A novel paradigm for Precision Medicine: sparse non-linear Neural Networks for end-to-end Genome Interpretation

In the last decade, the surge of high-throughput sequencing and big data technologies seemed almost unstoppable, to the point that even finally uncovering all aspects of our genome and enabling precision medicine seemed at reach. Ten years later, notwithstanding many advances in genetics, our genome is still hiding most of its secrets. When it comes to disorders with complex aetiology, the bottleneck has indeed only shifted from a data availability to a data interpretation problem, since the classical computational approaches have been largely unsuccessful in explaining these diseases.
Here we address this problem by building the foundations of a conceptually novel end-to-end Genome Interpretation (EtEGI) framework which aims at directly modeling the genotype-to-phenotype relationship using cutting-edge Machine Learning (ML) approaches, such as Neural Networks (NNs). EtEGI has just become feasible due to the critical mass of genomics data and the popularization of flexible NN libraries. We propose a fully differentiable framework of NN tailored solutions to deal with these unique kinds of genomics and phenotypic input and output data, addressing each NN modeling challenge in a biologically meaningful manner. EtEGI will be a fundamental step towards posing the basis for true precision medicine, and to do so we will address both ambitious ML and biological challenges related to modeling, accountability, scalability, interpretability and explainability of ML methods.