All human cells contain the same genome, but it works quite differently in each of them. Some produce enzymes, others hormones, and still others form the strong framework of bones and connective tissue. The difference arises not in the genes themselves, but in when and with what intensity they are turned on and operate. The answer to this fundamental question of biology has long interested scientists. Now researchers at the University of California, San Diego have found a new way to uncover it by applying artificial intelligence.
Under the leadership of Professor James T. Kadonaga, the team used machine learning to decipher critical 'start' regions of DNA—initiators that mark the site where gene reading begins. The study was led by graduate student Torrey Rhyne-Carigg, who, together with colleagues, performed high-throughput sequencing of about half a million initiator sequence variants and trained a machine learning model to recognize their characteristic patterns. The results showed that an active initiator is present in approximately 60% of focused human gene promoters—higher than previous estimates, which were based on searching for similar sequences and amounted to about 40–56%. The new figure is more reliable because it is based on experimental measurements rather than simple pattern matching.
For the first time, the resulting model made it possible to predict with high accuracy the presence of initiators in the genome and to decipher the minimal functional sequence required for transcription start. This means that scientists can now look at a DNA region and say whether it will act as an active initiator—such a tool was previously impossible.
The discovery has serious practical implications. Mutations in initiators disrupt normal gene activation and can contribute to the development of diseases, including cancer. The new model allows assessing the impact of such changes before conducting experiments on living cells—this saves time and resources and helps doctors more accurately predict the consequences of mutations for each individual patient. In addition, the deciphered initiator signature opens the way for biotechnologists to create synthetic promoters with predictable properties for targeted therapy and genetic engineering.
The human genome contains approximately 3 billion base pairs—a huge instruction in which the rules are encoded: where, when, and in what quantity each of the 20 000–25 000 genes should work. The initiator is just one of many regulatory elements that control this process. The model trained on initiators is the first step in deciphering the molecular 'language' of regulation. The scientists note that in the future they plan to expand the research to other important regulatory elements of the genome: enhancers (which increase gene activity) and silencers (which suppress them)—to build a complete picture of how genes are turned on and off in different cells.
The study, published in the authoritative journal Genes & Development on 31 July 2026, represents an important step in the long-term effort to build a complete artificial intelligence for deciphering the code of human gene expression—the full set of instructions that determine which genes are turned on, when, and in which cells. Practical applications are already visible in the near term: doctors will be able to more accurately predict the consequences of mutations, and biotechnologists will gain a tool for constructing artificial regulatory sequences with desired properties.
Understanding the mechanisms that decide which gene to 'turn on' at the right time and in the right amount brings us closer to more precise personalized medicine. This helps answer a fundamental question of biology: why the same DNA gives rise to such incredible diversity of cells, tissues, and organs in a single organism.
