AI Deciphers DNA Sequences Present in 60% of Human Genes
Researchers at the University of Toronto have developed an artificial intelligence model capable of decoding regulatory DNA sequences that account for 60% of the human genome. The tool, described in a study published in the journal Nature, provides a method to predict how specific genetic variations influence the expression of genes, a process previously difficult to map due to the complexity of non-coding "junk" DNA.
Decoding Non-Coding DNA
For decades, geneticists focused primarily on the 2% of the human genome that codes for proteins. The remaining 98%—often referred to as non-coding or "dark" DNA—contains the regulatory switches that determine when, where, and how much protein a gene produces. The new AI model, known as Sei, functions by analyzing these sequences to identify the biological "grammar" of the genome.
According to the study, the model acts as a genomic interpreter. It categorizes DNA sequences into distinct functional "bins," allowing researchers to determine whether a specific mutation in a non-coding region is benign or likely to disrupt a critical cellular process. This capability addresses a significant bottleneck in clinical genetics, where patients often present with rare diseases that cannot be linked to protein-coding mutations.
Clinical and Research Applications
The development of Sei offers a path toward interpreting the functional impact of variants identified in large-scale genome-wide association studies. By predicting the effects of non-coding mutations, the model assists researchers in prioritizing which genetic changes warrant further experimental investigation.
The research team, led by Jian Zhou, utilized deep learning architectures to train the model on vast datasets of genomic and epigenomic information. The system demonstrated an ability to predict the consequences of mutations across a wide range of human tissues, providing a comprehensive view of how regulatory landscapes vary across the body.
Limitations and Future Validation
While the model provides high-throughput predictions, the researchers noted that it serves as a predictive tool rather than a definitive diagnostic instrument. The functional outcomes suggested by the AI require subsequent laboratory validation to confirm biological activity in living cells.
The research highlights a shift in how computational biology approaches the interpretation of the human blueprint. By moving beyond the protein-coding sequences, the model provides a framework for understanding the regulatory architecture that underpins human development and disease susceptibility. The team has made the model available for further testing by the broader scientific community to facilitate ongoing genomic research.