DNA is a string of pearls enclosed in a tiny space. For decades, we've studied the sequence of these pearls: the A, T, C, and G bases that form the code. But what really matters is how this string folds. The loops, folds, and three-dimensional structures decide which genes are turned on and which remain off. MIX-HIC è the first multimodal model of artificial intelligence that can read this hidden architecture. Yes, sir, the 3D Genome. It does this by integrating two types of information: maps that show which parts of the DNA touch each other in space (called Hi-C) and chemical signals that indicate where the DNA is accessible (epigenomic signatures). What does this mean? Let's take a look together, calmly.
The 3D genome is not a detail: it is the mechanism
A liver cell and a neuron have the same DNA. Yet they do completely different things. The secret is not in the sequence, but in form. When DNA forms a chromatin loop, brings distant regions of the genome closer together. A chromatin loop is like a molecular bridge connecting a switch and the gene it's supposed to control. The shape is the switch. If the loop doesn't form, the gene stays off. If it forms in the wrong place, it turns on when it shouldn't.
Studying the 3D genome has so far been like putting together a puzzle with pieces from different boxes. Scientists had the contact maps of DNA, showing which regions touch each other in space. They also had the epigenomic traces, which reveal where the DNA is "open" and ready to be read. But these data were analyzed separately, with methods specialized for specific tasks. In practice, the overall picture was lost.
MIX-HIC learns two languages and the nuances of both
The team ofHong Kong University of Science and Technology has developed a system that completely changes the approach. MIX-HIC has been trained on over 1,2 million paired samples of Hi-C maps and epigenomic signatures: the largest dataset ever created for 3D genome study. The architecture is ingenious: the system doesn't simply "merge" the different types of data. It learns to recognize both the features common to both modalities and the unique features of each. It's like a translator who not only knows two languages but also understands cultural nuances.
In tests, MIX-HIC outperformed all existing methods in three key tasks. First: predict how DNA will organize itself in 3D in different cell types, with improvements up to 9,3% compared to the best previous systems. According to: identify chromatin loops with unprecedented precision. Third: predict how active a gene will be, essential information for understanding cell function.
When data is missing, imagine it
Hi-C data is expensive and difficult to obtain. It's often missing. MIX-HIC solves this problem too: thanks to its extensive training, the system can "imagine" how the 3D genome would be organized even when only epigenomic traces are available. It's like a detective reconstructing a crime scene with few clues, having learned from thousands of previous cases. This "unified semantics" approach allows the system to fill in gaps in the experimental data while maintaining the accuracy of its predictions. Of course, human supervision is required—it's no Oracle of Delphi—but the time savings are immense.
Diseases are not just wrong sequences
Many diseases (from cancers to rare genetic disorders) are caused not by simple errors in the DNA sequence, but by problems in the organization of the 3D genome. A gene important for suppressing tumors might be "switched off" because the 3D structure prevents regulatory factors from reaching it. Or a mutation might not directly change a gene, but rather alter a chromatin loop that controls it remotely.
With MIX-HIC, researchers can now analyze the 3D genome of specific patients to identify these structural problems. This opens the way to more precise diagnoses: Understanding exactly what's wrong with a patient's DNA organization. And targeted therapies that take into account not only which genes are mutated, but how the entire genomic architecture is altered. As is already happening in personalized nutrition, where genomics and AI are revolutionizing the way we care for our health.
3D Genome: Amplifies Capabilities, Not Replaces Them
The system accelerates basic research: instead of conducting lengthy and expensive experiments for each cell type, researchers can use MIX-HIC to virtually explore thousands of scenarios, identifying the most promising ones for laboratory testing. It's a tool that democratizes access to precision genomics, making it faster and more affordable. The methods developed to model these weak signals can also be used in astronomy, planetary defense, and monitoring the impact of human technology on our space environment.
MIX-HIC is an example of how artificial intelligence can amplify human capabilities in scientific research. It doesn't replace scientists: it offers them a tool to decipher one of biology's most complex puzzles: how the three-dimensional shape of our genome orchestrates the symphony of cellular life.
And when the music is out of tune, maybe he can tell us what we can do to bring it back into harmony.
