BindDM

Most of a binding pocket does nothing. Of the hundreds of protein atoms within reach of a bound ligand, only a handful — a hydrogen-bond donor here, a hydrophobic shelf there — hold the molecule in place, and the rest is scaffolding. Target-aware diffusion models for structure-based drug design typically hand the denoiser the whole protein–ligand complex as one flat graph, where every atom contributes on equal footing and the few interactions that decide affinity are diluted by everything around them. The usual remedy is to carve the problem up in advance, with predefined fragments, motifs or subpockets, but that imposes a vocabulary chosen before the target is seen.

BindDM (Binding-Adaptive Diffusion Models) instead learns which atoms matter, and re-decides at every step of generation. A learnable structural pooling — Binding-Adaptive Subcomplex Extraction, or BASE — scores every node of the complex graph and keeps the top-ranked half as a subcomplex. Two SE(3)-equivariant networks then run side by side, one over the full complex and one over the subcomplex, and a pair of cross-hierarchy interaction nodes shuttle pooled binding context between them so each view informs the other. Because BASE is applied inside every layer of the denoising network, the model's idea of "the essential binding site" keeps shifting as the ligand takes shape.

The model came out of Tsinghua University's Shenzhen International Graduate School with Peng Cheng Laboratory, Peking University, the University of Science and Technology of China, the University of Chinese Academy of Sciences and ByteDance, and was published at AAAI 2024. It builds on the TargetDiff codebase and data pipeline, keeping that model's equivariant score network but replacing its single-hierarchy denoiser.

#Key Features

• Binding-adaptive subcomplex extraction: A learnable pooling layer scores every atom from the complex adjacency and hidden states, then keeps the top half. Checked against a pretrained binding-affinity model's atom attributions, the retained set agrees about 65% of the time, against 50% for random selection.
• Cross-hierarchy interaction nodes: Two summary nodes, one per hierarchy, exchange information through a gated transmission module built from attention pooling, sum pooling, a gated fusion and a GRU, so context flows complex-to-subcomplex and back.
• Equivariant by construction: Shifting the protein atoms' center of mass to zero and parameterizing the Markov transition with an SE(3)-equivariant network gives a likelihood invariant to rotation and translation of both hierarchies.
• Single-stage generation: Coordinates and discrete atom types are sampled jointly and bonds assigned afterward, with no external software supplying decomposed priors and no fragment library.
• Arbitrary pockets at inference: One trained checkpoint samples ligands for any protein supplied as a PDB file, so a new target needs no retraining.

#Technical Details

The denoiser is a 9-layer SE(3)-equivariant transformer with 128-dimensional hidden states and 16 attention heads, operating on a k-nearest-neighbor graph (k = 32) rebuilt from the current coordinates at each layer. Edge features cross radial basis functions at 20 centers between 0 and 10 Å with a four-way bond-type vector. Generation runs 1000 diffusion steps with a sigmoid β schedule for coordinates and a cosine schedule for atom types, trained with Adam and the atom-type loss weighted 100× against the coordinate loss.

Training uses CrossDocked2020, filtered to poses under 1 Å docking RMSD and proteins under 30% sequence identity, giving 100,000 protein–ligand pairs and 100 held-out test proteins. On that split BindDM reaches an average Vina Score of −5.92 (median −6.81), Vina Min −7.29, and Vina Dock −8.41, with 64.8% of generated molecules binding more tightly than the pocket's reference ligand, alongside QED 0.51, SA 0.58 and diversity 0.75. Ablations isolate the mechanism: dropping the cross-hierarchy interaction entirely moves the average Vina Score to −5.04, and replacing BASE with a random half-selection of atoms gives −4.79. Drawing the subcomplex from the joint complex also beats drawing it from the pocket (−5.37) or the ligand (−5.46) alone.

#Applications

BindDM suits early hit generation against a target with a solved or modeled structure: given a pocket, it proposes 3D molecules already posed in the site, ready for docking, minimization or free-energy triage without a conformer-generation step. The subcomplex it selects is interpretable output in its own right, flagging which atoms the model treats as carrying the binding signal.

#Impact

BindDM's contribution is the observation that a generative model can decide for itself which part of a binding site to attend to, and that the decision should be revisited as generation proceeds — an alternative to the fragment- and subpocket-based decompositions common in pocket-conditioned generation. Its practical reach is limited by what was released: the repository ships training, sampling and evaluation code but no trained checkpoint, carries no license file, and hardcodes the authors' local dataset paths, so reproducing the numbers means retraining from scratch. The gains are in-silico and docking-scored; no generated molecule has been synthesized or assayed.