Exploring Ligand Flexibility in Nucleic Acid Scaffolds Using Graph Neural Networks.

Interactions between nucleic acids and ligands are vital for gene regulation and therapy development, yet accurate modeling is hindered by shallow, flexible pockets and conformational changes. Existing docking approaches typically assume that both ligands and nucleic acids are rigid, which results in sampling pools that lack near-native conformations and makes it more challenging for scoring functions to distinguish correct poses from decoys. We introduce ZHMolLigGraph, a two-stage graph-based deep learning framework that explicitly models ligand flexibility. Phase I (Flexibility Exploration) applies iterative atomic displacements to explore ligand conformational adaptability, allowing the ligand to traverse a wide range of structural states; Phase II (Feasibility Selection) screens these candidates for geometric and interaction plausibility, retaining only physically reasonable poses rather than ranking all candidates. Across various benchmarks, ZHMolLigGraph consistently improved near-native hit rates by 10.30-30.91% compared to conventional docking algorithms, while maintaining fast computational efficiency. ZHMolLigGraph provides a practical and scalable framework for exploring ligand flexibility in nucleic acid-ligand systems. Notably, the framework is extensible and can be expanded to incorporate receptor flexibility in the future, offering a path toward more faithful modeling of RNA-ligand recognition in realistic biological contexts.