Researchers at the University of California San Diego have used machine learning to decode the DNA sequence of the "initiator," a regulatory element that starts gene activation. The finding, based on an analysis of about 500,000 experimental data points, could help predict how DNA mutations contribute to diseases such as cancer.
The initiator is the site where the instructions coded in genes are first converted into functional products. When it fails to work correctly, cells can malfunction or develop into disorders. Until now, researchers did not know the exact DNA pattern that defines this sequence across human genes.
The team, led by graduate student Torrey Rhyne-Carrigg in Professor James T. Kadonaga's lab, measured gene expression activity for roughly half a million initiator variants using high-throughput DNA sequencing. They then trained an AI model on that data to identify the initiator's underlying sequence pattern. The model revealed that about 60% of human genes contain the initiator.
What the AI model found
"These AI models were found to provide, for the first time, strong predictions of the presence or absence of the initiator in human genes, and were thus able to decode the DNA base sequence pattern of the initiator," said Kadonaga, a professor in the UC San Diego Department of Molecular Biology, School of Biological Sciences.
With the initiator's sequence now known, researchers can scan human genomes for mutations in this region that may lead to disease. The data and models from the study could also support the design of synthetic promoters - sequences that turn genes on and off - with customized functions for research or therapeutic use.
The study, published in Genes and Development, was authored by Rhyne-Carrigg, Long Vo ngoc, Claudia Medrano, Kassidy E. Gillespie, and Kadonaga. Computational work was performed on the Expanse system at the San Diego Supercomputer Center, with support from the National Science Foundation and the National Institutes of Health.
Toward a complete gene expression code
Kadonaga sees the initiator model as one piece of a larger puzzle. The human genome contains roughly six billion DNA bases, and within them lies a code that determines when, where, and how much each gene is expressed. An AI model for the entire code would allow researchers to predict gene activity for different variants in different people.
"More globally, this work is a step forward in the combined use of laboratory experiments and AI to decipher the information that is embedded in the sequence of the DNA bases in humans," said Kadonaga. "Ultimately, within the six billion bases of DNA in each of our cells, there is a gene expression code that specifies when, where and to what extent each of our genes should be turned on or off. If we had an AI model for the entire gene expression code, we would be able to predict the activity of each of the different variants of genes in different people. The new AI model for the initiator is a small but important part of this gene expression code, and I am optimistic that we will expand our AI models of the human gene expression code in the not-too-distant future."
For researchers working in genomics or molecular biology, the study offers a concrete example of how machine learning can extract regulatory rules from large experimental datasets. The approach - combining high-throughput assays with AI - is transferable to other DNA elements whose sequence logic remains unknown. Scientists studying gene regulation, disease genetics, or synthetic biology can apply similar methods to their own systems of interest.
Those looking to build skills in this area may find AI Learning Path for Research Scientists useful for understanding how to apply machine learning to biological data. Broader resources on AI for Science & Research cover related applications across scientific disciplines.
Why this matters for science and research professionals
This work demonstrates a practical workflow: generate large-scale experimental data, train an AI model to find patterns, then use those patterns to make predictions about disease-related mutations. For researchers in genomics, the initiator model provides a new tool for interpreting variants in non-coding DNA - a region of the genome that has been difficult to analyze with traditional methods. The study also shows that AI models can be built from laboratory data rather than requiring pre-existing datasets, which may be relevant for labs considering similar approaches in other areas of gene regulation.
Your membership also unlocks:
