Fused Bayesian Flow Networks Advance Dual-Target Drug Design

Dual-target drug design is a promising strategy for discovering polypharmacological compounds to treat complex diseases, aiming to create 3D molecules that can bind to two target proteins simultaneously. While recent generative models have shown success in single-target drug design, current dual-target methods often fall short, either focusing on sequence generation or adding predictive drift terms to diffusion models, which limits their ability to fully integrate information from both targets. A new approach, FusedBFN (Fused Bayesian Flow Network), has been developed to overcome these limitations. FusedBFN frames dual-target generation as a distribution fusion problem within a continuous parameter space. It employs a product-of-experts formulation to seamlessly incorporate dual-target information throughout the entire generative process, ensuring comprehensive consideration of both targets. To address the common issue of scarce dual-target structural data, FusedBFN utilizes a pretrained target-aware BFN model as its shared backbone. It also introduces chemically aware prior-based alignment and prior-free pocket alignment strategies to construct accurately aligned dual-target contexts. Extensive experiments confirm that FusedBFN generates molecules with strong binding affinity to both targets while maintaining desirable molecular properties, marking a significant step forward in drug discovery.