salmon-model: statistical models used during quantification.

What these models are for

Sequencing is not a uniform sampler. A fragment's chance of being observed depends on how long it is, what sequence sits at its ends, how GC-rich it is, and where in the transcript it starts. Left uncorrected, those effects are read as differences in abundance, which they are not. Each model here measures one such effect from the data itself and hands the quantifier a correction factor:

• [fld] — the fragment-length distribution, and automatic library-type detection ([libdetect]);
• [seqbias] — sequence-specific bias, from primer/ligation preferences at fragment ends;
• [gcbias] — fragment GC bias, largely from PCR amplification efficiency;
• [posbias] — positional bias along the transcript body, e.g. 3' pileup from degraded RNA;
• [bias] — combines them into a corrected effective length per transcript.

Every model is observed vs expected: count what was actually seen, count what would have been seen under no bias, and take the ratio. That framing is why the dump files in [dumps] always come in obs/exp pairs.