SymfuseX: A framework bridging drug–target predictive modeling and target-specific molecular design via a symbiotic fusion mechanism

Drug discovery requires learning from heterogeneous biochemical information while supporting both predictive reasoning and design-oriented molecular generation. Existing drug–target interaction and affinity models can achieve strong predictive performance, but the learned cross-modal interaction knowledge is typically optimized only for scoring and is rarely transformed into a reusable target-conditioned representation for downstream molecular design. To address this limitation, we propose SymfuseX, a unified framework built around a Symbiotic Fusion Mechanism that bridges drug–target prediction and target-specific molecular design. It integrates bidirectional conditional modulation, multi-path factorized fusion, and adaptive path weighting to model reciprocal preferences between drug and protein representations and to capture complementary interaction evidence from both raw and modulated features. Through this mechanism, SymfuseX not only improves cross-modal biochemical interaction modeling, but also explicitly learns a transferable target-conditioned latent representation that can be reused beyond predictive scoring. SymfuseX integrates graph-based drug structure, sequence-based protein context, and rule-derived chemical prior features within a unified predictive-to-generative pipeline. Experimental results on benchmark DTI and DTA datasets show that SymfuseX achieves strong and stable predictive performance across multiple evaluation settings. In addition, the transferred target-conditioned latent representation supports competitive target- specific molecular generation and facilitates the identification of plausible lead compounds with favorable drug-likeness and synthetic accessibility.