Bayesian Statistics · Human Genetics · Prevention

Principled probabilistic models of human disease, from genome to clinic.

We combine the genome and the health record to estimate who is at risk and when, to find the genetic causes of disease, and to learn what actually changes its course.

Based at Massachusetts General Hospital, the Broad Institute, and Harvard.
What we work on

Prevention, discovery, cause.

Three questions, one Bayesian generative model of lifetime disease — Aladynoulli — applied to hundreds of thousands of individuals in the UK Biobank, Mass General Brigham, and All of Us.

1

Bayesian disease prevention

Calibrated, individual lifetime risk that combines polygenic scores, monogenic variants, and the clinical record — estimating not just whether disease will come but when, so screening and prevention can be timed to the person.

2

Genetic discovery

Sharing strength across conditions, tissues, and time to find what inherited variation does. Genes don't just tilt the odds — they reshape which diseases cluster and when they arrive, for common polygenic risk and rare pathogenic variants alike.

3

Causal connection

The genome is fixed at birth — a natural experiment upstream of everything in the record. We use genetic anchors to connect drug targets to disease, and individual disease signatures to ask who benefits from which therapy, with negative controls built in.

The approach

From genome and record to a readable model of a life.

Priors, posteriors, and borrowing strength — a Bayesian generative model fit at biobank scale, built so that every parameter means something a clinician and a geneticist can both interrogate.

Inputs

Genome + record

The genome and the longitudinal stream of diagnoses, labs, and timing.

Structure

Latent signatures

A small set of interpretable disease processes, shared across people and time.

Dynamics

Time-varying effects

Genetics warps the rate and order in which signatures emerge over a lifetime.

Output

Transferable model

Calibrated trajectories that move to a new health system — the model, not just the weights.

Selected work

The lineage of the method.

One idea — borrow strength across what looks unrelated, and keep it interpretable — at four scales: tissues, conditions, individuals, time.

Explore the Aladynoulli model & figures ↗
Tutorials & resources

Learn how the model thinks.

Everything is open: the framework, the figures, and the code. Start here whether you want to understand the method, reproduce a result, or build on it.

About

The person building it.

SU

Sarah Urbut, MD, PhD

Instructor in Medicine, Harvard Medical School · Massachusetts General Hospital

Sarah is a physician-scientist who builds parsimonious, interpretable statistical methods to decipher patterns in large-scale genomic and clinical data — working at the seam between statistics and genomics, science and medicine.

Her path runs from a Princeton thesis on adaptive false-discovery methods, through an MD–PhD at the University of Chicago — where, with Matthew Stephens, she developed mash, now a widely used tool for sharing evidence across conditions — to cardiology and the build-out of Aladynoulli. The same idea recurs at every scale she works on: borrow strength across what looks unrelated — tissues, conditions, individuals, time — and never lose the ability to say what each parameter means.

She views probabilistic modeling as a way to narrate the complex patient stories she encounters in the lab and the clinic. An avid cyclist who commutes between MGH, the Broad, and Cambridge — and a loyal, often heartbroken, Chicago White Sox fan.

View full CV →

In close collaboration with Giovanni Parmigiani (Harvard / Dana-Farber, Bayesian methodology), Alexander Gusev (Dana-Farber, computational genetics), and colleagues across MGH and the Broad Institute.
Supported by
NIH / NHLBIK08 Mentored Clinical Scientist Research Career Development Award
Burroughs Wellcome FundCareer Award for Medical Scientists
American Heart AssociationCareer Development Award

Join the lab.

We welcome students, postdocs, and clinicians interested in Bayesian methods, statistical genetics, and disease prevention. Backgrounds in statistics, computer science, genetics, or medicine are all a good fit.

ML / CS researchers

Generative models, probabilistic machine learning, and scalable inference for large health data.

Statisticians & geneticists

Bayesian methods, longitudinal and survival data, genetic discovery, and causal inference.

Quantitative clinicians

Physicians with quantitative training who want to bring clinical questions to large-scale data.

Email Sarah →