Orthogonal Evaluations Enable More Robust Predictions of Protein-Ligand Interactions

Joseph G. Wakim, Irene Kim, Adam T. Zemla

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: tiny / short paper (2-4 pages excluding references; extended abstract format)

Keywords: Protein-Ligand Interactions, Computer-Aided Drug Discovery, Structural Models, Physicochemistry, Protein Function, Embeddings

TL;DR: We combine three biologically linked, but independently derived, protein representations to predict and evaluate protein-ligand interactions.

Abstract: Computational models can predict protein‑ligand interactions (PLIs) at scales that far surpass experimental validation, which makes reliable confidence estimation critical. Existing approaches use protein structure and function as complementary, independently derived comparators for predicting and evaluating PLIs. However, function‑based evaluations perform poorly for promiscuous ligands, which target proteins with diverse functions. Accordingly, confidence estimation for modeled PLIs involving promiscuous ligands remains an open challenge. To address this gap, we introduce a novel physicochemical representation as an additional comparator for evaluating PLIs. Our representation encodes binding-pocket-specific features along a protein's surface, which influence its affinity for a ligand. In preliminary experiments on PLIs involving promiscuous ligands, we find that incorporating these features yields more robust confidence estimates compared to using structure and function alone. These results suggest that physicochemical representations capture meaningful biological signals for prioritizing high-quality drug leads, motivating a multimodal evaluation framework for drug discovery.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 70