Neural networks for classifying human traits based on RNA elements

• Applicant: Professor Dr. Michael Sammeth
• Subject Area: Bioinformatics and Theoretical Biology
• Term: since 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 439797931

In our future research we want to leverage machine learning, in particular artificial neuronal networks due to their transparent parameters, for investigating molecular circumstances (i.e., biomarkers) that shape human phenotypes. A first objective is to develop a universally applicable method that – as a groundbreaking difference to earlier approaches – integrates a quantitative representation of generic RNA processing patterns with DNA mutations that are retained in exonic sequences. So far, studies on biomarkers for phenotypes have been exclusively based on either gene expression or on DNA variants, compared to which our model is expected to produce higher signal-to-noise ratios. Our method is based on RNA haplotype graphs, an extension of our previously developed, graph-based framework for our long-standing studies of human transcriptomes. These haplotype graphs can be constructed from RNA-seq experiments and provide an intrinsic hierarchical organization of alternative RNA elements that we employ to describe a sparse neuronal architecture, in order to address limitations by the amount of available training data. We outline how to develop our neuronal network for general phenotype classification problems based on the RNA-seq data from the population-scale GEUVADIS and GTEx projects, to which we contributed during the last ten years. Our project will produce several technical insights about applying artificial intelligence to the classification of phenotypes from RNA-seq data, and we are convinced that we will find many novel and interesting biology along our way to study the RNA elements that contribute to human phenotype(s).

• DFG Programme: Research Grants