Scalable Ligand Pose Generation via QUBO-Guided Grid Sampling and Geometric Triplet Matching

A framework based on Quadratic Unconstrained Binary Optimization (QUBO) is presented for efficient ligand pose generation within protein binding pockets, providing a scalable approach for structure-based virtual screening. The method discretizes receptor pockets into three-dimensional grids and formulates and optimizes a QUBO model to identify energetically favorable and spatially distributed grid points before ligand placement. Each ligand is represented using a geometric contour defined by three heavy atoms, and candidate poses are generated by aligning ligand atom triplets to grid-point triplets through rigid-body transformations. The framework focuses exclusively on pose generation and produces ensembles of physically plausible candidate poses that can be evaluated by arbitrary downstream scoring functions. Benchmark evaluations were performed using a self-docking and cross-docking data set derived from the CASF-2016 benchmark. Performance was evaluated using the average root-mean-square deviation (RMSD) between predicted and experimental ligand coordinates, as well as recovery rates at RMSD thresholds of 1.6, 1.2, 0.8, and 0.4 Å. Increasing the number of generated candidate poses increases the probability that near-native poses are included in the ensemble. At the same time, optional steric-clash filtering can substantially reduce the size of the candidate-pose set by discarding geometrically incompatible configurations. Comparisons with established docking programs, including AutoDock4, AutoDock Vina, Gnina, and DiffDock, show that the proposed method achieves comparable pose-recovery performance, while maintaining a modular pose-generation strategy. These results demonstrate that QUBO-guided grid sampling combined with geometric triplet matching provides a scalable and flexible framework for ligand pose generation in modern virtual screening workflows.