I was perusing the changes that bumped rnaseq to use Salmon 2.*. According to the docs, in Salmon 2.*, the "classic error-model" is opt in and turned off by default. In alignment mode, the default behaviour now reads off alignment scores with from the

AS

tag and falls back to a uniform error model. According to this PR, the fallback is faster but performs worse github.com/COMBINE-lab/salmon/issues/1140. Performance can be rescued by opting-in into

--errorModel

. I noticed today that the STAR-Salmon route's transcriptome aligned bam records input to Salmon does not have AS tags, and

--errorModel

is not provided by default in the STAR-Salmon route. Has anyone here thought about or has taken a look at this quirk of the Salmon 1->2 migration?

I was blissfully unaware and, after you posted the link, tried to read through the concerns and benchmarks in the GitHub issue you linked. Frankly, it reads to me as two LLM agents communicating with each other: > Formal ask + implementation plan, so we don’t collide. All four PRs against

develop

, independent of each other, land in any order. Pointers are against

develop @ c498cbf0

. Unfortunately I cannot wrap my head around information presented in the “_What breaks silently when the default flips_” gibberish and I reckon I am not alone. If there is anything Rob wants us developers and users to know, I hope he still finds time to communicate it differently.

Thank you... I see. I can synthesize and be more affirmative. Based on my understanding, I believe that there the STAR-salmon path with Salmon 2, as implemented in rnaseq, will produce abundance estimates that are less accurate than the implementation with Salmon 1. This is because the "traditional" error modeling is now off by default and is off in the STAR-salmon path because STAR does not output

AS

tagged records for txome aligned records, and salmon 2 falls back to a less accurate uniform error model. My concern is that this is an unwanted regression.

Thank you for your attempt of being more affirmative. I understood your concern well, I just failed to quickly grasp from the issue how big the difference between the traditional and the new standard uniform model is. Since I am not a tool developer myself, I never bothered terribly much about the intricacies of the quantification algorithms. In my understanding tools like RSEM / Salmon / Kallisto (re)calculate assignment probabilities across isoforms based on their own sophisticated error models considering the nM tag, sequence GC content and a bunch of other magic sauce. What I can tell for sure is, that the

AS

tag never was part of the

Aligned.toTranscriptome.out.bam

files. The Migration README does not mention a respective caveat, so I do not think that any unwanted regression is to blame on the way how the pipeline applies the tool. The SAM attributes for transcriptomic alignments are hardcoded anyway, so unless one ports them from the genomic alignment they will not be available to Salmon. I think, this is a fair concern of yours, but something you should discuss with Rob first. We are certainly happy to fix the module / pipeline if needed.

Sounds good. I'll connect with Rob!