DualDiff

A drug that engages two proteins at once can suppress a pathway and its escape route together, which is why dual-target compounds are attractive against resistance in oncology. Structure-based generative models cannot be trained on that problem directly: the field's protein-ligand corpora are built from crystal structures of one ligand in one pocket, and there is no comparable body of ligands solved against two targets. Train on what exists and you get a model that knows one lock at a time.

DualDiff sidesteps the missing data by never training on it. It treats an already-trained single-target diffusion model as a factor in a product of distributions — the density over ligands that fit pocket one, multiplied by the density over ligands that fit pocket two — and samples from that product by composing the two conditional generative processes at inference. Because each factor is the same frozen network evaluated on a different pocket, no new parameters are fit and no dual-target examples are needed. The paper introduces two compositions: CompDiff, which averages the drift terms the network predicts at its output for the two pockets, and DualDiff, which builds two complex graphs sharing one set of ligand nodes and averages the SE(3)-equivariant messages at every layer of the network rather than only at the end.

The work is by Xiangxin Zhou and colleagues at the School of Artificial Intelligence, University of Chinese Academy of Sciences, and the Institute of Automation, CAS, with collaborators at the University of Illinois Urbana-Champaign, Tsinghua University and Peking University, published at NeurIPS 2024.

#Key Features

• Zero-shot over a frozen backbone: the pretrained checkpoint is used unmodified for every target pair. The authors demonstrate the framework with TargetDiff and introduce no new pretrained weights.
• Layer-wise message composition: DualDiff averages the equivariant hidden and coordinate updates from both complex graphs inside each layer, which recovers better samples than CompDiff's single composition at the output.
• Pocket alignment by binding priors: two pockets have no natural common frame, so a probe ligand is docked into both and the docked poses are superposed to place the pockets in one coordinate system.
• A dual-target benchmark: 12,917 target pairs over 438 unique targets, each with a reference ligand and a protein-ligand complex structure.
• Linker design as a baseline: fragments from the two reference ligands are docked and joined with DiffLinker and LinkerNet, formalizing the medicinal chemist's pharmacophore- combination strategy as a comparison point.

#Technical Details

The backbone denoises 3D atom coordinates and categorical atom types jointly with an SE(3)-equivariant network over a k-nearest-neighbour protein-ligand graph. DualDiff runs that network twice per step, on graph one and graph two, and averages the per-layer updates before continuing; CompDiff instead averages the predicted drifts at the output, with a strength hyperparameter set to 1/2, which the paper frames as Langevin sampling from a sum of the two pockets' energy functions.

The dataset is assembled from DrugCombDB synergistic combinations, targets resolved through DrugBank and the Therapeutic Target Database, and structures taken from PDBbind and the PDB where available and from the AlphaFold Protein Structure Database otherwise, with P2Rank locating pockets and AutoDock Vina producing the complexes. Across all 12,917 pairs, DualDiff generates molecules whose affinity exceeds both reference ligands 36.3% of the time, against 35.9% for CompDiff, 35.7% for LinkerNet, 24.6% for DiffLinker, 29.0% for TargetDiff sampling on one pocket and 0.2% for Pocket2Mol. Its worse-of-two-pockets Vina score averages -7.66 kcal/mol, versus -6.57 for TargetDiff, whose single-pocket scores collapse on the second target. Ablations show the alignment matters: aligning pockets by geometric centre instead of binding priors drops the dual-affinity rate to 30.0%.

#Applications

The natural targets are protein pairs whose simultaneous inhibition is already known to synergize — a kinase and a resistance paralog, or two nodes of one oncogenic pathway — where a chemist would otherwise fuse pharmacophores by hand. Since the method requires only a pretrained structure-based generator and two pocket structures, teams can point it at pairs for which no dual-target ligand has ever been solved.

#Impact

DualDiff established dual-target 3D ligand generation as a benchmarked task rather than a bespoke exercise, and its 12,917-pair dataset has been adopted by later work in the area; FusedBFN evaluates on it and treats DualDiff and CompDiff as the baselines to beat. The honest limits are that generality across pretrained generators is argued rather than demonstrated — TargetDiff is the only backbone tested — and that every result is in-silico docking, with no synthesis or binding assay reported.