FuseDiff

A molecule that inhibits two proteins at once has to do it in two different shapes: the same covalent graph, folded one way in the first pocket and another way in the second. Structure-based generative models are trained on corpora that never show this — every entry pairs one ligand with one pocket — so dual-target design has been assembled from pieces. A graph-first model proposes a 2D molecule conditioned on both targets, single-target predictors place it in each site separately, and any disagreement between the two poses is absorbed by a downstream conformational search.

FuseDiff learns the whole object at once: a conditional diffusion model over the triple of molecular graph, pose in pocket one, and pose in pocket two, denoising all three jointly given both pockets, so a single reverse trajectory emits a ligand together with the two target-specific conformations it adopts. The coupling lives inside the denoiser: Dual-target Local Context Fusion (DLCF) merges each ligand atom's neighbourhood from both pocket graphs into one augmented graph before message passing, then updates coordinates separately within each complex, letting the two poses share chemistry while adapting geometry independently.

The work comes from Sun Yat-sen University, with collaborators at Renmin University of China and Tsinghua University, and was presented at KDD 2026. Two other dual-target entries in the catalog take the opposite route: DualDiff composes a frozen single-target TargetDiff across two rigidly aligned pockets with no dual-target training data, and FusedBFN fuses two pocket-conditioned Bayesian flow distributions. FuseDiff instead derives a dual-target training set and fits the joint density end to end.

#Key Features

• Joint graph-and-pose generation: One reverse diffusion trajectory produces the shared molecular graph and both pocket-specific conformations, removing the pose-recovery stage that staged pipelines depend on.
• Dual-target Local Context Fusion: Each denoising layer builds a point-cloud graph per pocket, deduplicates the shared ligand nodes into one augmented graph, and passes messages over it, so every ligand atom sees both targets' local environments.
• Explicit bond generation: Bond types are diffused and denoised alongside atom types rather than inferred post hoc from coordinates, which keeps the two poses attributable to a single molecule.
• Symmetry by construction: The denoiser is permutation-invariant in the two targets and SE(3)-equivariant within each complex, while deliberately not constraining the two poses to a single rigid transformation.
• Pre-docking pose assessment: Target-specific poses come directly out of the model, so binding-affinity proxies can be read before any conformational search.

#Technical Details

The denoiser is an L-layer SE(3)-equivariant message-passing network. Gaussian diffusion corrupts ligand coordinates in both complexes; categorical diffusion corrupts atom and bond types under a bond-first schedule that absorbs bonds into a "none" type. Training data is BN2-DT, derived from the 197,488-ligand high-confidence subset of BindingNet v2: two single-target complexes form a dual-target tuple when they share a graph-isomorphic ligand across distinct pockets, giving 58,058 tuples over 801 unique protein targets and 21,721 unique ligands. Evaluation uses the DualDiff benchmark of 12,917 target pairs as a held-out test set, with 2.57% exact target-pair overlap against BN2-DT.

Across 10 ligands sampled per target pair, FuseDiff reports the best QED (0.64), synthetic accessibility (0.68) and Lipinski compliance (4.92) of the methods compared, Vina Dock averages of -9.11 and -9.16 in the two pockets, and a dual high-affinity rate of 49.2% against 40.5% for DualDiff. Dual-Validity — the share of samples whose two generated poses can both be assembled into the same chemically valid molecule — reaches 61%. Ablations show how load-bearing the design choices are: removing explicit bond generation collapses Dual-Validity to 3%, and removing DLCF to 2%.

#Applications

The target is polypharmacology: compounds that engage two proteins to improve efficacy or blunt resistance, where the chemist needs not just a molecule but a plausible binding mode in each site. The paper works a dual-kinase Alzheimer's case, GSK3β (PDB 6Y9S) and JNK3 (PDB 4WHZ); from 1,000 sampled molecules per method, 78 passed a combined drug-likeness filter against 15 from DualDiff, and a second pair, RORγt-DHODH, gave 41 against 5. Shipping two poses with each candidate lets triage on cross-target compatibility happen before any docking campaign.

#Impact

FuseDiff reframes dual-target design as one conditional density rather than a pipeline, and the ablations argue that topological consistency between the two poses is not something a post-hoc bond-inference step can recover. It also makes pre-docking dual-pose quality measurable, which earlier methods structurally could not report. The caveat sits in the training data, and the authors state it plainly: BN2-DT is a surrogate, built by pairing independently observed single-target complexes that happen to share a ligand graph, so it does not sample the true dual-target co-binding distribution. Code is released under MIT, with the trained checkpoint and processed training set distributed separately; no parameter count is reported.