FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design

A research team has introduced FuseDiff, a novel AI model for a critical challenge in modern drug discovery: designing a single molecule that can effectively bind to two different disease targets simultaneously. This approach, known as polypharmacology, can lead to therapies with improved efficacy and reduced drug resistance, but it requires generating a ligand with two distinct, target-specific 3D binding poses. Existing methods typically use staged pipelines that either treat the poses as independent or enforce overly rigid correlations, failing to capture the complex joint relationship.

FuseDiff addresses this with an end-to-end diffusion model that jointly generates the ligand's molecular graph and its two binding poses, conditioned on the 3D structures of both target protein pockets. Its core innovation is the Dual-target Local Context Fusion (DLCF) module within a message-passing neural network backbone. DLCF fuses local chemical context from both pockets for each atom, allowing expressive joint modeling while preserving the necessary symmetries of the problem. This ensures topological consistency across both poses under a shared molecular graph while permitting target-specific geometric adaptations.

The team created a dedicated dual-target training dataset and conducted experiments on a benchmark and a real-world system. Results show FuseDiff achieves state-of-the-art performance in docking metrics. Crucially, it enables the first systematic, AI-driven assessment of dual-target pose quality *before* running computationally expensive docking simulations, potentially streamlining the early drug design pipeline. This represents a significant step toward more rational and efficient design of multi-target therapeutics.