Given small molecules locally resemble amino acid sidechains, we (Artem Gazizov, Anna Lian, Casper Goverde, Nicholas Polizzi) wondered if AF2 could be used to predict binding sites, when provided with a structure and a string of AAs. (1/10) *[looking for PhD opportunities] 🤓 - ThreadSky

sokrypton.org • 500 days ago

Given small molecules locally resemble amino acid sidechains, we (Artem Gazizov, Anna Lian, Casper Goverde, Nicholas Polizzi) wondered if AF2 could be used to predict binding sites, when provided with a structure and a string of AAs. (1/10) *[looking for PhD opportunities] 🤓

Comments

rolanddunbrack.bsky.social•500 days ago

Sometimes everything really is just a nail :-).

sokrypton.org•500 days ago

oops, somehow lost the placement of *
Artem Gazizov, the first author of this work, is looking for PhD opportunities! Please reply here or email him (email address is in preprint) if you are looking for PhD students! 🤓

sokrypton.org•500 days ago

Though we don't expect the structure module to correctly dock the individual AAs (given each AA can go to multiple locations) we figured maybe the pair representation (which encodes the attention between protein and each "bait" amino acid) may have this information (2/10).

sokrypton.org•500 days ago

We train a simple logistic regression model that goes from the pair representation to binding site prediction (3/10)

sokrypton.org•500 days ago

We find the pair representation (between the target and 20 amino acids) to give more meaninful representations compared to those extracted from ESM2 (sequence) and ESM-IF1 (structure) models (4/10).

sokrypton.org•500 days ago

Since the representations directly map to the "bait" amino acids, we can directly assess which of the AAs contributed towards the binding prediction, giving us some clues about the ligand composition. (5/10) 😎

sokrypton.org•500 days ago

We spent a long time perfecting our training/valid/test split... making sure there is no overlap based on sequence (id, pfam, interpro), structure (TMscore, ecod, cath, scop) and ligand site. (It's suprisingly easy to overfit, even with this simple model...) (6/10)

sokrypton.org•500 days ago

This model generalizes to withheld sequences/structures. (7/10)

Comments

Posting Rules

Reply