A new AI model predicts how dangerous or effective a drug combination will be in patients, using only lab-scale data collected before anyone runs a clinical trial.
The model, called Madrigal, learns from four types of preclinical data: molecular structure, biological pathways, cell-viability tests, and gene-expression readouts. It aligns these signals for 21,842 compounds into a shared representation, then trains on 158 expert-curated and 795 patient-reported combination outcomes. In testing, it beat models that rely on a single data type or on other multimodal approaches, and it flagged elevated danger for combinations that share the same membrane transporters. Across 28 head-to-head trial comparisons, it picked the arm with more reported neutropenia, anemia, hair loss, or low blood sugar as the riskier one in 25 cases.
Most drug-interaction models still lean almost entirely on chemical structure and known drug targets, which misses how a compound actually behaves inside a cell. Folding in cell-viability and transcriptomic data lets Madrigal catch combination risks that structure alone would not flag, and do it before a single patient is enrolled. That is the expensive part of drug development that keeps producing late-stage failures.
Still, this is a retrospective analysis run on existing cohorts, not a live clinical decision tool, and getting the risk ranking right in 25 of 28 comparisons leaves real room for the misses that matter most.