From genomes to decisions

Mechanistic clinical trials

Clinical trials answer whether a treatment works. We design trials that also reveal how it works and for whom.

We embed genomic, proteomic, metabolomic, microbiome, and other omics profiling into clinical trials run within Massachusetts General Hospital and the Harvard-affiliated hospitals. Serial multi-omic measurements before and during treatment allow us to trace the molecular pathways through which an intervention acts, to identify responders and non-responders, and to discover biomarkers that can guide treatment selection.

Omics also make trials leaner. Omics-based enrichment selects participants whose biology makes a treatment effect most likely, and omics-based readouts register that effect at the molecular level long before clinical events accrue. Together these allow efficient trials with smaller sample sizes and shorter follow-up than conventional event-driven designs.

• Omics-augmented trial design, in which sample collection and molecular profiling are built into the protocol from the outset.
• Omics-enriched recruitment and molecular endpoints that shrink trial size and duration.
• Mechanistic studies of metabolic therapies, including the effects of the GLP-1 pathway on biological and organ-specific aging beyond weight loss.

Agentic biology

We use AI agents as collaborators in scientific discovery, not only as tools.

Large language models and agentic systems can now plan analyses, write and execute code, read the literature, and interpret results. We build and evaluate such agents for biomedical discovery, with the goal of compressing the cycle from question to validated finding. Several studies are underway in partnership with Google DeepMind and other collaborators.

• Agents that autonomously conduct genetic and multi-omic analyses across large biobanks and report reproducible findings.
• Benchmarks and evaluation frameworks that measure whether AI-generated biological hypotheses are correct, novel, and testable.
• Human-in-the-loop workflows that pair agents with clinicians and experimental biologists for target validation.

Drug development

Drugs with human genetic support are far more likely to succeed. We use genetics and AI to find the next targets and to anticipate their effects.

Our target discovery pipeline combines rare and common variant association from whole genome sequencing with multi-omic integration, Mendelian randomization, and machine learning to prioritize genes and pathways for cardiometabolic disease. Biological digital twins, computational models of an individual's molecular state, allow us to simulate the consequences of perturbing a target before an experiment is run. Prioritized targets are then validated with experimental collaborators and, where appropriate, carried toward AI-assisted drug design.

• Integrative multi-omic and multi-trait prioritization of therapeutic targets for dyslipidemia, obesity, and coronary artery disease.
• Genetic dissection of disease subtypes, such as obesity endotypes with distinct biology, to match targets to the patients most likely to benefit.
• Genetic safety assessment of existing drug targets, including on-target adverse effects revealed by human genetics.

Clinical informatics for precision medicine

A genome is most useful when it is read together with everything else we know about a patient.

We develop risk models that integrate germline and somatic genetics, plasma proteomics, and metabolomics with electronic health records, laboratory values, and imaging. These multimodal models improve prediction, diagnosis, and treatment selection across the spectrum of cardiometabolic conditions beyond what clinical scores alone can achieve, and they identify the individuals for whom early intervention has the greatest benefit. Our goal is to deploy such models within the health system so that they inform real clinical decisions.

• Machine learning-based polygenic and proteomic risk scores for cardiometabolic disease.
• Integration of genetic risk into EHR-based prevention pathways and treatment thresholds.
• Multimodal foundation models that learn jointly from omics, imaging, and clinical text.

Precision public health

Public health has always been population-level medicine. We want population genetics to become part of how it is practiced.

Drawing on the principal investigator's experience as a public health official and as a Lead Collaborator of the Global Burden of Disease Network, this program brings genetic and molecular evidence into the frameworks that governments and global health agencies use to set priorities. We collaborate closely with the Institute for Health Metrics and Evaluation, home of the Global Burden of Disease study, to incorporate genetic risk factors into burden estimates, and we use Mendelian randomization and large-scale epidemiology to establish which risk factors are causal and therefore actionable.

• Genetically informed estimates of disease burden and attributable risk across populations and ancestries.
• Causal evidence for modifiable risk factors, such as adiposity, across cardiovascular, metabolic, and gastrointestinal outcomes.
• Proteomic readouts of subclinical disease trajectory that measure how far, and how fast, individuals and populations are moving toward disease, so that prevention can begin before a diagnosis is made.