AlphaFold3 and RoseTTAFold All-Atom structures enable radiosensitizers discovery by targeting multiple DNA damage repair proteins

Abstract

Radioresistance remains a primary obstacle in tumor radiotherapy, with no clinically approved radiosensitizers due to toxicity concerns. To identify effective and safe radiosensitizers, a natural products database containing 79,263 compounds are docked against a hybrid target library of four DNA damage response (DDR)-related proteins, comprising both experimental and artificial intelligence (AI)-predicted structures generated by AlphaFold3 and RoseTTAFold All-Atom models. Retrospectively, AI-modeled structures show comparable AUC and logAUC values to experimental structures. Prospectively, compounds screened by AI-modeled structures versus those by experimental structures exhibit limited overlap, e.g., 10% for ataxia telangiectasia mutated (ATM), 22.2% for ATM- and Rad3-related (ATR), 7.7% for DNA-dependent protein kinase catalytic subunit (DNA-PKcs), and 40% for Poly (ADP-ribose) polymerase 1 (PARP1). This highlights structural complementarity of AI-modeled structures when docking against small-scale compound libraries. Two compounds exhibiting lower binding free energy than the DNA-PKcs co-crystallized ligand were selected and validated as effective radiosensitizers in tumor cells. Proteomic analyses reveal shared DDR dysregulation but distinct repair pathway vulnerabilities behind both compounds, which activate TP53-associated apoptosis and senescence as cellular endpoints by modulating the synergistic interplay between DDR and spindle checkpoints. These findings highlight their potential as context-dependent radiosensitizers, providing novel candidates and strategies to overcome tumor radioresistance.