Regulatory potential of flipons revealed by deep learning.

Poptsova M., Умеренков Д., Fedorov A., Pavlov F., Beknazarov N., Konovalov D., Данилова А. В., Кох В., Herbert A.

Flipons – non-B DNA conformations – have been shown to play an important role in various
genomic processes. Flipons identification and localization is difficult due to their dynamic
nature. We developed deep learning approaches to identify non-B DNA secondary structures
using available information from thousands of omics data sets. We created DeepZ models
based on CNN and RNN, and Z-DNABERT model based on transformer algorithm to predict
Z-flipons at the genome-wide scale. We showed Z-flipon enrichment in promoters and
telomeres and overlap quantitative trait loci for RNA expression, RNA editing, splicing and
disease associated variants. We applied the same approach to quadruplexes and triplexes and
generated whole-genome predictions. We detected that miR- and flipon-based mechanisms are
deeply connected. We found direct interaction of conserved miR and engagement of argonaute
proteins with experimentally validated flipons. Evidences where flipon variants affect
phenotype are provided by case studies.
 

Keywords: Computational biology, deep learning, DNA structure, Recurrent Neural Networks (RNN), CNN (Convolutional neural network), flipons

In book

Proceedings of 11th Moscow Conference on Computational Molecular Biology MCCMB'23

IITP RAS, 2023.