READ

Generating a ligand directly inside a protein pocket forces a generative model to satisfy two demands at once. It has to explore chemical space widely enough to find something new, and it has to respect valence rules, synthetic accessibility, and the tight geometry of the binding site. Pocket-conditioned diffusion models are good at the first and inconsistent at the second: they routinely emit molecules with implausible substructures or steric clashes, which is why many pipelines bolt on hand-written validity filters after the fact.

READ — Retrieval-Enhanced Aligned Diffusion — attacks that trade-off by giving the diffusion process something to imitate. Developed at Shenzhen University with collaborators at the University of Nottingham Ningbo China and the Hong Kong Polytechnic University, it combines an SE(3)-equivariant diffusion generator with retrieval-augmented generation. A contrastively pretrained atom-level encoder builds a latent manifold from six million force-field-optimized ZINC molecules; at inference, a structural index of protein pockets retrieves ligands bound to geometrically similar sites, and their pretrained embeddings are fused into the denoiser at every reverse-diffusion step. Real chemistry is injected as a prior exactly where the sampler would otherwise drift.

The model sits alongside catalog entries like PocketXMol in the structure-based drug design family, but it is the first to route retrieval-augmented generation through an equivariant diffusion process rather than stitching retrieved fragments together.

#Key Features

• Retrieval-guided denoising: Embeddings of ligands from structurally matched pockets are fused by cross-modal attention into each reverse-diffusion step, acting as attractors that pull the generated point cloud toward known pharmacophoric motifs.
• Contrastively pretrained latent space: An SE(3)-equivariant GNN encoder trained with InfoNCE across four augmentation views — atom masking, bond perturbation, subgraph removal, and coordinate noise — encodes validity constraints into the latent geometry rather than into post-hoc filters.
• Hierarchical pocket retrieval: A coarse alignment against seed pockets selects an entry point, followed by a K-nearest-neighbour refinement, so lookup cost stays negligible relative to sampling.
• Dual-space diffusion: Continuous 3D coordinates evolve under a variance-exploding SDE while discrete atom types follow an absorbing-state process.
• Compact model: Both released variants are 12M parameters, differing only in whether they run 1,000 or 2,000 denoising steps.

#Technical Details

The retrieval index is built from the CBGBench training split with TM-align and DaliLite, yielding a graph of 2,200 protein pocket nodes and 166,000 ligand nodes; the default search returns 40 neighbours and fuses the top four ligand embeddings. Training ran for 500,000 iterations on a single 48 GB NVIDIA A6000, roughly six days, with convergence around 350,000 iterations. On CBGBench, READ-2k takes the top spot in a Friedman weighted ranking over thirteen metrics and twelve baselines including TargetDiff, DecompDiff, MolCraft, VoxBind, Pocket2Mol, and DiffSBDD. It reaches a mean AutoDock Vina Dock energy of −7.79 kcal/mol and the best mean percentage binding gap at 12.33, with a validity rate of 0.98 — the highest in the comparison. Filtering the 100 generated ligands per pocket for drug-likeness, the top candidate beats the crystallographic reference ligand in 96.8% of pockets under Dock mode, with a mean improvement of 3.23 kcal/mol. Generating 100 ligands takes 18 minutes at 1,000 steps and 35 minutes at 2,000.

#Applications

READ targets early-stage hit generation for a protein target with a solved or predicted structure: given a pocket, it proposes ligands that dock well and pass drug-likeness filters without a downstream cleanup stage. The retrieval component makes it particularly suited to targets that resemble well-characterized pockets, where decades of medicinal chemistry are already encoded in the ligand database. It also serves as a template for adding retrieval to other geometric generative models.

#Impact

The paper's central claim — that retrieval and diffusion can be co-optimized rather than chained — is supported by an ablation showing that removing latent alignment roughly doubles the steric-clash rate. The authors are direct about the corresponding weakness: performance depends on a fixed retrieval graph of about two thousand pockets, and novel targets outside that coverage stand to benefit least, which is precisely the regime where generative design is most valuable. Pretraining is also separated from the diffusion stage rather than trained jointly. No source code or pretrained weights have been released, so the reported numbers have not been independently reproduced.