Uncertainty-Aware Drug-Target Affinity Prediction via Cognitive Memory Retrieval and Attraction-Repulsion Interaction

Effective drug-target affinity (DTA) prediction remains challenged by the limited integration of historical binding knowledge and the lack of reliable confidence estimates. We propose CogNet-DTA, a cognitive-inspired framework that couples an attraction-repulsion mechanism with global memory retrieval for robust affinity estimation. At its core, the Chemical Graph Memory Network (CGMN) utilizes a learnable memory bank to retrieve canonical binding patterns, mimicking expert-driven reasoning. To refine interaction modeling, we implement a dual-pathway head where the attraction pathway captures functional group potentials via fingerprints and evolutionary embeddings, while the repulsion pathway models steric hindrance through drug-graph super-nodes and protein contact maps. This representation is further optimized by spatial-aware attention leveraging distance-weighted contact information. For screening reliability, CogNet-DTA incorporates Monte Carlo Dropout sampling for uncertainty quantization. Evaluations on benchmark datasets demonstrate that CogNet-DTA achieves state-of-the-art performance, particularly in consistency-based metrics such as the r2m, validating that merging cognitive memory with biophysical modeling provides a superior strategy for drug-target interaction analysis.