salmon-align: alignment-based (BAM) quantification.
The alternative to mapping FASTQ reads: take a BAM of reads already aligned
to the transcriptome (each @SQ is a transcript) and quantify directly
from those alignments. References and their lengths come from the BAM
header, so no index or FASTA is required for basic quantification.
For each fragment (the contiguous run of records sharing a read name) the
set of transcripts it aligns to becomes an equivalence class, weighted by
alignment score (AS) via the same exp(-scoreExp·(best−score)) rule used
in the mapping path. Fragment lengths (TLEN) feed the fragment-length
distribution. The class set then flows through the shared EM
([salmon_infer]) to quant.sf. Mirrors salmon's quant -a mode (the
position-binned alignment error model is a later refinement).