Inferential replicates

Transcript abundance estimates carry inferential uncertainty: reads that map ambiguously to several transcripts can be apportioned in more than one way. salmon can quantify this uncertainty by producing inferential replicates — repeated abundance estimates that downstream tools (fishpond/swish) use to separate biological signal from inferential noise.

Bootstrap replicates

salmon quant -i salmon_index -l A \
  -1 r1.fq.gz -2 r2.fq.gz -p 16 \
  --numBootstraps 100 \
  -o out

Gibbs replicates

Gibbs sampling draws posterior samples from the abundance model. It is mutually exclusive with bootstrapping.

salmon quant -i salmon_index -l A \
  -1 r1.fq.gz -2 r2.fq.gz -p 16 \
  --numGibbsSamples 100 \
  --thinningFactor 16 \
  -o out

Output

Replicates are written to aux_info/bootstrap/:

• names.tsv.gz — the transcript names, in index order.
• bootstraps.gz — the replicate abundances as raw little-endian f64s (n_replicates × num_transcripts, contiguous).

aux_info/meta_info.json records num_bootstraps and samp_type ("bootstrap" or "gibbs"). This is the same format C++ salmon used, so it loads directly in the standard R tools. See the output format specification for the exact byte layout.

Loading in R

library(tximport)
library(fishpond)

# txi$infReps holds the inferential replicate matrices
txi <- tximport("out/quant.sf", type = "salmon", txOut = TRUE)

tximport reads the replicates automatically when they are present; fishpond and swish then use them for differential-expression testing with uncertainty.