Trunk sampling for increased diversity

[rui-yin]

Hi Boltz team,

Is it possible to increase diversity in boltz-1 predictions by doing multiple trunk sampling iterations (each time with a different seed), like what AlphaFold3 reported in their paper for 1,000 seed runs? From what I understand, Boltz-1 currently precomputes the trunk outputs once per complex and do multiple diffusion samples.

Thank you so much for your work on Boltz!

Best, Rui

[gcorso]

Hi Rui, yes it is possible but not currently supported on the repo. We will likely eventually get around to implementing it but if someone wants to contribute this as a PR we would be happy to merge! It requires resampling the MSA every X number of samples.

[rui-yin]

Thank you Gabriele for the prompt response!

To resample MSA, would exposing this this argument and setting it to True do the trick?

[PabloLemosP]

@rui-yin did you ever figure this out?

[GravityPHY]

Following up to check if anyone works on this :)