Supplementary Materials for CoMed

Adverse drug events (ADEs) from combination therapies remain a global challenge, particularly in low- and middle-income countries (LMICs) where clinical decision support infrastructure is often limited. While large language models (LLMs) show promise for biomedical understanding, their black-box nature and lack of contextual reasoning limit their utility in high-stakes clinical applications. Here, we present CoMed, an interpretable, LLM-based framework for combination medication risk profiling. CoMed combines chain-of-thought prompting with a modular multi-agent architecture to support clinically grounded and transparent risk assessments. Across a curated benchmark of 1,482 biomedical abstracts involving 95 drug pairs, CoMed-enhanced models significantly outperformed baseline LLMs in identifying co-administration scenarios (F1 up to 0.971, AUPRC 0.974). Beyond identification, CoMed synthesizes literature into multi-dimensional risk reports covering adverse effects, efficacy, indications, selectivity, and management considerations. Outputs are fully traceable to source evidence and designed for frontline clinical utility. Our results demonstrate that CoMed enables accurate, interpretable, and scalable analysis of combination drug risks, offering a practical tool to enhance medication safety in resource-constrained healthcare settings . To support broader adoption, we have released an open-source python package of CoMed, including documentation and full installation instructions, available at https://github.com/studentiz/comed.