Structure-based drug design framework that scores interaction-aware fragments against protein subpockets, then diffuses a 3D scaffold to link them.
No providers recorded yet. Browse all providers
Generative models for structure-based drug design split roughly into two camps. Atom-level models — including pocket-conditioned diffusion approaches such as Apo2Mol — place every heavy atom individually, which gives maximum freedom but routinely yields molecules that are hard to synthesize and offers no account of why a particular atom went where it did. Fragment-level models assemble molecules from chemically sensible building blocks, which fixes synthesizability, but typically choose those fragments on shape or a score that collapses the physics of binding into one opaque number.
This framework, from Peng Cheng Laboratory with South China University of Technology, the University of Kentucky, and the University of Agder, takes the fragment-level route and makes the fragment-selection step explicit. It decomposes a binding site into subpockets and a ligand into arms — the fragments that make direct contact — and trains a concept-based neural network to predict, for a given arm-subpocket pair, a set of named physical quantities: hydrogen bonds, hydrophobic contacts, π-π stacking, salt bridges, water bridges, halogen bonds, and π-cation interactions, alongside two geometric concepts covering surface complementarity and how much of the subpocket's non-polar volume the arm fills.
Those predicted concepts are what rank candidate arms. Only after the arms are chosen and positioned does a second stage — an E(3)-equivariant diffusion model conditioned on the target structure and the selected arms — generate the scaffold that ties them into a single connected molecule. Chemistry is decided first and geometry second, and the concept scores that drove each choice remain available for inspection.
The arm encoder is SchNet, capturing atomic and geometric properties; the protein subpocket encoder uses continuous-discrete convolution. Subpockets are defined by running PLIP over experimental complexes, and the arm library is produced by BRICS fragmentation of ligands from PDBbind2020 and CrossDocked, filtered to poses under 1 Å RMSD from targets sharing under 30% sequence identity. The concept model trains on roughly 400,000 arm-subpocket pairs; the diffusion scaffold generator trains on roughly 100,000 filtered CrossDocked2020 protein-ligand complexes.
Evaluation uses 100 held-out novel protein targets, against AR, GraphBP, TargetDiff, and DecompDiff. The model reports the best QED (0.54 versus 0.51 for the next best) and synthetic accessibility (0.63 versus 0.60), and the strongest minimized and redocked binding scores, at −7.07 Vina Min and −8.57 Vina Dock against DecompDiff's −7.04 and −8.39. Its un-minimized Vina Score of −4.79 trails both diffusion baselines, and diversity at 0.69 sits slightly below TargetDiff and GraphBP. Jensen-Shannon divergence between generated and reference bond-distance distributions is lowest across every bond type measured, indicating more realistic local geometry. Ablations are informative about which concepts matter: replacing sampled arms with the reference ligand's own arms drops Vina Dock to −6.89 and diversity to 0.37, while removing the hydrogen-bond concept alone costs the most, dropping Vina Dock to −5.80.
The intended use is hit generation against a protein target with a resolved or modeled binding site, in settings where medicinal chemists need candidates they can plausibly make. Because generation begins by choosing fragments from a retrosynthetically derived library, output molecules decompose into recognizable building blocks, and the retained concept predictions give a chemist a specific, checkable hypothesis — which contacts a fragment was selected to form — rather than a bare score.
The work's contribution is to move interaction physics out of an implicit scoring function and into an explicit, named intermediate representation that both drives generation and survives it. The ablations support that this is doing real work: the hydrogen-bond and spatial concepts each account for a large share of the binding-affinity result, and sampled arms outperform the reference ligand's own arms on every metric. Limitations are concrete. The method degrades when a subpocket is too small to seat an arm with enough atoms, which weakens binding affinity in tightly constrained sites. Evaluation is entirely in silico, on docking scores and cheminformatic proxies rather than assays. The paper is an arXiv preprint, and no code or trained weights have been released, so the results cannot currently be reproduced or the model applied to new targets.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.