ИСТИНА

Estimating the impact of different mRNA features is important for genetic engineering and designing gene therapy. In this work, we propose a novel computational approach for predicting the impact of different 5' and 3' UTRs on gene expression. To this end, we used the data generated by a Massively Parallel Reporter Assay (MPRA). In the experiment, 20,000+20,000 human 5' and 3'UTR fragments (respectively) were cloned into a plasmid containing two genes (eGFP and mCherry), which were then packaged into a lentivirus that infects human cells. The effect of a plasmid insertion was evaluated by the ratio of eGFP/mCherry fluorescence. For predicting the impact on the expression level of a particular inserted fragment, we adapted the LegNet neural network as a soft classifier that predicts the probability to find a particular sequence in a fixed cell sorting bin. By supplying additional information to the network and using sequence augmentations, we managed to obtain a model with Pearson r = 0.760 and 0.722 for 5'UTR and 3'UTR sequences, respectively.