Genome-alignment quantification

Most RNA-seq pipelines align reads to the genome (STAR, HISAT2, minimap2), producing spliced alignments. salmon’s usual alignment mode (-a) expects a BAM aligned to the transcriptome, so using it would mean a second, transcriptome- centric alignment. Genome-alignment mode removes that step: give salmon a genome-aligned, name-grouped BAM plus a GTF/GFF annotation and it projects each spliced alignment into transcriptome coordinates — using bramble — then quantifies exactly as it would from a transcriptomic BAM.

salmon quant -a genome.bam --annotation anno.gtf -l A -p 16 -o out

The presence of --annotation is what switches the -a branch into genome-projection mode; without it, -a quantifies a transcriptomic BAM as before.

Requirements

• A name-grouped (query-sorted) BAM. salmon needs a read’s records adjacent, so collate first: samtools collate (or samtools sort -n), or aligner output in read order (e.g. STAR --outSAMtype BAM Unsorted). A coordinate-sorted BAM is rejected.
• An annotation (--annotation <gtf|gff>) whose transcript models match the genome the reads were aligned to. Transcripts absent from the annotation (e.g. ALT-haplotype/scaffold transcripts not in a primary-assembly GTF) cannot be quantified — a real, inherent limit of genome-over-primary quantification, not a projection artifact.

Options

Option	Description
--annotation <gtf|gff>	Required for genome mode. Transcript models used to build the genome→transcriptome map.
--genome <fasta>	Genome FASTA. Optional; supply it to enable bias correction — transcript sequences are reconstructed from exon slices so --seqBias/--gcBias/--posBias work.
--juncMissDiscount <f>	Penalty for a spliced read whose junction is not supported by the annotation (bramble junc_miss_discount; default 1.0 = no penalty).

Short reads

Genome projection targets short-read (Illumina) data. For long reads (PacBio/ONT), use oarfish, COMBINE-lab’s dedicated long-read quantifier.

How it works

bramble builds a genome→transcriptome index from the annotation and the BAM’s @SQ reference names, then projects each fragment’s genomic alignment onto every compatible transcript. salmon turns each projected placement into a RAD record (transcript id, transcript-relative position, fragment length, orientation, and a per-placement score derived from bramble’s similarity), then hands the RAD to the same deterministic quantifier used everywhere else. As a result:

• it is inherently deterministic — byte-identical across thread counts, like the rest of RAD-based quantification;
• it composes with the whole feature set for free — bias correction (with --genome), --numBootstraps/--numGibbsSamples, -g/--geneMap, and -l A library-type inference all run through the shared tail;
• there is no alignment error model — bramble exposes no projected CIGAR, so the projected similarity is the placement’s quality signal (genome mode is implicitly --noErrorModel).

Accuracy

On simulated data with known truth, genome-projected quantification tracks direct transcriptomic quantification closely, and matches a reference genome→ transcriptome projection (e.g. STAR’s own --quantMode TranscriptomeSAM). The residual gap versus read-based transcriptome quantification is inherent to genome alignment: a transcriptomic aligner sees a read against an entire sequence-similar transcript family at once, whereas a genome aligner commits the read to a single locus, so paralogous/retained-intron isoforms are assigned differently. This gap is shared by any genome-projection tool and is not specific to bramble.

Note

Genome mode reuses alignment mode’s inference and output, so quant.sf, quant.genes.sf (with -g), and the inferential-replicate outputs are identical in format to every other mode.