Machine learning discovery of variants that alter gene regulation in cardiovascular development and disease

The National Heart, Lung, and Blood Institute (NHLBI), part of the National Institutes of Health within the Department of Health and Human Services, awarded J. David Gladstone Institutes $925,796 on August 1, 2026, under the Cardiovascular Diseases Research program (CFDA 93.837) to develop machine learning methods for identifying noncoding genetic variants that cause congenital heart disease.

Gladstone will create a deep learning toolkit to predict how rare and de novo noncoding variants—including single nucleotide changes and large structural variants—alter gene regulation during cardiovascular development. The research addresses the gap that while genome sequencing has identified damaging variants in heart development genes, the majority of children with congenital heart disease lack a genetic diagnosis. Each project aim develops a machine learning model targeting one genetic mechanism: three-dimensional chromatin interactions, local nucleosome positioning, or transcription factor binding sites. The methods will predict candidate causal variants and generate mechanistic hypotheses about how high-scoring variants disrupt gene regulation, while identifying target genes, upstream regulators, and affected cell types.

The award is structured as a Project Grant, the standard assistance type for NIH research funding. Work is performed in San Francisco, California, with a period of performance through April 30, 2030.

Abstract

Congenital heart disease (CHD) affects ~1% of live births in the US and runs in families. Genome sequencing of children with CHD has identified damaging variants in dozens of heart development genes. These are enriched for transcription factors and chromatin regulators. Despite this progress, the majority of children with CHD lack a genetic diagnosis, and the full spectrum of causal pathways remains unknown. Emerging evidence suggests a role for rare and DE novo noncoding variants. Unfortunately, IT is challenging to guess which noncoding sequence changes will perturb cardiovascular development and via what mechanisms. This limits our ability to test candidate causal variants outside of protein-coding genes, despite possessing the necessary tools for engineering patient alleles into human cells. To overcome these obstacles, we propose to create a deep learning toolkit for heart development that can predict diverse functional effects of noncoding variants. We will apply these methods to different types of rare and DE novo variants detected in the genomes of children with CHD, ranging from single nucleotide changes to large structural variants. Our overarching hypothesis is that some rare and DE novo noncoding variants cause CHD by changing three-dimensional chromatin interactions, local nucleosome positioning, and/or transcription factor binding sites. Each aim of this project develops a machine learning model for one of these genetic mechanisms, enabling US to predict candidate causal variants along with a hypothesis about how each high-scoring variant alters gene regulation in heart development. Importantly, our methods will identify candidate target genes, upstream regulators, and cell types for these variants, which would be impossible using standard bioinformatics analyses or genetic association tests. The top hits for cardiomyocytes will be engineered into isogenic cell lines and evaluated for the predicted effects using advanced genomics techniques, including HI-C, single-molecule nucleosome occupancy, and massively parallel reporter assays, plus RNA sequencing. We anticipate that this machine-guided strategy will reveal new genetic mechanisms, causal variants, and CHD associated loci. Our trained models and code will be freely shared to enable others to predict the effects of sequence variation on gene regulation in heart development.