PROTEUS: Predicting How Post-Translational Modifications Alter Drug Binding Affinity

Bryan Cheng, Austin Jin, Jasper Zhang, Arnav Pemmaraju, Joshua Chang, Brendan Lo

Keywords: multimodal deep learning, drug-target interaction, post-translational modifications, foundational benchmark, graph neural networks, protein language models, uncertainty quantification, binding affinity prediction

TL;DR: SOTA multimodal framework fusing protein LM embeddings, 3D contact graphs, PTM annotations, and drug molecular graphs to predict how PTMs alter drug binding affinity. Establishes the first benchmark and method for this foundational task.

Abstract: Post-translational modifications (PTMs) dynamically reshape the binding landscapes of protein drug targets, yet no computational method predicts how specific PTMs alter drug binding affinity. We introduce PROTEUS, a multimodal deep learning framework that fuses protein language model embeddings, 3D protein contact graphs, PTM-site annotations, and drug molecular graphs to directly predict ΔΔG---the change in binding free energy caused by a PTM. A Normal-Inverse-Gamma evidential head provides calibrated epistemic and aleatoric uncertainty without ensemble overhead. To enable systematic study of this previously unaddressed prediction task, we construct PTM-BIND-Bench, the first benchmark linking PTM states to quantitative binding affinity changes: 831 curated protein-PTM-drug triples spanning 48 kinases, 120 inhibitors, and 5 PTM types. Under protein-level GroupKFold cross-validation, PROTEUS achieves Spearman ρ = 0.527 and concordance index = 0.715, outperforming all baselines with non-overlapping 95% CIs. On 10 experimentally measured ΔΔG values, the model attains Pearson r = 0.967 with 100% direction accuracy. Zero-shot evaluation on 17 novel drugs yields 88.9% direction accuracy, and split conformal prediction provides distribution-free 90% intervals of ±0.64 kcal/mol.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 129