A Hybrid Physics-Deep Learning Framework for Combinatorial De Novo Design of Small-Molecule Binding Proteins

Abstract

Engineering small-molecule binding proteins de novo remains a significant challenge as even advanced generative models struggle to model the atom-level details of protein-ligand interactions with sufficient accuracy. Higher experimental success rates have resulted from methods that explicitly scaffold predefined binding interactions into helical bundles. Here we introduce a scaffolding strategy that generalizes to alpha-beta architectures. By screening thousands of combinatorially assembled protein-ligand interactions against diverse de novo backbones with finely varied pocket geometries, the protocol allows for high-fidelity accommodation of target interaction geometries. Our protocol then integrates physics-based and deep learning methods for optimization of interfacial interactions and sequence-structure compatibility, considerably improving in silico design metrics. Applying this method to two chemically similar steroids achieved a notable experimental success rate (4/26 designs bind their targets), and NMR structures of two designs are in good agreement with design models. Our generalizable, atomically precise approach offers a robust framework for small-molecule binder design, effectively eliminating the need for high-throughput screening.