search

jwohlwend / boltz

Official repository for the Boltz-1 biomolecular interaction model
MIT License
1.29k stars 135 forks source link

Finetuning Boltz with no MSA. #75

Open AlexanderMath opened 1 day ago

AlexanderMath commented 1 day ago

Any interest in PR / model weights finetuned to improve single-sequence performance? (e.g. using ESM activations as done in CHAI).

Asking after trying Crambin (1EJG), saw 0.4 RMSD increase to 8-10 RMSD - curious if other see the same, scared I'm just doing something wrong.

AlexanderMath commented 1 day ago

Also, curious if anyone finetuned ESM to minimize |ESM(seq).MSA_head - MSA(seq)| - very hacky way to circumvent MSA pipeline engineering pain.

wckdouglas commented 1 day ago

Also, curious if anyone finetuned ESM to minimize |ESM(seq).MSA_head - MSA(seq)| - very hacky way to circumvent MSA pipeline engineering pain.

@AlexanderMath maybe not quite minimize |ESM(seq).MSA_head - MSA(seq)| but vcmsa uses prot-t5 embeddings to infer msa