Question: ways to improve conditional MSA generation

[yzhang-github-pub]

I played with evo-diff on msa conditional generation. Structures of query and generated are not as similar as I expected. Actually many times they are quite different.

Below is a comparison of 3D structures of query and generated sequence from your example jupyter notebook under "Evolutionary guided sequence generation with EvoDiff-MSA":

query=MNIFEMLRIDEGLRLKIYKDTEGYYTIGIGHLLTKSPSLNAAKSELDKAIGRNTNGVITKDEAEKLFNQDVDAAVRGILRNAKLKPVYDSLDAVRRAALINMVFQMGETGVAGFTNSLRMLQQKRWDEASVNLAKSRWYNQTPNRAKRVITTFRTGTWDAYKNL
generated=MDLRSSLVEHEGLRWKVYNNAEYVPTIGLGQIHNRPSQYWDYPVPLPEQYAEKDQISWSLETIQAVFDERYTKAKSEMVNLETIGKNFDDLPSEHTNAVTDMMFQLGTDHLSEFHKMITALKNNTYEEACREMKSSFWTRQMGNRCTRYLNDALEENYFFFNHH

Structures are from colab alphafold2 with default settings. Arrows pointing to regions where structures are not similar:

My question is how to generate sequences with more similar structures to that of query. Should input a3m include very similar sequences? Any parameters to test in generation, like number of sequences to subsample? Thanks.