Reconstructing gene regulatory networks with neural relational inference

Discovering structures of gene regulatory networks (GRNs) based on large-scale high throughput data is a fundamental research challenge in systems biology. The inference of accurate GRNs from data is challenging not only because of the large number of components involved, but also their possible mutual interactions. The inferred network models can provide system-level understanding of the mechanism of cell transformations, which is crucial for the comprehension of the progression of complex diseases such as Parkinson’s disease (PD), and the affected cell types.

GENERIC takes an interdisciplinary approach towards GRN inference, integrating theories, techniques and tools from biology and computer science. It aims to devise data-driven deep learning models for inferring GRNs with high accuracy from single-cell gene expression data. In particular, with this proposed project we will be the first to develop novel GRN inference methods by exploiting neural relational inference (NRI) to discover latent interactions directly from data. This paves a new way for accurate and efficient reconstruction of comprehensive real-life GRNs.

18Our inference methods combine an iterative graph generation process and reinforcement learning into NRI in order to increase its efficiency and accuracy, and as well to incorporate prior biological knowledge of the studied GRN and well-established GRN structural properties. The development of our methods will be driven and accompanied by application of the methods to single cell gene expression data from differentiation of human dopaminergic neurons (hDAns), a cell type with high biomedical relevance for example in PD. Moreover, the effectiveness and accuracy of the inferred GRNs will be validated in wet-lab using ChIP-seq and CRISPR-dCas9 inhibition experiments in hDAns derived from induced pluripotent stem cells (iPSCs). The newly identified networks will give rise to better control of hDAn differentiation.