search

facebookresearch / esm

Evolutionary Scale Modeling (esm): Pretrained language models for proteins
MIT License
3.26k stars 643 forks source link

Question about confidence determination #426

Closed alpha-omega-labs closed 1 year ago

alpha-omega-labs commented 1 year ago

Hello, Thank you very much for ESM Fold! During prediction here is a sample output:

Predicted structure for X with length 337, pLDDT 84.5, pTM 0.687 in 9.0s. 1 / 2 completed

Question: How exactly pTM counted in case of ESM Fold prediction? With what structures/fragments alignments are made or other way? Thank you!

tomsercu commented 1 year ago

pTM captures the model's confidence, it stands for "predicted TM-score". It is well-calibrated with actual TM-score to target structure which is presented during supervised training, but is a metric that does not need a target itself.

alpha-omega-labs commented 1 year ago

Thank you for a fast response. Is there a way to measure (with ESM Fold current presented models) pTM for 3rd party .pdb? Roughly compare other party predicted structure with the same sequence predicted by ESM Fold.

tomsercu commented 1 year ago

if you have an existing (sequence, structure) pair from your .pdb file, you can give the sequence to ESMFold and it will give you a structure prediction and pTM score. With the predicted structure, you can use TMalign to compute TM-score to the reference structure from the original pdb file.