Most genetic differences linked to disease do not occur within genes themselves, but instead in the vast stretches of non-coding DNA that regulate how genes turn on and off. Interpreting the effects of these variants has remained one of the major challenges in genomics and precision medicine because experimentally testing every possible DNA change across the human genome is not feasible.

In this study, researchers developed MPAC, a machine learning framework trained on data from massively parallel reporter assays (MPRAs), powerful experiments that simultaneously measure how thousands of DNA variants alter gene regulation. By learning patterns from these large-scale datasets, MPAC can predict the regulatory impact of non-coding variants at unprecedented scale. The team applied the model to all known human variants - 575M in all, spanning inherited variation, disease-associated mutations, cancer genomes, and evolutionary datasets.

The results demonstrate that MPAC can accurately identify functional variants linked to complex traits and diseases, detect potentially pathogenic non-coding mutations in clinical databases, flag candidate cancer-driving mutations, and reveal how evolutionary forces have shaped regulatory regions of the genome.

For scientists and clinicians, this work represents an important step toward making sense of the non-coding genome, transforming vast amounts of genetic variation into actionable biological insight. Yale Center for Research Computing enables this type of discovery by providing the computational infrastructure needed to train large-scale machine learning models and analyze hundreds of millions of genomic variants across diverse datasets.