Madrigal

Combination therapies dominate the treatment of cancer, metabolic disease, and infection, but every added agent multiplies the chance of an adverse drug-drug interaction, and the number of possible pairs grows far faster than any trial programme can test. The decision that matters happens early, when a compound is still preclinical and the clinical safety data that would settle the question does not yet exist. Models built on molecular structure alone are cheap to apply at that stage but miss pharmacodynamic effects that structure does not encode.

Madrigal, from the Zitnik Lab at Harvard Medical School with AstraZeneca, the Broad Institute, Dana-Farber, and Carnegie Mellon, predicts drug-combination outcomes by fusing four preclinical views of each compound: molecular structure, position in a biomedical pathway knowledge graph, cell viability across a screening panel, and transcriptomic response to perturbation. It covers 953 clinical outcomes across 21,842 compounds, spanning approved drugs and agents still in development.

Its defining design problem is missing modalities. The compounds that most need a prediction — novel, early-stage, sparsely characterized — are exactly the ones lacking pathway annotations or perturbation screens, and methods assuming complete profiles discard them. Madrigal handles absent modalities at training and inference by aligning every view to the one that is always available.

#Key Features

• Structure-anchored contrastive alignment: An InfoNCE objective pulls pathway, viability, and transcriptomic embeddings toward the structural embedding, so an observed modality can stand in for a missing one at inference.
• Attention bottleneck fusion: Bottleneck tokens inserted between the modality embeddings regulate cross-modal information flow and keep the many transcriptomic cell lines from swamping the sparser modalities.
• Robustness to incomplete profiles: Predictions degrade gracefully when modalities are absent, which is the operating regime for compounds in preclinical or early clinical development.
• Hard generalization splits: Evaluation holds out entire drugs by shared target profile and by anatomical therapeutic class, not just unseen pairs, mirroring a novel compound paired with an approved partner.
• Patient-level extension: Drug embeddings combine with clinical covariates and genomic profiles to rank combinations for an individual rather than a population.

#Technical Details

Each modality gets a purpose-built encoder: a Graph Isomorphism Network over RDKit molecular graphs, a Heterogeneous Graph Transformer over the PrimeKG precision medicine knowledge graph with drug-drug and drug-phenotype edges removed to prevent leakage, an autoencoder over 559-dimensional PRISM cell-viability vectors, and chemCPA encoders — a chemical extension of CPA — over LINCS Connectivity Map profiles, with each of 16 cell lines treated as its own modality and represented by a 1,956-dimensional vector of landmark-gene changes at two timepoints. Training runs in three stages: modality-specific pretext tasks, contrastive alignment, then fine-tuning of the encoders, the attention bottleneck fusion module, and the prediction head.

Supervision comes from TWOSIDES (4,656,138 combinations, 1,457 drugs, 795 outcomes, derived from FAERS reports) and DrugBank (1,188,371 combinations, 3,632 drugs, 158 expert-curated outcomes). Under the most stringent split — held-out drugs sharing minimal targets with training drugs — Madrigal reached AUROC 0.789 and AUPRC 0.640 on TWOSIDES and AUROC 0.836 and AUPRC 0.772 on DrugBank, improving on structure-based baselines (DeepDDI, CASTER, GMPNN-CS) by 10.7% and 6.2% on average across metrics, and on multimodal knowledge-graph models (MUFFIN, TIGER) by 22.5% and 12.8% AUROC. GMPNN-CS matched its DrugBank AUROC to within 0.001 in that split. Ablations confirm that both the multimodality and the alignment step contribute beyond structure alone.

#Applications

Madrigal is aimed at combination triage in drug development: ranking which pairings of a new agent with an approved partner carry elevated risk of specific adverse events, before a trial commits patients. It recovers transporter-mediated interactions, and its risk ordering agrees with observed differences in alopecia, anemia, hypoglycemia, and neutropenia across 35 head-to-head trial arms. In metabolic disease it ranks resmetirom, the first FDA-approved MASH drug, second among candidate partners for type 2 diabetes therapies. Personalized use is demonstrated on primary acute myeloid leukemia samples, patient-derived xenografts, a longitudinal EHR cohort, and an oncology cohort where its embeddings reached 0.68 average AUROC across 13 outcomes. Where PKAG-DDI generates free-text descriptions of an interaction, Madrigal scores risk across a fixed outcome vocabulary.

#Impact

Madrigal makes a case that perturbation readouts already generated at scale in pharma — viability panels and transcriptomic screens — carry clinically predictive signal about combination safety that structure-based models miss, and that the missing-modality problem blocking their use is tractable. Code is released under an MIT license with pretrained checkpoints on Hugging Face and benchmark splits on Harvard Dataverse, making the comparisons reproducible. The weights are CC BY 4.0, though easy to miss — the model card declares it in a trailing YAML block Hugging Face does not parse, so no license badge appears. Its ceiling is set by its labels: TWOSIDES outcomes derive from spontaneous adverse-event reports and DrugBank outcomes from text extraction, so real risks absent from those vocabularies — the authors cite tropifexor's dose-related pruritus and firsocostat-associated hypertriglyceridemia — cannot be predicted at all. The work remains a preprint awaiting peer review.