Brute-Forcing About 500,000 Variants Through Machine Learning Revealed 'Start Signals' in About 60% of Human Genes

Explanation

A research team at the University of California, San Diego, deciphered the function of short DNA sequences called "initiators" by combining an experiment that measured about 500,000 variants at once with machine learning. As a result, it was predicted that roughly 60% of human genes have an initiator. The paper was published in Genes & Development, with graduate student Torrey Rhyne-Carrigg as first author and Professor James T. Kadonaga of the Department of Molecular Biology as corresponding author.

An initiator is a sequence that marks the position where the information in a gene actually begins to be read. Human cellular DNA contains about 6 billion base pairs, and unless it is determined where to start reading, the information in a gene cannot be used. However, this marker cannot be described simply, such as "an initiator is wherever ATGC appears in this order"; many slightly different sequences all perform similar roles. This has made it difficult to determine from sequence alone whether a given gene has an initiator.

The method the research team took combined brute force with machine learning. First, they created about 500,000 variants with slightly altered initiator sequences, and measured the function of each at once using massively parallel DNA sequencing. This produces data on "which sequences function well and which do not" that is far too large for humans to derive rules from by eye. Machine learning was then applied to extract the patterns determining function.

The resulting model was able to predict whether a human gene has an initiator. According to the research team, this is the first time an AI model has been able to do this.

It should be kept in mind that machine learning did not come up with a new biological principle. The experiment first generated 500,000 sets of answer data, and machine learning extracted the patterns within it. The part that generated the data was handled by experiment. Professor Kadonaga himself stated that what was learned this time is a "small but important part" of the overall mechanism of gene expression.

This model can be used to predict how disease-related mutations affect gene function, or to design artificial promoters that turn genes on and off under targeted conditions.

Paper: Genes & Development, DOI: 10.1101/gad.353623.125
Source: Phys.org (August 22, 2026) https://phys.org/news/2026-08-ai-decodes-dna-sequence-human.html