Deep-learning pharmacophore modeling from a protein pocket alone, scoring ligands by graph matching for ultra-large virtual screening on one CPU.
No providers recorded yet. Browse all providers
A medicinal chemist looking at a binding site does not enumerate every atom pair. They notice that one subpocket wants an aromatic ring, that a backbone carbonyl wants a hydrogen-bond donor a few angstroms away, that a charged residue wants a counter-charge. That abstraction is a pharmacophore, and it has traditionally been built by hand or from the crystal structure of a ligand already known to bind. Neither route helps when the target is new, or when the structure came from a predictor rather than a diffractometer.
PharmacoNet, from Seonghwan Seo and Woo Youn Kim at KAIST, derives that abstraction from the protein alone. The binding site is voxelized into a 3D image and handed to an instance segmentation network: the model first decides which protein functional groups are genuine interaction hotspots, then paints, for each one, the region of empty space where a complementary ligand group would sit. The result is a set of typed pharmacophore points with positions and radii — a hypothesis about what a good binder looks like, produced once per target before any compound is considered.
Scoring then avoids deep learning entirely. A candidate's functional groups and the pharmacophore points are each treated as a complete graph, and a matching algorithm finds the feasible correspondences under distance constraints, ranked by a distance-likelihood function with only seven parameters. The separation matters: the network describes a pocket, a task the PDB covers well, while affinity ranking falls to an analytical function that cannot overfit a small training set. Where the deep docking model KarmaDock loses screening power on targets below 0.5 sequence similarity to its training proteins, PharmacoNet holds steady, and likewise for ligands dissimilar to anything it trained on.
The pocket is rendered at 0.5 Å resolution into a 64×64×64 grid of atom, residue and functional-group channels, encoded by a 3D Swin Transformer V2 backbone inside a feature pyramid network. Heads predict short- and long-range cavity masks, a hotspot score per functional-group token, and a density mask per hotspot covering six interaction types: hydrophobic contact, hydrogen bond, halogen bond, π–π stacking, cation–π and salt bridge. Training used PDBbind v2020 — 17,658 complexes for training, 1,500 for validation, the 285 CASF-2016 complexes held out — supervised by complex-based pharmacophores extracted with PLIP. On the PDBbind core set PharmacoNet is 3,956× faster than AutoDock Vina and 34,117× faster than GLIDE SP, and the margin widens for large ligands. Screening power on DEKOIS2.0 and the decoy-free LIT-PCBA benchmark exceeds AutoDock Vina, KarmaDock and docking-free scoring models including PLAPT, and beats the earlier automated protein-based pharmacophore methods Apo2ph4-Pharmit and PharmRL.
The intended role is pre-screening: run the whole library through PharmacoNet, keep the top few percent, and spend docking or a heavyweight scoring model only on those. The paper filters 1.6 million ChEMBL compounds against dihydrofolate reductase and an EGFR double mutant in under ten minutes on one 32-core CPU, retaining 40% and 70% of the top-10 hits at a 95% filtration rate. Because subtle pocket differences change the pharmacophore, the same machinery supports selectivity campaigns: screening 187 million ZINC20 compounds against the CB2 cannabinoid receptor took about 20 hours, and re-scoring the top 1% against CB1 surfaced 434 compounds with over 100-fold predicted selectivity. OpenPharmaco wraps the pipeline in a PyMOL-backed desktop interface for users without scripting expertise.
PharmacoNet's durable contribution is the division of labour between a learned pocket description and an analytical score, trading a modest amount of accuracy for generalization that holds on unseen targets and unseen chemistry. Its pocket encoder is reused as a frozen module by other groups and by the authors' later work, including RxnFlow, and the repository ships pretrained docking-proxy models built on it. The limits are stated by the authors: pharmacophore-level abstraction cannot distinguish salt bridges of differing strength or account for ligand internal energy, so it is a filter rather than a replacement for docking, and every reported result is computational, with no experimental confirmation of the cannabinoid hits.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.