Closed jadolfbr closed 1 year ago
It might depend on the difficulty of the backbone, but I would design 8 sequences per backbone and predict all of them using AlphaFold and then depending on the true LDDTs (or predicted LDDTs) design more sequences if all of them fail to be predicted by AlphaFold confidently.
This is extremely helpful. Thank you Justas! Is this what was done for the designs in your paper, basically the number would vary based on difficulty/fwd folded results? Would you always keep a temp of .1?
On Mon, Oct 17, 2022 at 12:01 AM Justas Dauparas @.***> wrote:
It might depend on the difficulty of the backbone, but I would design 8 sequences per backbone and predict all of them using AlphaFold and then depending on the true LDDTs (or predicted LDDTs) design more sequences if all of them fail to be predicted by AlphaFold confidently.
— Reply to this email directly, view it on GitHub https://github.com/dauparas/ProteinMPNN/issues/24#issuecomment-1280246341, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZDHRHTWVZ5R3Y6C23GHVLWDTFQTANCNFSM6AAAAAAREVGW6A . You are receiving this because you authored the thread.Message ID: @.***>
I think for most of the designs in the paper only one or couple of sequences were generated because it was easy to get AlphaFold to repredict those structures with those sequences. Yes, a temp of 0.1
What would you recommend for the number of sequences to create for design? What have you used for production runs? In the examples, these are usually one or two, but what is generally recommended?