AI Decodes Hidden DNA Initiator Sequence in ~60% of Human Genes

The initiator (Inr) is a core promoter element where RNA polymerase II transcription begins. Prior to this study, the precise DNA sequence pattern of the initiator was unknown, limiting understanding of how genes are activated. The new AI model provides the first strong predictive capability for identifying the initiator in human genes, enabling researchers to predict how mutations affecting this region may contribute to disease the initiator's DNA identity unmasked. This work builds on decades of research into core promoters but solves a long-standing puzzle about the exact sequence requirements of the initiator decades as a component of core promoters. The identification of this pattern in 60% of human genes represents a major advance in understanding gene regulation roughly 60% contain the initiator.

The UC San Diego model represents the current state-of-the-art for initiator identification, achieving strong predictions of presence/absence in human genes for the first time strong predictions of the presence or absence of the initiator. The model was trained on experimental data from approximately 500,000 initiator sequence variants, making it the largest experimentally-validated training dataset for this task analyzing about 500,000 DNA sequences. This approach of combining high-throughput experimentation with machine learning represents a significant methodological advance over previous bioinformatics approaches that relied on computational prediction alone high-throughput DNA sequencing technology to determine the gene expression activity.

The initiator element has been studied for decades as a component of core promoters, but its exact DNA sequence pattern remained elusive decades as a component of core promoters. This study, led by graduate student Torrey Rhyne-Carrigg in Professor James T. Kadonaga's laboratory at UC San Diego, combined high-throughput DNA sequencing with machine learning to crack this long-standing problem machine learning to create an AI model that decoded the initiator's signature DNA pattern. The research was published in Genes & Development in July 2026 Machine learning analysis of the human initiator region reveals key features of different types of core promoters. The study measured gene expression activity across approximately 500,000 different versions of the initiator gene expression activity of approximately 500,000 different versions.

This research was conducted at UC San Diego's Department of Molecular Biology, School of Biological Sciences, funded through standard academic research grants Department of Molecular Biology. No specific funding information was disclosed in the publications. The broader application of AI in genomics represents a significant investment area, with the global AI in life sciences market continuing to grow. The study's authors noted potential commercial applications in synthetic biology and gene therapy design design synthetic promoters. This work builds on the combined use of laboratory experiments and AI to decipher information embedded in DNA sequence combined use of laboratory experiments and AI.

The study's findings could enable commercial applications in synthetic biology. The ability to design synthetic promoters with customized functions could be valuable for biotech companies developing gene therapies, biomanufacturing, and therapeutic protein production design synthetic promoters, sequences that turn genes on and off. Companies working in gene therapy delivery and synthetic DNA design may benefit from these insights synthetic promoters with customized functions. The complementary approach of this initiator model with generative AI systems like DNA-Diffusion could enable creation of novel regulatory sequences from scratch DNA-Diffusion generates new regulatory sequences.

The study was conducted in the laboratory of Professor James T. Kadonaga at UC San Diego Professor James T. Kadonaga's laboratory. Lead author Torrey Rhyne-Carrigg, along with researchers Long Vo Ngoc, Claudia Medrano, and Kassidy E. Gillespie, published their findings in Genes & Development Machine learning analysis of the human initiator region reveals key features of different types of core promoters. The research used high-throughput DNA sequencing to measure gene expression activity across 500,000 different versions of the initiator, then trained a machine learning model to identify the characteristic DNA pattern gene expression activity of approximately 500,000 different versions. The study reveals key features of different types of core promoters key features of different types of core promoters.

Near-term (0-12 months): The AI model will be applied to predict the effects of disease-causing mutations in initiator regions, particularly in cancers and genetic disorders. Researchers will also use the model to design synthetic promoters for biotechnology applications. Mid-term (1-3 years): Expansion of AI models to decode other core promoter elements beyond the initiator, moving toward a comprehensive gene expression code. Integration with existing deep learning models like DNA-Diffusion for generating new regulatory sequences DNA-Diffusion generates new regulatory sequences. Long-term (3-10 years): Development of complete AI models capable of predicting gene expression patterns across the entire human genome, enabling personalized medicine applications where individual gene variants can be assessed for functional impact. The researchers express optimism about expanding these models to cover the full gene expression code within the six billion bases of human DNA expand our AI models of the human gene expression code in the not-too-distant future. This could ultimately enable prediction of the activity of each variant of genes in different people predict the activity of each of the different variants of genes.