UC San Diego

SurGBSA combines molecular dynamics pretraining with fine-tuning on snapshot-level MMGBSA energies, enabled by a dataset of short MD simulations of docking poses. The MMGBSA fine-tuned models adapt to binding affinity and decoy pose ranking tasks. Protein–ligand interactions are inherently dynamic, yet most structure-based drug-design methods rely on static structural representations. Although atomistic molecular dynamics (MD) simulations offer a means of sampling these dynamics, datasets and learning frameworks that incorporate MD-derived energetic supervision remain limited. Here, we introduce Surrogate MMGBSA (SurGBSA), a dataset and application framework for learning molecular representations from MD simulations using frame-level MMGBSA estimates as supervision. We generated 10-ns MD trajectories for crystal structures and multiple docking poses from CASF-2016, a representative subset of PDBbind, producing approximately 1.37 million conformational snapshots paired with MMGBSA labels. The resulting resource includes ligand poses, MD trajectories, frame-level MMGBSA estimates, dataset splits, and model checkpoints for developing and evaluating MD-informed protein–ligand models. Using existing graph-based architectures, including a pretrained EGMN model, we performed k -fold cross-validation to identify suitable model, hyperparameter, and initialization configurations for learning from the resulting trajectories. We then evaluated the selected configuration using sequence-identity-based splits for MMGBSA prediction and downstream fine-tuning tasks, including binding-affinity prediction and docking-pose ranking. The results demonstrate the feasibility of using frame-level MMGBSA supervision to learn representations from MD trajectories and provide evidence that the resulting models capture information beyond a single static protein–ligand structure. However, performance varies across sequence-similarity regimes and downstream tasks, highlighting the need to balance structural coverage with dense conformational sampling. Model checkpoints, code, trajectories, MMGBSA annotations, and dataset splits are made publicly available to support further development of MD-informed molecular representation-learning methods.