Decoding Molecular Binding Equilibria through Uncertainty-Aware Memory Retrieval

While accurate drug-target affinity prediction accelerates drug discovery, prevailing deep learning approaches are often bottlenecked by poor generalization and an inability to quantify prediction confidence. Here, we introduce CogNet-DTA, a novel memory-augmented framework that synergizes historical binding priors with biophysical constraints. Our approach introduces a Chemical Graph Memory Network that facilitates reasoning by analogy through retrieving canonical binding motifs from historical data. To capture precise structural contexts, we leverage ESM-based evolutionary embeddings combined with Contact-Weighted Attention, which injects ESM-derived residue-residue contact probabilities as spatial biases to refine protein-intrinsic structural context. Distinctively, CogNet-DTA utilizes a Chemo-Geometric Routing Module to organize multimodal inputs into positive and negative predictive pathways under a shared memory prior, providing a conceptual, data-driven decomposition of molecular binding affinity. By integrating Monte Carlo Dropout, our model systematically quantifies uncertainty, providing a critical mechanism to filter out low-confidence, high-risk predictions during virtual screening. Extensive evaluations across diverse benchmark datasets demonstrate that CogNet-DTA achieves highly competitive performance in DTA prediction. Notably, it exhibits robust generalization to novel drugs and targets under strict cold-start and homology-based split settings, successfully mitigating sequence similarity bias.