ABSTRACT Accurately predicting drug–target affinity (DTA) is crucial for accelerating virtual screening and guiding lead optimization in drug discovery. However, current computational approaches face a critical trade‐off: interaction‐free models lack fine‐grained binding details, while interaction‐based models overlook higher‐order contextual and functional patterns. This limitation hinders both prediction performance and real‐world generalization. To overcome this, we propose MF‐Net, a unified hierarchical multiscale fusion framework that integrates sequence‐, atomic‐, and fragment‐level representations to model drug–target interactions across complementary scales. MF‐Net achieves state‐of‐the‐art performance on the PDBBind v2016 benchmark and demonstrates strong early enrichment across multiple virtual screening datasets. Additionally, ADP‐Glo assays confirm that the MF‐Net‐guided virtual screening pipeline identifies seven novel nanomolar inhibitors targeting hematopoietic progenitor kinase 1 (HPK1). Among them, one compound achieves sub‐nanomolar activity (IC50 = 0.41 nM), outperforming the positive control inhibitor Sunitinib. These results demonstrate that MF‐Net not only excels on standard benchmarks but also delivers tangible lead discovery outcomes, underscoring its practical value for structure‐based drug design.
Shuo Liu, Xiang Zhang, Haixia Feng et al.· Advancement of science· 0 citations
Targeted drug discovery is fundamentally bottlenecked by the challenge of accurately modeling complex biomolecular interactions, ranging from small-molecule ligand binding to high-order macromolecular assemblies. While traditional physics-based computational methods provide profound mechanistic insights, their clinical utility is frequently hampered by prohibitive computational costs and scalability limitations when addressing highly flexible, cross-scale systems. Conversely, the rapid emergence of pure deep learning offers unprecedented computational speed but suffers from a fundamental “black-box” nature, sometimes yielding physically improbable conformations—often referred to as “hallucinations”—that can pose challenges in real-world experimental validation. To bridge this critical translational gap, the integration of physical principles with artificial intelligence—Physics-Informed Deep Learning (PIDL)—is currently driving a fundamental transition from purely empirical approximations to rational, physically grounded design. This review constructs a strategic framework to critically evaluate these transformative advances, structured around three methodological pillars: (1) Physics-constrained optimization, which integrates thermodynamic principles and integrative experimental restraints at the output level to decode macromolecular dynamics; (2) Physics-encoded architectures, which embed appropriate SE(3) or E(3) geometric symmetries directly into neural network topologies for precise structural recognition; and (3) Physics-guided representations, which project discrete sequences into continuous physicochemical manifolds to enhance interaction prediction. By delineating how physics-based priors synergize with data-driven representation learning, this review not only synthesizes current algorithmic breakthroughs but also provides a comprehensive roadmap for generating physically plausible and thermodynamically stable therapeutics, ultimately accelerating the transition of computationally designed molecules from in silico blueprints to viable clinical candidates.
Hao-Bo Xie, Hao Wang, Xiaojun Yao et al.· The Innovation Drug Discover...· 0 citations
CoBind is presented, a multitask deep learning framework that jointly predicts RNA–compound interactions and nucleotide-level binding-site probabilities within a unified architecture and provides complementary nucleotide-level binding-site localization, supporting a site-aware view of RNA–ligand recognition under distribution shift.
Shihang Wang, Lin Wang, Wei Zhao et al.· Journal of Medicinal Chemist...· 0 citations