![]() |
ИСТИНА |
Войти в систему Регистрация |
ИПМех РАН |
||
The diversity of eukaryotic cells is determined by the controlled gene expression through a complex gene regulatory network. In turn, the stage for the rational design of regulatory sequences is being set by the advanced computational modeling of the transcriptional regulatory grammar. In this study, we used the cold diffusion approach to train a generative neural network model to iteratively correct artificial noise by reversing computationally introduced single-nucleotide substitutions found in sequences with known expression levels. The model was then used to 'correct' completely random sequences so that upon full correction with single-nucleotide substitutions the resulting sequence provides a desired expression level. The Pearson and Spearman correlation between the target (requested) and actually generated (predicted by the original neural network) expression reached 0.839 and 0.843, respectively. This result demonstrates the proof of concept for the effective application of diffusion-like models in the process of generating promoter sequences.