Physics-informed graph neural network that scores protein-ligand binding affinity and screens compounds from 3D complexes, trained on augmented poses.
No providers recorded yet. Browse all providers
A structure-based scoring function is asked to do two jobs that pull against each other. In virtual screening it must separate the few true binders from thousands of docked decoys; in lead optimization it must rank close analogs of one scaffold by affinity. Deep learning models trained only on crystal structures regress affinity well but are easily fooled by computer-generated poses, while models trained to reject decoys lose their grip on affinity. The root cause is data: PDBbind holds only a few thousand high-quality complexes with measured affinities, too few for a network to learn which geometric features actually matter.
PIGNet2, developed by Woo Youn Kim's group at KAIST with HITS Inc. and published in Digital Discovery in 2024, attacks that problem from two directions. First, it builds physics into the architecture: the network does not output an affinity directly but predicts the parameters of pairwise interaction energies between protein and ligand atoms, which are then summed. Second, it introduces positive data augmentation (PDA). A crystal structure is one snapshot at a local minimum of the energy surface, so poses that are geometrically and energetically close to it should carry the same affinity label. PIGNet2 generates such near-native poses and trains on them as if they were crystal structures, alongside the decoy-based negative augmentation used by its predecessor, PIGNet (Chemical Science, 2022).
Protein pocket atoms (residues within 5 Å of the ligand) and ligand atoms pass through a feedforward embedding, a gated graph attention network and an interaction network that exchanges information between the two molecules; the joint features parameterize the four physics terms. Training used the PDBbind v2020 refined set minus the CASF-2016 core set and redundant entries: 5,046 crystal complexes, expanded to about 375,000 PDA poses and roughly 1.7 million negative poses from re-docking, cross-docking and random docking. Reported results are an ensemble of four randomly initialized models.
On CASF-2016, the ensemble reached a Pearson R of 0.747 for scoring, a Spearman ρ of 0.651 for ranking, 93.0% top-1 docking success and a top-1% enrichment factor of 24.9 for screening. On DUD-E, which it was not trained on, the top-1% enrichment factor was 31.2 using ten poses per ligand. On the 2015 derivative benchmark of eight congeneric series, its average R of 0.64 exceeded MM-GB/SA (0.40) and Glide SP (0.29); on the 2020 Merck FEP series it averaged 0.43. Ablations showed that adding PDA to a model trained with negative augmentation raised the DUD-E enrichment factor from 21.1 to 24.3 and the derivative-benchmark R from 0.39 to 0.64. The model treats solvation only implicitly, and it anti-correlates on the Eg5 series, where flexible chains extend into solvent.
PIGNet2 is a rescoring function for medicinal chemists and computational groups: dock a library with a conventional tool, minimize poses with Smina as the authors recommend, then score with the four-model ensemble to prioritize hits or rank a congeneric series before committing to free energy calculations. Because each prediction breaks down into per-atom-pair energy terms, it can also point to which contacts a proposed modification gains or loses. HITS's HyperLab platform cites PIGNet2 as the physics-informed model behind the rigid-receptor mode of its Hyper Binding feature, the sibling of the Hyper Binding Co-folding model.
PIGNet2 showed that a model trained on a few thousand crystal complexes can match task-specialized networks on both scoring and screening when physical inductive bias is combined with augmentation on both sides of the decision boundary, and its near-native augmentation idea is independent of the architecture. Later structure-based scorers, including PBCNet2.0 and OmniScore, benchmark against it. Code, the four pretrained checkpoints and the augmented training and benchmark data are public, so the results can be reproduced and the model retrained. It depends on an input pose, so its accuracy is bounded by the docking or structure prediction that produces one.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.