SurGBSA

Molecular Mechanics Generalized Born Surface Area (MMGBSA) is one of the workhorse endpoint methods of structure-based drug design: it estimates protein-ligand binding free energy from an explicit force field plus an implicit solvent term, and it routinely outperforms fast docking scoring functions at telling a correct binding pose from a plausible decoy. Its cost is the problem. A single-point evaluation takes seconds of CPU time per pose, putting it out of reach for the millions-of-compounds screens that motivate it. Meanwhile, most learned molecular representations are built from static crystal structures and discard the conformational ensemble that determines binding thermodynamics.

SurGBSA (SURrogate mmGBSA) attacks both gaps at once by learning a neural surrogate of the MMGBSA energy function directly from molecular dynamics trajectories. Researchers at Lawrence Livermore National Laboratory and UC San Diego ran atomistic MD for every pose in a protein-ligand benchmark, scored every resulting snapshot with physics, and trained an equivariant graph network to reproduce those scores from atom types and 3D coordinates alone. The result, posted as an arXiv preprint in September 2025, reproduces the physics-based method's pose-ranking accuracy to within half a percentage point while running roughly four orders of magnitude faster.

The approach is complementary to co-folding systems such as Boltz-2, which predict a structure and an affinity from sequence. SurGBSA assumes the poses already exist — from docking or from a co-folding model — and replaces the expensive physics-based rescoring step that follows.

#Key Features

• Physics-informed pretraining on MD ensembles: The model is trained on over 1.4 million simulation frames, each carrying its own single-point MMGBSA label, so the network sees how the scoring function responds to thermal motion.
• Equivariant graph matching encoder: An EGMN backbone consumes atom types and coordinates for two time steps and connects them with cross-graph edges, giving the model an explicit representation of how the complex moves between snapshots.
• Near-parity pose ranking: On the pose-ranking task the surrogate identifies the correct top pose 44.7% of the time against 45.1% for single-point MMGBSA, a difference of -0.4 percentage points, with a higher rank correlation to the reference energies (Spearman 0.702 vs. 0.54).
• Four orders of magnitude speedup: Inference takes 0.0003 seconds per pose on an H100 GPU versus 8.378 seconds for the physics-based calculation, a roughly 27,927x reduction.
• Fully released artifacts: Code is MIT-licensed on GitHub, and a 101 GB HuggingFace dataset ships the raw Amber trajectories, per-frame MMGBSA scores, ML-ready NumPy tensors, cross-validation splits, and the trained model weights.

#Technical Details

The pretraining corpus comes from 10 nanoseconds of atomistic MD for each of 1,464 starting poses drawn from the PDBBind CASF-2016 benchmark — a crystallographic pose plus five docked poses per complex — with snapshots saved every 10 picoseconds to yield 1,000 frames per trajectory. Every frame was scored with Amber's MMPBSA.py using Generalized Born model 5 in implicit water; 1,070,272 frames (76% of those collected) survived filtering into the training set. The best-performing encoder is an Equivariant Graph Matching Network with 6 layers, a hidden dimension of 128, and 5,237,445 parameters, initialized from a released ProtMD checkpoint before being trained for up to 600 epochs to regress the MMGBSA energy. Graph embeddings are average-pooled and passed through an MLP to a scalar. Non-equivariant GNN and E(n)-equivariant EGNN encoders were trained as ablations; the 445,312-parameter GNN baseline trails the EGMN substantially, and the authors note that improving the EGMN's inductive biases is the main lever for scaling to larger structures.

#Applications

The immediate use is rescoring: a virtual screening pipeline can dock a library with a fast scoring function, then apply SurGBSA where MMGBSA would otherwise be run, recovering most of the discrimination between correct and decoy poses at a cost that scales to large libraries. Because the checkpoint is applied to new poses without per-target refitting, it slots into existing structure-based design workflows as a drop-in rescoring stage, and its learned embeddings offer a starting point for downstream protein-ligand property models. The released trajectory dataset is independently useful as a labeled MD corpus for anyone building representation learners over molecular dynamics.

#Impact

SurGBSA is the first demonstration that pretraining on MD trajectories, rather than static structures, measurably improves a learned model of protein-ligand energetics, and it is unusually complete in what it releases: weights, code, splits, and the full simulation corpus that produced them. The limitations are equally clear. The preprint has not been peer reviewed; the pretraining corpus is derived from a single benchmark family, so transfer to targets unrelated to CASF-2016 has not been demonstrated at scale; and the evaluation covers pose ranking rather than absolute binding affinity. The surrogate also inherits whatever systematic errors MMGBSA itself carries — it reproduces the physics method, it does not improve on it.