Researchers Develop Method for Direct Generation of Regulatory DNA

© iStock

Researchers at HSE University have developed a model for generating promoters and enhancers—DNA sequences that regulate gene activity. The model works directly with DNA nucleotides, without first transforming them into a continuous numerical representation. This solution could be useful for applications in synthetic biology and gene therapy. The study results were presented at the ICLR 2026 Workshop ‘Generative AI in Genomics (Gen^2): Barriers and Frontiers.’

A skin cell and a neuron contain almost the same DNA, yet they look and function differently, because each cell type has its own set of active genes. Which genes are expressed is largely determined by regulatory regions of DNA: promoters, which help initiate gene transcription, and enhancers, which regulate gene activity.

Researchers are learning to design regulatory regions artificially, with the aim of controlling gene activity in future therapeutic applications. Generative models can be used for this purpose, trained on naturally occurring gene sequences. Once trained, these models identify patterns in DNA and generate new sequences that resemble naturally occurring genomic regions and perform the desired function without directly copying them.

The approach works well with images: pixel values can be changed smoothly as the image is generated. DNA, however, presents a challenge. Because many algorithms are designed to work with continuous data, DNA nucleotides must first be converted into a continuous numerical representation and then converted back into a sequence of letters. Errors can occur during these conversions.

Researchers at the AI and Digital Science Institute of the HSE Faculty of Computer Science investigated whether Discrete Flow Matching could be used to design regulatory DNA sequences. The method retains the idea of gradual generation while working directly with DNA nucleotides, without converting them into a continuous numerical space. The algorithm was trained on three datasets: 100,000 human promoters, 89,000 enhancers from melanoma cells, and 105,000 enhancers from Drosophila brain tissues. The authors tested how well the approach generates promoters and enhancers and compared its performance with four existing models: Bit Diffusion, the Dirichlet Diffusion Score Model, Dirichlet Flow Matching, and Fisher Flow Matching.

The proposed model was tested in two modes: unconditional generation, in which it generated realistic sequences resembling natural ones, and conditional generation, in which it had to take into account promoter activity or tissue type.

When generating promoters, the model performed on par with Fisher Flow Matching and outperformed the other models included in the comparison. It also showed strong results in unconditional enhancer generation. On the melanoma dataset, it outperformed all the other flow-based methods. On the Drosophila brain tissue dataset, it surpassed Dirichlet Flow Matching but performed slightly worse than Fisher Flow Matching.

During conditional generation, the results varied across the three cell groups: the discrete model outperformed Dirichlet Flow Matching in one group but underperformed it in the other two. The authors attribute this to the fact that the mechanism used to guide generation according to a given condition has not yet been fully adapted to discrete data. In the future, the researchers plan to refine the model to enable more precise control over the properties of generated sequences. 

Maria Poptsova

'We showed that the discrete approach works and achieves competitive performance with existing generative methods. Designing regulatory regions directly at the nucleotide level avoids errors that can arise when translating DNA into a continuous representation and back again. In the future, this approach could make it possible to design promoters and enhancers with specified properties, enabling researchers to target or fine-tune the activity of specific genes in particular cell types. Such tools are in demand in synthetic biology and gene therapy, where precise control of gene activity is essential,' explains co-author Maria Poptsova, Director of the Centre for Biomedical Research and Technologies of the HSE AI and Digital Science Institute. 

The study was supported by a grant for research centres in the field of AI provided by the Ministry of Economic Development of the Russian Federation and implemented at HSE University.