Polygenic risk
Evaluating Polygenic Risk Scores at Scale
We test published risk scores in large biobanks and release the results, methods, mappings, and score files.
Fritsche Lab · University of Michigan
We test what genomic and health data can reliably tell us.
We study prediction models built from genomic data, health records, and patient-reported information. Much of our work begins after a model has been built: we test how well it performs in new populations and health systems, and why performance varies.
01Develop
One cohort
02Test
An independent cohort
Independent evaluationTesting models with data that were not used to build them
Longitudinal contextLinking genomics with health records and patient reports
Reusable resultsPublishing methods, score files, mappings, and comparisons
Team scienceWorking with clinical, statistical, and computational colleagues
Research
Do findings hold up in a new setting?
This question connects our work in polygenic risk, biobanks linked to electronic health records (EHRs), mental health, and pain.
Biobanks
Biobanks Linked to Electronic Health Records
We link genomic data with health records collected over time to study disease patterns, treatment, and outcomes.
Mental health
Precision Mental Health
We study mental health treatment response using genetic, clinical, and patient-reported data.
Pain research
Pain Genetics and Phenotyping
We study the genetic basis of pain and define pain phenotypes using clinical records, imaging, and patient-reported data.
Current collaboration · Precision mental health
Why does the same mental health treatment help one person more than another?
COMPASS studies treatment response using genetic, clinical, and patient-reported information. The principal investigators are Amy Bohnert, Srijan Sen, and Lars Fritsche. Lars leads the genetics and EHR data work.
Research resource
Compare published polygenic risk scores across biobanks
PRSweb reports how published scores performed in the Michigan Genomics Initiative and UK Biobank. The three archived releases remain unchanged, so their results and source files continue to match the papers that describe them.
PRSweb Research Portals
Scores, comparisons, and source files in one place
Each release includes performance results, phenome-wide associations, phenotype mappings, methods, and downloadable score files.
Software and teaching
Tools for reproducible research
Software, practical guides, and teaching materials for data analysis, cluster computing, and coding.
PheWAS software
phewasFlow
An R package for reproducible phenome-wide association studies in both directions, with local and cluster workflows, multiple-testing correction, and Manhattan and volcano plots.
Research computing guide
Fritsche Lab SLURM Playbook
The SLURM Playbook covers submitting and monitoring CPU jobs and job arrays on U-M clusters, diagnosing failures, and rerunning only failed tasks.
Research practice guide
Practical GenAI Coding Guide
An eight-step workflow for AI-assisted research code, covering planning, prompting, review, testing, version control, and documentation.
Selected research
Papers from our current and earlier work
2026
Polygenic risk scores for prediction of immune checkpoint inhibitor thyroid toxicity in diverse populations
Clinical Cancer Research
Why it matters: Treatment-related outcomes include adverse effects as well as benefits. In 4,289 veterans treated with checkpoint inhibitors, an updated hypothyroidism polygenic score predicted thyroiditis in both the non-Hispanic White and Black patient groups, but not in chemotherapy controls. The older score, which was derived from European-ancestry data, did not predict thyroiditis in the Black patient group. This shows why scores need to be validated in each population and treatment setting.
2025
Assessing the Clinical Utility of Published Prostate Cancer Polygenic Risk Scores in a Large Biobank Data Set
European Urology Oncology
Why it matters: Researchers tested 16 published prostate cancer scores in the Michigan Genomics Initiative (MGI) and used detailed biopsy data to evaluate their clinical utility. Even the best-performing score separated cases from controls only modestly, and none distinguished aggressive from indolent disease. Predicting a diagnosis is not the same as identifying the cancers that most need treatment.
2024
Improving prediction models of amyotrophic lateral sclerosis (ALS) using polygenic, pre-existing conditions, and survey-based risk scores in the UK Biobank
Journal of Neurology
Why it matters: Genetics, diagnoses, and survey measures did not contribute equally. In UK Biobank, polygenic scores for amyotrophic lateral sclerosis (ALS) offered modest discrimination. Adding diagnoses recorded before onset improved discrimination, while adding the exposure score did not. In this analysis, combining more types of data did not necessarily improve prediction.
People
A team working across disciplines
Our researchers at Michigan Medicine and the School of Public Health bring experience in statistical genetics, clinical research, and research software development.
Lars Fritsche, PhDLab Head
Ryan Welch, PhDApplication Programmer/Analyst Lead
Sarah Fazal, PhDResearch Fellow
Shivalika Pathania, PhDResearch Fellow
Tai Yang, MSGraduate Student Research Assistant (GSRA) and PhD Candidate
Latest
Research stories, talks, and project news
· Research story
U-M researchers aim to shorten the path to effective mental health care
U-M Research explains how COMPASS studies differences in responses to mental health treatment.
· Video
MGI 2024: Lars Fritsche
In this U-M symposium talk, Lars Fritsche reviews a decade of EHR-linked genetics research, including PheWeb and PRSweb.
· Funding
U-M announces NIMH-backed COMPASS precision mental health grant
NIMH awarded $17.9 million to the COMPASS precision mental health collaboration.