Stanford researchers develop AI model to map gene-regulating DNA switches across cell types

The human genome contains roughly 20,000 genes, but the real complexity lies not in the genes themselves but in the switches that turn them on and off. These switches—stretches of DNA called enhancers—sit scattered across the genome, sometimes far from the genes they control, and they determine not just whether a gene activates but when it does so and how forcefully. Understanding how enhancers work is central to understanding why some genetic variants cause disease while others do not. The problem is that enhancers are not universal. The same DNA sequence might act as a powerful switch in a liver cell and remain silent in a neuron. This cell-type specificity has made it nearly impossible to build accurate maps of which enhancers control which genes.

For years, researchers have tried to solve this problem computationally. They built models that use measurements of chromatin structure and three-dimensional DNA contacts to predict enhancer-gene interactions, and these models have produced maps spanning hundreds of cell types and tissues. But the validation has lagged far behind. Confirming whether these predictions are correct requires direct experiments, and such experiments have been completed in only a handful of cell types. The field has been working largely in the dark, confident in its methods but uncertain of their accuracy.

A team at Stanford University, led by Maya Sheth and Jesse Engreitz, has now developed a new approach that sidesteps some of these limitations. Published in Nature Genetics, their work introduces scE2G—a set of machine learning models trained to predict enhancer-gene interactions using single-cell data. The key innovation is the training method. Rather than relying on indirect measurements, the researchers used CRISPR experiments to directly test more than 10,000 candidate enhancer-gene pairs. They then fed this ground-truth data into their models, allowing the algorithms to learn the actual patterns that distinguish real regulatory interactions from false positives.

Once trained, scE2G can be applied to single-cell datasets from any tissue or cell type. This matters because single-cell sequencing has become far more common than the bulk tissue methods that earlier models relied on. Researchers have already collected thousands of single-cell datasets across different tissues and conditions. More importantly, single-cell data allows researchers to separate rare cell types computationally—cells that would be impossible to isolate in sufficient quantities for traditional experiments. A tissue might contain only a few hundred cells of a particular type, but if you sequence 100,000 cells from that tissue, you can identify and analyze those rare cells. scE2G can build enhancer maps for these previously inaccessible populations.

The Stanford team tested their models on datasets of varying sizes and sequencing depths, confirming that scE2G performs reliably even when data quality is inconsistent. They then applied the approach to complex tissues to trace disease-associated genetic variants to their target genes. In one example, they linked two genes—INPP4B and IL15—to variation in lymphocyte counts in the blood. This connection would have been extremely difficult to make using noncoding DNA sequences alone, because the variants sit in enhancer regions far from the genes themselves. By mapping which enhancers actually regulate which genes in the relevant cell types, the researchers could bridge that gap.

As single-cell genomics continues to expand, the implications become clearer. The human body contains thousands of distinct cell types, each with its own regulatory landscape. Current enhancer maps cover only a fraction of this diversity. If scE2G and similar models can be applied across the growing archive of single-cell datasets, researchers could eventually chart enhancer-gene regulation across the full spectrum of human cell types. That map would transform how we interpret genetic variants associated with disease, moving from abstract statistical associations to concrete mechanistic understanding: this variant disrupts an enhancer that controls this gene in this cell type, and that disruption causes this phenotype. The work is still in its early stages, but the direction is clear.