Questions about steering functionality

[zoidpuff]

Hi and thanks for the wonderful work!

The recently added steering feature seems very interesting and I had some questions employing it in practice.

• Is it generally advisable to always use it when generating ensembles?
• Is computationally expensive?
• Is it preferable to relaxing unphysical structures after generation?
• Should only use it if I first see a large fraction of structures are filtered out? (i.e. start with a hundred samples with no steering and then enable it if I see large fraction of unphysical backbones)

Thank you!

[ludwigwinkler]

Hi @zoidpuff ,

• That depends on the potentials that you define towards which steering would guide the samples. Physicality potentials are generally a good recommendation, and we recovered ensembles quantitatively with the TICA plots in the bioemu benchmarks with steering. But for more complex or targeted potentials your mileage might vary.
• Sequential Monte Carlo steering is computationally almost free as it computes denoised samples from the already predicted score but is also less powerful. FKC steering is markedly more capable but requires up to 2x more compute and memory if you choose to backpropagate through the denoiser.
• It's dependent on whether your sequences are prone to unphysicality when generating with BioEmu. Using the physicality potentials should already greatly reduce the number of unphysical samples. We were able to supress unphysicality down to 1-2 samples per 1000 generated samples for structures from 30 per 1000. You can always add relaxation afterwards to be sure.
• Yes, that would be a good approach. Both SMC and FKC are sequential Monte Carlo algorithms, so you should be aware of the diversity trade-off when applying resampling too aggressively. We saw good performance by using 5 particles between which we resample if the resampling condition is triggered.