Scientific report «Machine learning methods in functional genomics»

The speaker is Maria Poptsova, Head of the International Laboratory of Bioinformatics

With the help of neural network deep learning methods, it became possible to aggregate information about functional elements of different levels of cellular organization — genomics, epigenomics, proteomics, metabolomics — and other "omics”, in order to predict functional elements for which experiments either did not reach the desired quality or were not available.

In this talk, I will discuss the deep learning methods developed at the International Laboratory of Bioinformatics for predicting secondary DNA structures. Models based on convolutional (CNN), recurrent (RNN), and generative adversarial (GAN) networks have been developed, as well as transfer learning methods with domain adaptation for G-quadruplex and Z-DNA prediction problems. I will also introduce the approaches developed in the laboratory from the field of explainable artificial intelligence (XAI), which allow us to identify significant patterns of association of the epigenetic code and the code of secondary DNA structures.