Combination therapies are widely used to treat diseases, from hypertension to cancer and infectious diseases. Computational models typically use structural data or target profiles to identify drug combinations with high efficacy and low toxicity, yet often overlook complex biological interactions. Addressing these gaps requires integrating diverse modalities—such as structural data, pathways, cell viability, and transcriptomics—to capture the complexity of clinical drug combinations from multimodal preclinical data.

Here, we introduce Madrigal, a unified multimodal AI model for predicting drug combinations across 953 clinical outcomes and 21,842 compounds, including approved drugs and novel compounds in development. Madrigal uses a bottleneck module with modality alignment to transfer information across modality-specific encoders and address the all-stage missing modality challenge, a problem that plagues multimodal learning in medicine.

Madrigal outperforms single-modality methods and state-of-the-art models in predicting adverse drug interactions. We demonstrate Madrigal for identifying transporter-mediated drug interactions, drug combination virtual screening of anti-cancer combinations, and identifying polypharmacy needs for type II diabetes and metabolic dysfunction-associated steatohepatitis (MASH). Madrigal ranks resmetirom, the first and only FDA-approved drug for metabolic dysfunction-associated steatohepatitis (MASH), among the safest candidates. Madrigal applies to individualized cancer therapies, as demonstrated in primary acute myeloid leukemia samples and patient-derived xenograft models, where it predicts personalized drug combination efficacy based on genomic profiles. By integrating Madrigal with a large language model, users can describe clinical outcomes in natural language. Madrigal offers a multimodal approach to designing safe combination therapies.