviation given by --fldSD).


Note : This option is only important when running Salmon with single-end reads.

Since the empirical fragment length distribution cannot be estimated from the mappings of single-end reads, the --fldSD allows the user to set the expected standard deviation of the fragment lenth distribution of the sequencing library. This value will affect the effective length correction, and hence the estimated effective lengths of the transcripts and the TPMs. The value passed to --fldSD will be used as the standard deviation of the assumed fragment length distribution (which is modeled as a truncated Gaussan with a mean given by --fldMean).


Use the variational Bayesian EM algorithm rather than the “standard” EM algorithm to optimize abundance estimates. The details of the VBEM algorithm can be found in [3]. While both the standard EM and the VBEM produce accurate abundance estimates, there are some trade-offs between the approaches. The EM algorithm tends to produce sparser estimates (i.e. more transcripts estimated to have 0 abundance), while the VBEM, in part due to the prior, tends to estimate non-zero abundance for more transcripts. Conversely, the prior used in the VBEM tends to have a regularizing effect, especially for low abundance transcripts, that leads to more consistent estimates of abundance at low expression levels. We are currently working to analyze and understand all the tradeoffs between these different optimization approaches. Also, the VBEM tends to converge after fewer iterations, so it may result in a shorter runtime; especially if you are computing many bootstrap samples.

The default prior used in the VB optimization is a per-nucleotide prior of 0.001 per nucleotide. This means that a transcript of length 1000 will have a prior count of 1 fragment, while a transcript of length 500 will have a prior count of 0.5 fragments, etc. This behavior can be modified in two ways. First, the prior itself can be modified via Salmon’s --vbPrior option. The argument to this option is the value you wish to place as the per-nucleotide prior. Additonally, you can modify the behavior to use a per-transcript rather than a per-nucleotide prior by passing the flag --perTranscriptPrior to Salmon. In this case, whatever value is set by --vbPrior will be used as the transcript-level prior, so that the prior count is no longer dependent on the transcript length. However, the default behavior of a per-nucleotide prior is recommended when using VB optimization.


Salmon has the ability to optionally compute bootstrapped abundance estimates. This is done by resampling (with replacement) from the counts assigned to the fragment equivalence classes, and then re-running the optimization procedure, either the EM or VBEM, for each such sample. The values of these different bootstraps allows us to assess technical variance in the main abundance estimates we produce. Such estimates can be useful for downstream (e.g. differential expression) tools that can make use of such uncertainty estimates. This option takes a positive integer that dictates the number of bootstrap samples to compute. The more samples computed, the better the estimates of varaiance, but the more computation (and time) required.


Just as with the bootstrap procedure above, this option produces samples that allow us to estimate the variance in abundance estimates. However, in this case the samples are generated using posterior Gibbs sampling over the fragment equivalence classes rather than bootstrapping. We are currently analyzing these different approaches to assess the potential trade-offs in time / accuracy. The --numBootstraps and --numGibbsSamples options are mutually exclusive (i.e. in a given run, you must set at most one of these options to a positive integer.)


Passing the --seqBias flag to Salmon will enable it to learn and correct for sequence-specific biases in the input data. Specifically, this model will attempt to correct for random hexamer priming bias, which results in the preferential sequencing of fragments starting with certain nucleotide motifs. By default, Salmon learns the sequence-specific bias parameters using 1,000,000 reads from the beginning of the input. If you wish to change the number of samples from which the model is learned, you can use the --numBiasSamples parameter. Salmon uses a variable-length Markov Model (VLMM) to model the sequence specific biases at both the 5’ and 3’ end of sequenced fragments. This methodology generally follows that of Roberts et al. [2], though some details of the VLMM differ.

Note: This sequence-specific bias model is substantially different from the bias-correction methodology that was used in Salmon versions prior to 0.6.0. This model specifically accounts for sequence-specific bias, and should not be prone to the over-fitting problem that was sometimes observed using the previous bias-correction methodology.


Passing the --gcBias flag to Salmon will enable it to learn and correct for fragment-level GC biases in the input data. Specifically, this model will attempt to correct for biases in how likely a sequence is to be observed based on its internal GC content. This bias is distinct from the primer biases learned with the --seqBias option. Though these biases are distinct, they are not completely independent. When both --seqBias and --gcBias are enabled, Salmon will learn a conditional fragment-GC bias model. By default, Salmon will learn 3 different fragment-GC bias models based on the GC content of the fragment start and end contexts, though this number of conditional models can be changed with the (hidden) option --conditionalGCBins. Likewise, the number of distinct fragment GC bins used to model the GC bias can be changed with the (hidden) option --numGCBins.

Note : In order to speed up the evaluation of the GC content of arbitrary fragments, Salmon pre-computes and stores the cumulative GC count for each transcript. This requires an extra 4-bytes per nucleotide. While this extra memory usage should normally be minor, it can nonetheless be controlled with the --gcSizeSamp option. This option takes a positive integer argument i, such that Salmon stores the values of the cumulative GC count only at every ith nucleotide. The cumulative GC count at values between the sampled positions are recomputed on-the-fly when necessary. Using this option will reduce the memory required to store the GC information by a factor of i, but will slow down the computation of GC-fragment content by a factor of i/2. Typically, the --gcSizeSamp can be left at its default value of 1, but larger values can be chosen if necessary.


Passing the --posBias flag to Salmon will enable modeling of a position-specific fragment start distribution. This is meant to model non-uniform coverage biases that are sometimes present in RNA-seq data (e.g. 5’ or 3’ positional bias). Currently, a small and fixed number of models are learned for different length classes of transcripts, as is done in Roberts et al. [2]. Note: The positional bias model is relatively new, and is still undergoing testing. It replaces the previous –useFSPD option, which is now deprecated. This feature should be considered as experimental in the current release.


When evaluating the bias models (the GC-fragment model specifically), Salmon must consider the probability of generating a fragment of every possible length (with a non-trivial probability) from every position on every transcript. This results in a process that is quadratic in the length of the transcriptome — though each evaluation itself is efficient and the process is highly parallelized.

It is possible to speed this process up by a multiplicative factor by considering only every ith fragment length, and interploating the intermediate results. The --biasSpeedSamp option allows the user to set this sampling factor. Larger values speed up effective length correction, but may decrease the fidelity of bias modeling. However, reasonably small values (e.g. 10 or less) should have only a minor effect on the computed effective lengths, and can considerably speed up effective length correction on large transcriptomes.


Passing the --writeUnmappedNames flag to Salmon will tell Salmon to write out the names of reads (or mates in paired-end reads) that do not map to the transcriptome. When mapping paired-end reads, the entire fragment (both ends of the pair) are identified by the name of the first read (i.e. the read appearing in the _1 file). Each line of the umapped reads file contains the name of the unmapped read followed by a simple flag that designates how the read failed to map completely. For single-end reads, the only valid flag is u (unmapped). However, for paired-end reads, there are a number of different possibilities, outlined below:

u   = The entire pair was unmapped. No mappings were found for either the left or right read.
m1  = Left orphan (mappings were found for the left (i.e. first) read, but not the right).
m2  = Right orphan (mappinds were found for the right read, but not the left).
m12 = Left and right orphans. Both the left and right read mapped, but never to the same transcript.

By reading through the file of unmapped reads and selecting the appropriate sequences from the input FASTA/Q files, you can build an “unmapped” file that can then be used to investigate why these reads may not have mapped (e.g. poor quality, contamination, etc.). Currently, this process must be done independently, but future versions of Salmon may provide a script to generate this unmapped FASTA/Q file from the unmapped file and the original inputs.


Passing the --writeMappings argument to Salmon will have an effect only in mapping-based mode and only when using a quasi-index. When executed with the --writeMappings argument, Salmon will write out the mapping information that it then processes to quantify transcript abundances. The mapping information will be written in a SAM compatible format. If no options are provided to this argument, then the output will be written to stdout (so that e.g. it can be piped to samtools and directly converted into BAM format). Otherwise, this argument can optionally be provided with a filename, and the mapping information will be written to that file. Note: Because of the way that the boost options parser that we use works, and the fact that --writeMappings has an implicit argument of stdout, if you provide an explicit argument to --writeMappings, you must do so with the syntax --writeMappings=<outfile> rather than the synatx --writeMappings <outfile>. This is a due to a limitation of the parser in how the latter could be interpreted.


Compatible mappings

The mapping information is computed and written before library type compatibility checks take place, thus the mapping file will contain information about all mappings of the reads considered by Salmon, even those that may later be filtered out due to incompatibility with the library type.

What’s this LIBTYPE?

Salmon, has the user provide a description of the type of sequencing library from which the reads come, and this contains information about e.g. the relative orientation of paired end reads. As of version 0.7.0, Salmon also has the ability to automatically infer (i.e. guess) the library type based on how the first few thousand reads map to the transcriptome. To allow Salmon to automatically infer the library type, simply provide -l A or --libType A to Salmon. Even if you allow Salmon to infer the library type for you, you should still read the section below, so that you can interpret how Salmon reports the library type it discovers.


Automatic library type detection in alignment-based mode

The implementation of this feature involves opening the BAM file, peaking at the first record, and then closing it to determine if the library should be treated as single-end or paired-end. Thus, in alignment-based mode automatic library type detection will not work with an input stream. If your input is a regular file, everything should work as expected; otherwise, you should provide the library type explicitly in alignment-based mode.

Also the automatic library type detection is performed on the basis of the alignments in the file. Thus, for example, if the upstream aligner has been told to perform strand-aware mapping (i.e. to ignore potential alignments that don’t map in the expected manner), but the actual library is unstranded, automatic library type detection cannot detect this. It will attempt to detect the library type that is most consistent with the alignment that are provided.

The library type string consists of three parts: the relative orientation of the reads, the strandedness of the library, and the directionality of the reads.

The first part of the library string (relative orientation) is only provided if the library is paired-end. The possible options are:

I = inward
O = outward
M = matching

The second part of the read library string specifies whether the protocol is stranded or unstranded; the options are:

S = stranded
U = unstranded

If the protocol is unstranded, then we’re done. The final part of the library string specifies the strand from which the read originates in a strand-specific protocol — it is only provided if the library is stranded (i.e. if the library format string is of the form S). The possible values are:

F = read 1 (or single-end read) comes from the forward strand
R = read 1 (or single-end read) comes from the reverse strand

An example of some library format strings and their interpretations are:

IU (an unstranded paired-end library where the reads face each other)
SF (a stranded single-end protocol where the reads come from the forward strand)
OSR (a stranded paired-end protocol where the reads face away from each other,
     read1 comes from reverse strand and read2 comes from the forward strand)


Strand Matching

Above, when it is said that the read “comes from” a strand, we mean that the read should align with / map to that strand. For example, for libraries having the OSR protocol as described above, we expect that read1 maps to the reverse strand, and read2 maps to the forward strand.

For more details on the library type, see Fragment Library Types.


For details of Salmon’s different output files and their formats see Salmon Output File Formats.


Salmon, in quasi-mapping-based mode, can accept reads from FASTA/Q format files, or directly from gzipped FASTA/Q files (the ability to accept compressed files directly is a feature of Salmon 0.7.0 and higher). If your reads are compressed in a different format, you can still stream them directly to Salmon by using process substitution. Say in the quasi-mapping-based Salmon example above, the reads were actually in the files reads1.fa.bz2 and reads2.fa.bz2, then you’d run the following command to decompress the reads “on-the-fly”:

> ./bin/salmon quant -i transcripts_index -l <LIBTYPE> -1 <(bunzip2 -c reads1.fa.gz) -2 <(bunzip2 -c reads2.fa.bz2) -o transcripts_quant

and the bzipped files will be decompressed via separate processes and the raw reads will be fed into Salmon. Actually, you can use this same process even with gzip compressed reads (replacing bunzip2 with gunzip or pigz -d). Depending on the number of threads and the exact configuration, this may actually improve Salmon’s running time, since the reads are decompressed concurrently in a separate process when you use process substitution.

Finally, the purpose of making this software available is for people to use it and provide feedback. The paper describing this method is published in Nature Methods. If you have something useful to report or just some interesting ideas or suggestions, please contact us ( and/or If you encounter any bugs, please file a detailed bug report at the Salmon GitHub repository.


[1]Roberts, Adam, and Lior Pachter. “Streaming fragment assignment for real-time analysis of sequencing experiments.” Nature methods 10.1 (2013): 71-73.
[2](1, 2) Roberts, Adam, et al. “Improving RNA-Seq expression estimates by correcting for fragment bias.” Genome biology 12.3 (2011): 1.
[3]Patro, Rob, et al. “Salmon provides fast and bias-aware quantification of transcript expression.” Nature Methods(2017). Advanced Online Publication. doi: 10.1038/nmeth.4197