Diffusion model predicts 3D genomic structures

by MIT News

12 February 2025

From sequence to structure

Inside the cell nucleus, DNA and proteins form a complex called chromatin, which has several levels of organization, allowing cells to cram 2 meters of DNA into a nucleus that is only one-hundredth of a millimeter in diameter. Long strands of DNA wind around proteins called histones, giving rise to a structure somewhat like beads on a string.

Chemical tags known as epigenetic modifications can be attached to DNA at specific locations, and these tags, which vary by cell type, affect the folding of the chromatin and the accessibility of nearby genes. These differences in chromatin conformation help determine which genes are expressed in different cell types, or at different times within a given cell.

Over the past 20 years, scientists have developed experimental techniques for determining chromatin structures. One widely used technique, known as Hi-C, works by linking together neighboring DNA strands in the cell’s nucleus. Researchers can then determine which segments are located near each other by shredding the DNA into many tiny pieces and sequencing it.

This method can be used on large populations of cells to calculate an average structure for a section of chromatin, or on single cells to determine structures within that specific cell. However, Hi-C and similar techniques are labor-intensive, and it can take about a week to generate data from one cell.

To overcome those limitations, Zhang and his students developed a model that takes advantage of recent advances in generative AI to create a fast, accurate way to predict chromatin structures in single cells. The AI model that they designed can quickly analyze DNA sequences and predict the chromatin structures that those sequences might produce in a cell.

“Deep learning is really good at pattern recognition,” Zhang says. “It allows us to analyze very long DNA segments, thousands of base pairs, and figure out what is the important information encoded in those DNA base pairs.”

ChromoGen, the model that the researchers created, has two components. The first component, a deep learning model taught to “read” the genome, analyzes the information encoded in the underlying DNA sequence and chromatin accessibility data, the latter of which is widely available and cell type-specific.

The second component is a generative AI model that predicts physically accurate chromatin conformations, having been trained on more than 11 million chromatin conformations. These data were generated from experiments using Dip-C (a variant of Hi-C) on 16 cells from a line of human B lymphocytes.

When integrated, the first component informs the generative model how the cell type-specific environment influences the formation of different chromatin structures, and this scheme effectively captures sequence-structure relationships. For each sequence, the researchers use their model to generate many possible structures. That’s because DNA is a very disordered molecule, so a single DNA sequence can give rise to many different possible conformations.

“A major complicating factor of predicting the structure of the genome is that there isn’t a single solution that we’re aiming for. There’s a distribution of structures, no matter what portion of the genome you’re looking at. Predicting that very complicated, high-dimensional statistical distribution is something that is incredibly challenging to do,” Schuette says.

Rapid analysis

Once trained, the model can generate predictions on a much faster timescale than Hi-C or other experimental techniques.

“Whereas you might spend six months running experiments to get a few dozen structures in a given cell type, you can generate a thousand structures in a particular region with our model in 20 minutes on just one GPU,” Schuette says.

After training their model, the researchers used it to generate structure predictions for more than 2,000 DNA sequences, then compared them to the experimentally determined structures for those sequences. They found that the structures generated by the model were the same or very similar to those seen in the experimental data.

“We typically look at hundreds or thousands of conformations for each sequence, and that gives you a reasonable representation of the diversity of the structures that a particular region can have,” Zhang says. “If you repeat your experiment multiple times, in different cells, you will very likely end up with a very different conformation. That’s what our model is trying to predict.”

The researchers also found that the model could make accurate predictions for data from cell types other than the one it was trained on. This suggests that the model could be useful for analyzing how chromatin structures differ between cell types, and how those differences affect their function. The model could also be used to explore different chromatin states that can exist within a single cell, and how those changes affect gene expression.

“ChromoGen provides a new framework for AI-driven discovery of genome folding principles and demonstrates that generative AI can bridge genomic and epigenomic features with 3D genome structure, pointing to future work on studying the variation of genome structure and function across a broad range of biological contexts,” says Jian Ma, a professor of computational biology at Carnegie Mellon University, who was not involved in the research.

Another possible application would be to explore how mutations in a particular DNA sequence change the chromatin conformation, which could shed light on how such mutations may cause disease.

“There are a lot of interesting questions that I think we can address with this type of model,” Zhang says.

The researchers have made all of their data and the model available to others who wish to use it.

The research was funded by the National Institutes of Health.