Every cell in the human body shares the same genetic sequence, but each one expresses only specific genes, distinguishing a brain cell from a skin cell. These unique gene expression patterns are influenced by the three-dimensional arrangement of the genetic material, which determines gene accessibility. Researchers at MIT have developed a novel method to analyze these 3D genome structures using generative artificial intelligence, allowing them to predict thousands of structures within minutes. This advancement significantly accelerates the process compared to traditional experimental techniques. The study's lead author, associate professor Bin Zhang, aims to relate DNA sequences to their corresponding 3D genome structures. The new technique competes with cutting-edge experimental methods, presenting promising research opportunities. Inside cell nuclei, DNA and proteins form chromatin in various organizational levels, compacting 2 meters of DNA into a nucleus a mere one-hundredth of a millimeter wide. Epigenetic modifications attached to DNA affect chromatin folding, determining which genes are expressed in different cell types or at different times. While methods like Hi-C have been developed over the past two decades to determine chromatin structures, they require extensive time and effort, often taking about a week for data from a single cell.

To address this, Zhang and his team created a model that leverages deep learning and generative AI, allowing for swift and precise predictions of chromatin structures from DNA sequences. Their model, ChromoGen, comprises a deep learning model that analyzes DNA information and a generative AI model trained on over 11 million chromatin conformations. This integrated system captures the relationship between DNA sequences and chromatin structures, generating multiple possible structures for each sequence due to the inherent disorder in DNA. ChromoGen allows for rapid predictions—where existing methods may take six months to achieve a few dozen structures, the model can produce a thousand structures in about 20 minutes. After training, the researchers used the model to predict structures for over 2, 000 DNA sequences, confirming that the generated structures closely matched experimental data. Furthermore, the model demonstrated accuracy with data from cell types outside of its training set, indicating potential applications for analyzing chromatin structure variations across cell types and within individual cells. This capability could also facilitate research into how DNA mutations might alter chromatin conformation, potentially linking to disease mechanisms. The team has made their research data and model publicly available for further exploration. The study was funded by the National Institutes of Health.