Physics-informed graph neural network scoring protein-ligand binding affinity as a sum of learned atom-pair energy terms, for docking and screening.
No providers recorded yet. Browse all providers
A deep learning scoring function trained on crystal structures learns to predict affinity from complexes that are, by construction, already in their most stable pose. Virtual screening asks it something different: to pick the native pose out of dozens of docked decoys, and to pick a true binder out of thousands of molecules that do not bind at all. Structure-based networks trained only on PDBbind often score well on affinity correlation yet fail at both tasks, because nothing in their training data tells them what a wrong pose or a non-binder looks like.
PIGNet, developed by Woo Youn Kim's group at KAIST with HITS Inc. and published in Chemical Science in 2022, addresses both gaps. Instead of outputting an affinity directly, the network predicts the parameters of physics-shaped pairwise interaction equations, modeled on the empirical terms of AutoDock Vina, for every protein–ligand atom pair, and sums them into a total binding free energy. It is then trained on a greatly enlarged dataset of computer-generated decoy poses and non-binding complexes, with losses that require those to score worse than the true complex. Its successor, PIGNet2, kept this design and added augmentation with near-native poses.
Protein pocket residues within 5 Å of the ligand and the ligand atoms form one graph with two adjacency matrices, one for covalent bonds and one for intermolecular contacts. Three gated graph attention layers update node features over covalent bonds, three interaction-network layers exchange information across the interface, and fully connected layers map atom-pair features to the energy-term parameters. Training used 4,514 complexes from the PDBbind v2019 refined set after removing the CASF-2016 core set, plus 1,652,085 augmented complexes generated with Smina: 292,518 docking decoys, 831,885 random screening complexes built from the IBS compound library, and 527,682 cross-binding complexes pairing PDBbind ligands with other targets.
On CASF-2016, the ensemble reached 87.0% top-1 docking success, a top-1% screening enrichment factor of 19.6 with a 55.4% success rate, scoring Pearson R of 0.761 and ranking Spearman ρ of 0.682. AutoDock Vina scored 84.6% docking success with an enrichment factor of 7.7, and GlideScore-SP an enrichment factor of 11.4. On the external CSAR NRC-HiQ sets 1 and 2, R was 0.768 and 0.800. Ablations separated the two ideas: augmentation raised the single model's screening enrichment from 6.6 to 18.5 at a small cost in scoring R (0.792 to 0.749), while an otherwise identical graph network without the physics equations reached only 66.6% docking success and an enrichment factor of 10.2 with the same augmented data.
PIGNet is a rescoring function for structure-based virtual screening and pose selection: dock a library with a conventional program, then rank the resulting complexes by predicted affinity and use the uncertainty estimate to set aside unreliable hits. Because each prediction decomposes into atom-pair and substructure energies, medicinal chemists can see which moiety of a lead the model credits with its affinity. The repository includes training and benchmark scripts and a pretrained checkpoint, with the augmented training data and CASF-2016 inputs on Zenodo under CC BY 4.0. The code repository carries no license file.
PIGNet showed that physics-constrained output and training on deliberately wrong complexes each improve generalization, and that they add to each other: augmentation alone could not recover the docking performance lost by removing the physics equations. It also argued that affinity correlation on crystal structures is a poor proxy for docking and screening performance. The architecture was carried forward in PIGNet2, and later scorers such as OmniScore benchmark against the PIGNet family. It depends on an input pose, and its treatment of solvation and entropy is limited to what the four energy terms and the rotor penalty capture.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.