Accepted_test

Genomic variants within the coding regions of genes can lead to defective protein synthesis, which can cause disease. However, changes in gene expression can occur with mutations in non-coding genomic regions. Thus, the task of assessing changes in the transcriptional activity of genes as a result of such genomic mutations becomes relevant. In recent years, deep learning algorithms have been actively used in genomics, including the prediction of gene transcriptional activity. State-of-the-art deep learning model Enformer [Avsec et al. 2021] was developed to predict gene transcriptional activity by DNA sequence.  We aimed to check how well the Enformer model deals with predicting changes of gene transcriptional activity on real experimental examples. We analyzed the literature and created a dataset that includes mutations in promoters and enhancers of genes leading to critical gene expression changes. We have shown that in the most investigated cases the Enformer model predicts right direction and level of transcriptional activity changes. This fact allows creating a tool for interpreting non-coding genomic variants that will expand the available information about genetic anomalies obtained by different genomic analyses. This may help medical genetics to put forward new hypotheses regarding the molecular mechanisms associated with the pathological phenotype.