HSE University researchers have developed a generative neural network that designs promoters and enhancers – DNA regions that control gene activity. Simply put, they act as buttons and regulators: they help turn on the necessary genes and change the intensity of their work in specific cells.
The main difference of the development is that it works directly with nucleotides, i.e., genetic "letters". Other models first convert DNA into a numerical format and then reassemble the sequence, which can lead to errors. The scientists bypassed this stage using the discrete flow matching method.
For training, about 100 thousand human promoters, 89 thousand enhancers from melanoma cells, and 105 thousand sequences from the Drosophila fly brain were used. In tests, the model outperformed most competitors in creating promoters and showed one of the best results when working with enhancers.
In the future, AI will help assemble DNA regions with specified properties. The technology can be useful in the development of gene therapy, as well as in synthetic biology – for example, to tune cells to produce drugs, enzymes, and other useful molecules.