Output interpretation

[Mossy-Frog]

Hello !
Thank you for this awesome tool.
I had some questions regarding the output, especially in the RFdiffusion pipeline. I struggle to find on which criteria to filter the sample sequences generated by PMPNN based on a single structure. I have a vague idea of what the score and global score mean, but I am still confused. Which of these criterias would be the best to filter on ? Also, I tried several sequences analysis, based on the AA composition, often skewed towards certain amino acids. Is a protein with 30% of the same AA a good candidate ? How about when sometimes homopolymers of 6-10 same AA ? I checked #46 and #61 , but even with relaxation, the probem does not disappear entirely, thus the need of filtering.

What are your opinions ? I would love in both users and devs ideas on the matter !
Cheers

[roccomoretti]

There's now quite a few publications demonstrating successful design campaigns with ProteinMPNN. You should be able to look at most recent publications from the Baker lab (and others) to see how they did filtering and selection in particular cases.

Generally speaking, though, the primary filter used for design selection is ML refolding (typically AlphaFold). You're looking for designs where the ML structure prediction program predicts the same structure that was input to ProteinMPNN. If AlphaFold (etc.) predicts a structure which is substantially different from the input, or if it predicts a structure with a low confidence, then it's highly likely that the ProteinMPNN-designed sequence will not fold how you expect it to.

Past the refolding, it really depends on your particular use case. For your particular target system, there's likely experimental and performance characteristics which you're interested in. You can/should add on additional filters which look for those system-specific characteristics. -- One way to think about it is that you're looking for a "swiss cheese model" filtering approach. You're not looking for a design which is perfect by any one particular metric, but instead you're looking for a set of diverse designs which do decently well at a range of relevant metrics. Computation filtering approaches are imperfect, so loose thresholds on a bunch of different filters may be better than stringent thresholds on a single filter. And success rates are still not 100%, so you'll want to pick a number of different possibilities in the hope that at least one will be successful.

One thing to keep in mind is that generation of new designs is generally fast & cheap when compared to experimental testing. Often times it may be worth generating more designs and filtering, rather than go forward with experimental testing of "meh" designs.