MIRAGE-DTA: conflict-aware multimodal evidence arbitration for drug-target affinity prediction.

MIRAGE-DTA improves calibrated drug-target affinity prediction under missing evidence, cutting expected calibration error by 0.0524–0.1018 on CHEMBL_ASSAY splits with up to 35% sequence/text missingness.

• Why it matters: Multimodal DTA models struggle when modalities are missing, entities are cold-start, or assay distributions shift, making evidence reliability inconsistent and calibration degrade.
• What they did: The method adds a reliability-oriented evidence-arbitration layer that maintains two retrieval chains (current-sample and training-set), uses modality masks in fusion, and resolves disagreement via an interaction gate with a validation-selected anchor.
• The result: On 10-seed paired evaluation across BindingDB_Kd, DAVIS, KIBA, and a 2010–2024 CHEMBL_ASSAY benchmark (35,238 interactions), the two-branch model lowers RMSE and yields significantly lower risk at 80% coverage on target-cold and temporal splits.

AI summary of the abstract · may contain errors · not medical advice · read the paper.