Explore how ProteinEBM does at recovering alternate states

[etherealsunshine]

https://github.com/jproney/ProteinEBM

[marcuscollins]

@etherealsunshine can you provide some more information/context about this?

[etherealsunshine]

yeah sure, Unlike the current models being used, ProteinEBM has an explicit energy function rather than an implicit diffusion prior, which would give us ensemble reranking comes for free from the energy scalar rather than other confidence metrics.

It might be interesting to see how an EBM that scores conformations by learned energy rather than sampling frequency might produce less biased ensembles. The energy function could also serve as a LossFunction for guiding other predictors.

ProteinEBM does currently work at backbone resolution only, so sidechain-only altlocs wouldn't be evaluable(yet). But the backbone altloc segments should be a fair comparison. I personally was interested in seeing if an all-atom extension helps with recovering those conformations.