Machine Learning Penetrance Score Predicts Genetic Variant Impacts

September 2, 2025

By Bio-IT World Staff

September 2, 2025 | Researchers at the Icahn School of Medicine at Mount Sinai have published a machine learning model to predict penetrance of genetic variants—how likely the variants are to cause disease. The tool combines genomic and clinical phenotype data to evaluate penetrance at scale, offering refined, individualized disease risk estimates. The study was published last week in Science (DOI: 10.1126/science.adm7066).

Penetrance is hard to gauge and has historically been determined based on literature review, in silico prediction tools, segregation analysis, population-based allele frequency, and expert analysis.

“The American College of Medical Genetics and Genomics (ACMG) has published guidelines to help standardize the variant interpretation process,” the authors write in the paper. “Yet the absence of large accurate truth sets of variant disease risk and reliance on expert review that can be time-consuming and laborious result in sometimes discordant classifications, false positives, and uncertainty, especially for diseases associated with incomplete and age-dependent penetrance.”

Current approaches, including in silico tools such as AlphaMissense, don’t enrich predictions with real-world clinical data of disease manifestations associated with specific variants within a patient population. Instead they predict pathogenicity based on protein structure, sequence conservation, or molecular interactions.

But the authors believe that adding phenotypic data from electronic health records will help flesh out the variant data with physician notes, test results, diagnosis codes and other data. Creating a model that includes EHR data also allows for the inclusion of more diverse patient data. “The availability of EHRs from unrelated individuals unselected for specific diseases allows for a less biased assessment of disease risk associated with genetic variants, supporting more accurate variant interpretation and risk assessment,” the authors write.

ML penetrance, the measurement they developed, comes from machine learning models trained on clinical data, designed to quantify an individual’s risk and trajectory of disease given their genetic variant. ML penetrance is more useful than conventional penetrance, the authors argue, because it incorporates clinical factors specific to each individual into its assessment and leverages continuous disease scores instead of conventional binary case-versus-control classifications.

Introducing ML Penetrance

The authors previously tested this approach by creating an in silico score for coronary artery disease (ISCAD), a single disease score that combines clinical and genetic risk, atherosclerosis, disease sequela, and mortality. They posit that the same approach would characterize the penetrance of genetic variants as a holistic measure capturing genetic, clinical, lifestyle, and other factors.

To build the measure, the researchers constructed machine learning (ML) models for 10 autosomal dominant diseases that are incompletely penetrant. The diseases they included were arrhythmogenic right ventricular cardiomyopathy (ARVC), familial breast cancer (FBC), familial hypercholesterolemia (FH), hypertrophic cardiomyopathy (HCM), adult hypophosphatasia (HPP), long QT syndrome (LQTS), Lynch syndrome (LS), monogenic diabetes (MD), polycystic kidney disease (PKD), and von Willebrand disease (VWD). Within these 10 diseases, the team evaluated ML penetrance of 1,648 rare variants in 31 autosomal dominant disease-predisposition genes.  

Using diagnoses from clinical guidelines, they identified cases and controls in three distinct cohorts of 1,376,251 participants. They trained and internally validated ML models to predict the diseases in one cohort, tested them in a second cohort, and applied them to a third cohort with linked exome sequence data to obtain ML penetrance estimates.

The ML penetrance measure was validated by comparing its values for ClinVar-classified pathogenic (P) versus benign (B) variants and rare versus common variants, assessing its association with disease-relevant clinical outcomes in heterozygous individuals, and corroborating it with multiple lines of experimentally measured functional data. The team also compared ML penetrance with the conventional method of calculating penetrance using binary case-versus-control phenotypes, diagnosis-based (DX) penetrance.

“Notably, ML penetrance correlated with disease-relevant clinical outcomes, such as risk of end-stage renal disease for PKD variants and heart failure for HCM variants. ML penetrance also aligned with experimentally derived measures of variant function, reinforcing its biological relevance,” the authors report in their article summary.

“Importantly, ML penetrance aided in the evaluation of VUS and previously unknown LoF variants by delineating clinical trajectories—individuals with highly penetrant variants showed perturbed vital signs, electrocardiogram measures, and disease biomarkers over time. The researchers were also able to classify previous variants of unknown significance (VUS) and loss of function (LoF) variants and prioritize those that alter the clinical profiles of carriers for detailed analysis.”

Impacts and Future Value

The authors see myriad valuable impacts of ML penetrance. “ML penetrance could be used for early disease detection and screening, identifying high-risk individuals before symptoms appear or prioritizing VUS for further investigation … It has the potential to improve personalized disease risk assessments by genetic counselors and tailor monitoring and preventive care, particularly for conditions with gradual onset such as cancer or cardiomyopathy. Additionally, ML penetrance could aid in identifying candidates for clinical trials, refining population health initiatives, and ensuring efficient resource allocation within health systems,” they write.

The tool is not meant to be definitive, the authors caution. Rather, ML penetrance is meant to be one of several complementary approaches—functional mechanistic studies and prospective follow-up—to further assess the reliability and confidence of our predictions, determine causality, and inform clinical decision-making.

As always, they note, the reliability of the machine learning model depends on the quality of the EHR data on which it is trained. They also highlight the need for future studies on additional diverse populations and non-European groups, and for future studies to consider other types of genetic conditions, different modes of inheritance, and a wider range of genes.”

But this is a valuable first step. “This study helps lay the groundwork for ML-driven genetic models and tools to be developed and added to the modern geneticist’s armamentarium,” the researchers write.