Skip to content

Strandedness and STAR 2 pass

Ali Altintas edited this page Aug 27, 2026 · 2 revisions

Strandedness and STAR 2-pass

How strandedness is resolved

  1. Per-sample value from the strandedness column, if filled in; otherwise auto.
  2. Only auto samples take their value from RSeQC (samtools sort/index -> GTF converted to BED12 -> infer_experiment.py -> determine_strandedness.py). The forward- or reverse-strand fraction must reach strandedness.min_fraction to call forward/reverse; otherwise the library is called no (unstranded).
  3. For tetranscripts_diffexp, all samples in a contrast must resolve to the same strandedness value or the run fails with a clear error.

TEcount always receives raw, unsorted STAR output, per the TEtranscripts authors' recommendation.

The strandedness check

Strandedness is the most damaging silent error in the pipeline. Getting it wrong never fails a job — it just makes TEcount count the wrong strand, roughly halving gene counts and inverting the antisense signal, and every number downstream inherits that. It is also easy to get wrong: kit documentation often says "stranded" without saying which direction, and forward/reverse mean opposite things in different vocabularies.

So by default (strandedness.check_provided: true) RSeQC runs on every sample, including those that declared a value, and the MultiQC report gains a Strandedness check table comparing the two — modelled on nf-core/rnaseq's check:

Column Meaning
Declared the strandedness column (auto when left blank)
Inferred what infer_experiment.py says the reads look like
Used the --stranded value actually passed to TEcount/TElocal
Status OK, auto-detected, or MISMATCH (flagged red)
Forward / Reverse / Failed frac the RSeQC fractions behind the call

A mismatch is reported, never applied. An explicit sample sheet value always wins for quantification; the check tells you to go look rather than quietly changing your analysis. If the inference is right, the run needs repeating with a corrected samples.csv.

When reading a mismatch, check the Failed fraction first: a low failed fraction with one strand clearly dominant is strong evidence against the sample sheet. A high failed fraction, or two similar fractions, means RSeQC could not tell — the pipeline calls that unstranded because nothing else can be justified, which is not the same as being confidently unstranded, and the section says so when it happens.

Because RSeQC now runs on every sample, the report's built-in RSeQC → Infer experiment plot also covers the whole cohort instead of only the auto samples.

Set strandedness.check_provided: false to run RSeQC only where it is actually needed (the auto samples), skipping the extra jobs on a large cohort. The check table still renders, covering just those samples.

STAR 2-pass mapping

With star.two_pass: none, STAR alignment is guided only by the junctions annotated in ref.gtf -- a read spanning a real but unannotated splice junction can end up soft-clipped or reported as a spurious chimeric split alignment instead of a clean spliced alignment. star.two_pass trades extra STAR compute for more sensitive novel-junction detection (STAR manual section 9):

cohort is the default for now, while this feature is being evaluated -- set star.two_pass: none to opt out and get the original, single-pass-only DAG.

  • per_sample -- STAR's own --twopassMode Basic. Each sample independently discovers its own novel junctions in a first pass, inserts them into its own genome index, then re-aligns in a second pass. No other rules change; roughly doubles star_align's runtime per sample.
  • cohort -- novel junctions are pooled across every sample before any sample's final alignment (the STAR manual's "multi-sample 2-pass" recipe). Adds two stages: star_align_pass1 runs a throwaway, BAM-less alignment per sample (--outSAMtype None) purely to collect results/star_pass1/{sample}_SJ.out.tab; star_merge_junctions pools every sample's novel (non-annotated) junctions, summing uniquely-mapping read support across samples and keeping only those reaching star.two_pass_min_unique_reads (default 3) into results/star_pass1/merged_SJ.out.tab. Every sample's final star_align then passes that file to STAR via --sjdbFileChrStartEnd. A junction too weakly supported in any single sample to be caught by per_sample mode can still survive here if it recurs (even weakly) across the cohort -- the tradeoff is a DAG sync point (no sample's final alignment starts until every sample's pass-1 has finished) and roughly 2x STAR compute plus the pass-1 jobs themselves.

Both modes only affect alignment sensitivity to unannotated splicing; they do not produce a transcript/exon annotation (unlike e.g. StringTie), so exons.bed/genes.bed for the chimera screen still come entirely from ref.gtf either way.

Clone this wiki locally