Patient centered prediction of clinically important outcomes arising from pathogenic variants

Federal Cooperative Agreement Award Summary

Vanderbilt University Medical Center received a $3.07 million Cooperative Agreement awarded on August 15, 2025, under the National Human Genome Research Institute's Human Genome Research program (CFDA 93.172) to develop a machine learning/artificial intelligence (ML/AI) tool that predicts clinical outcomes in individuals carrying pathogenic genetic variants. The project, which runs through July 31, 2027, addresses the clinical challenge of incomplete penetrance and variable expressivity in genetic disease by leveraging longitudinal patient data from electronic health records (EHR) and biobanks. The awardee will create computable phenotypes that accurately represent clinical trajectories in pathogenic variant carriers, model factors affecting variant expressivity, and employ Bayesian transfer learning methodologies to synthesize multimodal data for generating individualized risk predictions of key clinical outcomes.

The deliverables include refined understanding of disease prediction mechanisms, identification of early phenotypic indicators of disease progression through longitudinal patient trajectory analysis, and a knowledge-guided ML/AI framework designed for clinical translation. This work directly supports the National Human Genome Research Institute's mission to advance the application of genomics to medical science and clinical care by providing practitioners with evidence-based decision-support tools to guide clinical interventions for individuals with uncertain health impacts from discovered pathogenic variants.

Abstract

As genetic testing becomes more common in medical care, an increasing number of individuals will discover that they carry a pathogenic variant. For some, this result will explain current symptoms and aid in diagnosis. For others, the future health impacts of a pathogenic variant may be less clear, causing uncertainty but also presenting an opportunity to prevent or mitigate serious health outcomes in the future. This uncertainty arises from incomplete penetrance and variable expressivity. Whether a symptom manifests in a carrier depends on numerous individual factors (only some of which are known), creating a dilemma for practitioners regarding whether and how to intervene. Using what we can learn from past clinical experience in variant carrying individuals captured in a growing number of resources like biobanks, we can address this clinical dilemma by creating a machine learning / artificial intelligence (ML/AI) tool that predicts the likelihood a pathogenic variant carrier will develop disease. Building on our extensive experience in creating computable and portable phenotypes from data in the electronic health record (EHR) that accurately represent clinical trajectories in pathogenic variant carriers, we propose to refine our understanding of disease prediction by modeling factors affecting variant expressivity and using longitudinal patient trajectories to identify early, often subtle, phenotypic indicators of disease progression. Finally, we will use a bayesian transfer learning approach to synthesize multimodal data for generating individualized predictions of risk of key clinical outcomes in the context of a given pathogenic variant. To develop a viable model for clinical translation that addresses the significant risks and challenges associated with developing ML/AI tools, we will employ a knowledge-guided framework that incorporates input from ethical, legal and social implications (elsi) experts, clinicians, statisticians, geneticists, and informaticians at every stage of the design process, from defining key clinical outcomes, selecting and engineering model inputs, and developing approaches to communicate predictions to patients and providers. Our framework will allow US to synthesize current knowledge of pathogenic variants with the patterns mined from real-world data while addressing the significant elsi concerns inherent in ML/AI tool development. Our proposal brings together a transdisciplinary team of experts in informatics, ethics, machine learning, cloud computing, genomics, and clinical medicine, and leverages exceptional local data resources to bring the potential of ML/AI genomic medicine. With this innovative proposal, we aim to create resources that will enable the development and validation of valuable genomic medicine tools for the future.