Bioactivity Deep Learning for Complex Structure-Free Compound-Protein Interaction Prediction

Abstract

Protein–ligand binding affinity assessment plays a pivotal role in virtual drug screening, yet conventional data-driven approaches rely heavily on limited protein–ligand crystal structures. Structure-free compound-protein interaction (CPI) methods have emerged as competitive alternatives, leveraging extensive bioactivity data to serve as more robust scoring functions. However, these methods often overlook two critical challenges that affect data efficiency and modeling accuracy: the heterogeneity of bioactivity data due to differences in bioassay measurements and the presence of activity cliffs (ACs)─small chemical modifications that lead to significant changes in bioactivity, which have not been thoroughly investigated in CPI modeling. To address these challenges, we present CPI2M, a large-scale CPI benchmark data set containing approximately 2 million bioactivity data points across four activity types (K_i, K_d, EC₅₀, and IC₅₀) with AC annotations. Moreover, we developed GGAP-CPI, a complex structure-free deep learning model trained by integrated bioactivity learning and designed to mitigate the impact of ACs on CPI prediction through advanced protein representation modeling. Our comprehensive evaluation demonstrates that GGAP-CPI outperforms 12 target-specific and 7 general CPI baselines across 4 scenarios (general CPI prediction, rare protein prediction, transfer learning, and virtual screening) on 7 benchmarks (CPI2M, MoleculeACE, CASF-2016, MerckFEP, DUD-E, DEKOIS-v2, and LIT-PCBA). Furthermore, GGAP-CPI is able to not only deliver stable bioactivity predictions but also measure prediction uncertainty and enrich binding pocket residues and interactions, underscoring its applicability to real-world bioactivity assessments and virtual drug screening.