DATR: DDI-Aware Therapeutic Structure Reconstruction for Safer Medication Recommendation

ICLR 2026 Conference Submission12630 Authors

Keywords: Medication Recommendation, Drug-Drug interaction, Drug Structure

TL;DR: We propose DATR, a medication recommendation framework that jointly optimizes therapeutic accuracy and drug safety by therapeutic-aware structure reconstruction and proactive DDI avoidance.

Abstract: Medication recommendation systems play a critical role in clinical decision support, where ensuring both predicting accuracy and safety, particularly drug-drug interaction (DDI) avoidance, is essential. While recent studies have explored drug molecular structures to enhance accuracy, they often overlook the semantic gap between chemical structures and therapeutic outcomes, leading to suboptimal recommendation. Moreover, existing DDI mitigation strategies typically operate in a post-hoc manner, limiting their ability to proactively prevent adverse interactions. In this work, we propose DDI-Aware therapeutic Structure Reconstruction (DATR), a novel framework that jointly models drug structures, therapeutic intent, and safety profiles. DATR conditionally encodes drug structures based on ATC-derived therapeutic labels, enabling intent-aware representation learning, and introduces a selectivity potential DDI constraint to proactively reduce interaction risk. Experiments on two real-world datasets and evaluations by clinical experts demonstrate that DATR achieves superior performance in recommendation accuracy and DDI reduction. Code is available at https://anonymous.4open.science/r/DATR-7EA8.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 12630