Integrating prior knowledge in neural network methods to reduce the dimensions of single-cell RNA-Seq Data

Single cell RNA-seq data provides valuable insights into cellular heterogeneity which may significantly improve our understanding of biology and human disease. Nonetheless, there are some challenges to identify new cell types and cell states. The common methods to address this issue are unsupervised approaches, based on clustering methods together with methods to reduce the huge dimensionality of data. Although they are widely using, the lack the ability to capture the complex patterns behind single cell data can lead to poor performance or misleading interpretations. In this project, neural network architectures are using to reduce the dimension of data, finding complex patterns, in a supervised framework. To avoid the lack of interpretability of networks (as called black box), several types of prior biological information which are protein-protein interactions, protein-DNA interactions and pathway are integrated in neural network architectures. The goal of this project is to integrate biological information to create an interpretable neural network.