Comparative AI-predicted and experimental structure-based screening identifies DNA-PKcs-binding candidate radiosensitizers

AI-predicted and experimental structures were cross-checked for screening. • Structure-source effects on ligand enrichment were target-dependent. • Multiple DNA-PKcs structures prioritized partly distinct candidate sets. • L-365,260 and Fmoc-Tic-OH showed apparent micromolar DNA-PKcs binding. • Both compounds produced radiosensitizing phenotypes in NSCLC cells. Radioresistance remains a major challenge in non-small-cell lung cancer (NSCLC) radiotherapy, motivating the search for compounds that enhance tumor-cell responses to irradiation. Artificial intelligence (AI)-predicted structures expand access to structure-based virtual screening, but how they compare with and complement experimentally determined structures remains unclear. Here, we compared experimental, AlphaFold2, AlphaFold3 and RoseTTAFold All-Atom structures of four DNA damage-response proteins—ATM, ATR, DNA-PKcs and PARP1—using structural analyses and retrospective active-decoy enrichment. AI-predicted and experimental structures of DNA-PKcs were then used to screen 79,263 natural-product-related compounds from ZINC20. The prospective screen yielded 13 filtered candidates and showed that the three DNA-PKcs structures prioritized partially overlapping chemical sets. Docking, ADMET prediction, molecular dynamics and molecular mechanics/Poisson–Boltzmann surface area (MM/PBSA) calculations were used to rank candidates, and L-365,260 (ZINC000003797417) and Fmoc-Tic-OH (ZINC000000621966) were selected for experimental characterization. Surface plasmon resonance yielded apparent steady-state DNA-PKcs dissociation constant (KD) estimates of 11 and 24 μM for L-365,260 and Fmoc-Tic-OH, respectively. In A549 and H1299 cells, the compounds produced model-derived sensitizer enhancement ratios at 50% survival (SER50) ranging from 1.87 to 2.17. Proteomic profiling, apoptosis, cell-cycle and γH2AX analyses characterized treatment-associated cellular responses. Within this DNA-PKcs case study, the results indicate that AI-predicted and experimental structures can provide complementary views of a binding site and prioritize partially non-overlapping candidate sets. This comparative workflow offers a practical framework for cross-checking structural sources and prioritizing compounds in structure-based screening. Download: Download high-res image (212KB) Download: Download full-size image