Benchmark sensitivty tradeoff of ungapped GPU

[keiran-rowell-unsw]

Many substasks related to to when the ungapped is not performant and CPU sensitivity parameters must be sued

[keiran-rowell-unsw]

Tom points of this is worth doing particular with uniref30, because it is paired, and through not just mmseqs-(gpu) but explicitly through colabsearch because there's down stream alignment expansion into a cluster which helps match species because you gather far more hits.