FusedBFN

Most structure-based drug design models follow a "one target, one drug" premise: condition on a single binding pocket and sample ligands that complement it. Many complex diseases respond better to a compound engaging two proteins at once, a polypharmacological strategy that can improve efficacy and blunt resistance. Dual-target generation resists direct training because structural data for ligands bound to two targets is scarce, so prior 3D work repurposed single-target diffusion models by adding a predictive drift term pushing the reverse trajectory toward a second pocket. That correction acts on the sampled molecule rather than the underlying distributions, leaving both pockets' structural information loosely combined.

FusedBFN, from the School of Computer Science at Shanghai Jiao Tong University with the Guangdong Laboratory of Artificial Intelligence and Digital Economy in Shenzhen, reframes the problem as distribution fusion. Bayesian flow networks generate by updating the parameters of a data distribution through Bayesian inference rather than denoising samples, giving one continuous parameter space for both continuous atom coordinates and discrete atom types. FusedBFN combines the two target-conditioned sender distributions in that space with a product-of-experts formulation, amplifying molecular patterns supported by both pockets and suppressing modes compatible with only one.

The framework's contribution is entirely at inference time. The network it drives is a frozen, pretrained single-target Bayesian flow model of the MolCRAFT family, reused unmodified across every target pair with no additional training or fine-tuning. It sits beside catalog entries sharing the parameter-space idea: MolPIF, which interpolates distribution parameters with flow matching, and CByG, which adds gradient guidance atop the same pretrained backbone family.

#Key Features

• Product-of-experts fusion: The sender distributions conditioned on each pocket are multiplied, yielding closed-form fused Bayesian updates — a Gaussian mean averaging the two coordinate estimates, and a softmax over summed categorical evidence for atom types.
• Accuracy additivity preserved: Sender accuracies stay additive under fusion for both variable types, so the Bayesian flow keeps its tractable marginal form.
• SE(3)-equivariant generation: Equivariance holds when the shared network is SE(3)-equivariant and the aligned pocket pair's center of mass is shifted to zero, retaining the geometric inductive bias 3D molecule generators depend on.
• Chemically aware pocket alignment: Aligning pockets via docked probe ligands weights heavy atoms above light ones, since heavy atoms carry the ligand's topological scaffold and give the more distinctive geometric constraints.
• Prior-free alignment fallback: Without a probe ligand, P2Rank extracts pocket surface atoms, RANSAC performs coarse registration and ICP refines it — reaching a mean Max Vina Dock of -7.86 kcal/mol against -8.15 for the prior-based default.

#Technical Details

The shared network is an SE(3)-equivariant graph neural network predicting a molecular estimate from the current parameter state for each pocket, with sampling carried out directly in parameter space to avoid re-injecting noise. Evaluation uses a dual-target benchmark of 12,917 target pairs, curated from synergistic drug combinations rather than arbitrary pairings and involving 438 unique drugs, generating 10 molecules per pair.

FusedBFN reaches a mean Max Vina Dock of -8.02 kcal/mol — the worse of a molecule's two docking scores, and the metric for balanced dual binding — against -7.60 for DualDiff, -7.50 for CompDiff, -7.30 for MolCRAFT, and -6.56 for TargetDiff. Dual high affinity, meaning stronger binding than the reference ligand on both targets at once, reaches 57.8% of molecules versus 51.2% for DualDiff, significant at p < 0.05. Mean QED is 0.56 and SA 0.69, while diversity of 0.69 sits marginally below the baselines. Median strain energy is 194 against 262 for MolCRAFT and 1112 for DualDiff, and mean steric clashes on the second target fall to 12.0 from DualDiff's 18.3. Ablations show parameter-space fusion beating sample-space averaging of the two molecular estimates, at -8.15 versus -7.78 mean Max Vina Dock.

#Applications

The immediate use is generating candidate scaffolds for a chosen pair of targets in polypharmacology campaigns — kinase pairs, or a primary target plus a resistance-conferring paralog — where structures exist for both pockets but no dual-active chemical matter does. Because the backbone is frozen and applied zero-shot, a medicinal chemistry team can point it at a new pair without assembling training data, and the prior-free alignment path covers targets with no suitable probe ligand.

#Impact

FusedBFN shows that a dual-target constraint can be enforced inside the generative mathematics itself, by fusing distributions in a Bayesian flow network's parameter space, rather than patched on as a drift correction to a single-target sampler. Its gains in balanced binding and conformational quality argue that the fused parameter-space route is the stronger formulation, and the construction extends toward the multi-target settings the authors identify as future work. Two limitations bound the claims: the work is a preprint awaiting peer review, with no code, weights, or license released, and all results are in-silico docking and cheminformatics proxies rather than measured binding.