TargetDiff

Placing the last carbon of a benzene ring is easy: five coplanar neighbours already fix where it goes. Placing the first atom of a molecule inside an empty binding pocket is not, because almost no context exists yet. Pocket-conditioned generators that grow a ligand atom by atom inherit that asymmetry — early placements are guesses the rest of the molecule must live with, and the model never scores the whole 3D structure. The alternative, painting molecules into a voxel grid of atomic densities, traded that problem for two others: the network is not equivariant to rotation, and voxel count grows with the cube of the pocket size.

TargetDiff answers both by denoising the entire ligand at once. Developed by Jiaqi Guan, Wesley Wei Qian and colleagues at the University of Illinois Urbana-Champaign with Peking University and the Institute for AI Industry Research at Tsinghua University, and presented at ICLR 2023, it represents pocket and ligand alike as sets of atoms carrying 3D Cartesian coordinates, then learns a denoising diffusion process that recovers a ligand from noise conditioned on the protein atoms. Shifting the protein's centre of mass to the origin and parameterising the reverse transition with an SE(3)-equivariant graph network makes the likelihood of the generated complex invariant to how that complex happens to be oriented in space.

The trained generator also doubles as an affinity model. Feeding a real protein-ligand complex through the same network with the coordinate branch frozen yields per-atom hidden states and a denoised atom-type distribution whose entropy tracks measured binding strength — a scoring signal obtained without labelled affinity data and without retraining.

#Key Features

• Joint diffusion over geometry and chemistry: Gaussian noise corrupts atom coordinates while uniform categorical noise corrupts one-hot atom types, and a single reverse process recovers both, so element identity and position are resolved together rather than in stages.
• Invariant likelihood by construction: centring on the pocket's centre of mass and using an equivariant transition network gives a generative distribution provably invariant to global rotation and translation of the binding complex.
• Non-autoregressive sampling: every atom is refined at every step, which aligns training with sampling, removes exposure bias, and lets the model reason about the molecule as a whole rather than committing to an arbitrary atom order.
• Affinity ranking without labels: the entropy of the denoised atom-type distribution acts as a scoring function for candidate ligands, carrying information complementary to physics-based scores such as AutoDock Vina.
• Sampling into an arbitrary pocket: a released script takes a 10 Å region around a reference ligand in any PDB file and generates binders from the distributed checkpoint.

#Technical Details

The denoiser is nine equivariant interaction layers in which atom hidden embeddings and ligand coordinates update alternately, implemented as graph attention with 16 heads and 128 hidden features; protein coordinates are masked out of the coordinate update. The atom count is drawn from a prior estimated from training complexes with similar pocket sizes, and OpenBabel infers bonds from the finished coordinates. Training used CrossDocked2020, whose 22.5 million docked complexes were filtered to poses under 1 Å RMSD with sequence identity below 30%, leaving roughly 100,000 training complexes and 100 held-out pockets. On those test pockets the paper reports mean and median Vina Dock energies of -7.80 and -7.91 kcal/mol, with 58.1% of generated molecules binding more tightly than the reference ligand; drug-likeness (QED 0.48) and synthetic accessibility (SA 0.58) sit at or below the reference molecules rather than above them. Bond-length distributions track reference more closely than the baselines compared, at Jensen-Shannon divergence 0.369 for single C-C bonds. Augmenting an EGNN affinity regressor with TargetDiff's hidden features improved RMSE from 1.445 to 1.374 and Pearson correlation from 0.648 to 0.680 on a PDBBind v2020 time split.

#Applications

The practical workflow is hit generation against a target whose pocket geometry is known from crystallography or structure prediction: supply the pocket, sample candidate ligands, and rank them. Because the same checkpoint provides a scoring signal, generation and triage share one model, and the extracted features can also be bolted onto a supervised affinity predictor. Bonds are recovered post hoc rather than generated, so candidates need a validity check before they reach a chemist.

#Impact

TargetDiff established diffusion as the default formulation for pocket-conditioned 3D molecule generation, and its CrossDocked2020 protocol — 100 test pockets, Vina Score/Min/Dock, QED, SA, diversity — became the comparison table nearly every later structure-based generator reports. Successors including MolPilot, DrugFlow and TransDiffSBDD are measured against it, and its MIT-licensed codebase is reused directly as a diffusion backbone. The caveats are that evaluation is entirely in silico through docking scores, that post-hoc bond inference yields an excess of seven-membered rings relative to reference ligands, and that checkpoints and preprocessed data are distributed through Google Drive rather than a versioned registry.