For years, mapping the three-dimensional structure of the genome—the tangled, folded mass of DNA that determines which genes are active in a given cell—has been an expensive, time-consuming process. However, a new generative AI model developed by MIT chemists could make that approach obsolete.

Their model, called ChromoGen, can predict thousands of chromatin structures in minutes, a task that would take weeks using current experimental techniques. The researchers detailed their findings in Science Advances, highlighting how this AI-driven approach could unlock new insights into everything from gene regulation to disease mechanisms.

The 3D Genome Problem

If DNA were just a simple code, understanding gene expression would be easy. Instead, inside every cell nucleus, DNA doesn’t just lie flat—it folds into complex structures that dictate which genes can be accessed and turned on. The same genetic blueprint produces neurons, liver cells, and skin cells, all because of how their DNA is packaged.

For the past two decades, scientists have relied on experimental methods like Hi-C, which chemically links nearby segments of DNA, allowing researchers to infer their spatial arrangement. The problem? It’s slow, expensive, and typically yields an average structure for a group of cells rather than capturing the full diversity of genome configurations.

This is where AI comes in.

What MIT’s AI Model Does Differently

Bin Zhang, an MIT associate professor of chemistry, and his team built ChromoGen to sidestep these experimental limitations. Their approach combines deep learning with generative AI—a technology similar to what’s used in image generation models but applied to genomic data.

Here’s how it works: one part of the model analyzes raw DNA sequences and chromatin accessibility data, learning how genes are likely arranged based on their underlying sequence. The second part, a generative AI component, then predicts what those structures actually look like, drawing on a training set of over 11 million chromatin conformations.

The result? An AI system that can take a DNA sequence and instantly generate multiple possible 3D structures, reflecting the natural variation found in actual cells.

The Speed Advantage

The biggest advantage of ChromoGen is its speed. “Whereas you might spend six months running experiments to get a few dozen structures, you can generate a thousand structures in a particular region with our model in 20 minutes on just one GPU,” says Greg Schuette, an MIT graduate student and lead author of the study.

When the researchers tested their AI-generated structures against real experimental data, the results were nearly identical. Even more impressive, ChromoGen worked for cell types it hadn’t explicitly been trained on, suggesting it could be used broadly across different biological systems.

Why This Matters

This isn’t just an academic breakthrough. The ability to rapidly model genome architecture has major implications for drug discovery, cancer research, and even gene editing. Many diseases—including some cancers—are driven not just by genetic mutations but by changes in how DNA folds. With AI-powered predictions, researchers could simulate these changes in silico rather than relying on slow, costly lab experiments.

“We’re now at a point where AI can predict genome structures with accuracy on par with cutting-edge experimental techniques,” says Zhang. “That opens up a lot of interesting opportunities.”

The researchers have made their model and data publicly available, inviting scientists around the world to test and build upon their work. As AI continues to reshape biology, breakthroughs like this suggest we’re only beginning to scratch the surface of what’s possible when machine learning meets genomics.

What Comes Next?

The future of genomic research won’t just be about sequencing DNA—it will be about understanding its three-dimensional structure at scale. With AI models like ChromoGen, researchers can now generate thousands of genome configurations in the time it used to take to analyze just one. That could transform how we study gene regulation, cellular identity, and disease mechanisms.

If AI can decode the genome’s hidden architecture faster than any human could, the next big question is: what else is it about to uncover?