Deep learning models for modeling cellular transcription systems

Cellular signal transduction system (CSTS) plays a fundamental role in maintaining homeostasis of a cell by detecting changes in its environment and orchestrates response. Perturbations of CSTS lead to diseases such as cancers. Almost all CSTSs are involved in regulating the expression of certain genes and leading to signature changes in gene expression. Therefore, the gene expression profile of a cell is the readout of the state of its CSTS and could be used to infer CSTS. However, a gene expression profile is a convoluted mixture of the responses to all active signaling pathways in cells. Therefore it is difficult to find the genes associated with an individual pathway. An efficient way of de-convoluting signals embedded in the gene expression profile is needed.\ud At the beginning of the thesis, we applied Pearson correlation coefficient analysis to study cellular signals transduced from ceramide species (lipids) to genes. We found significant correlations between specific ceramide species or ceramide groups and gene expression. We showed that various dihydroceramide families regulated distinct subsets of target genes predicted to participate in distinct biologic processes. However, it’s well known that the signaling pathway structure is hierarchical. Useful information may not be fully detected if only linear models are used to study CSTS. More complex non-linear models are needed to represent the hierarchical structure of CSTS. This motivated us to investigate contemporary deep learning models (DLMs).\ud Later, we applied various deep hierarchical models to learn a distributed representation of statistical structures embedded in transcriptomic data. The models learn and represent the hierarchical organization of transcriptomic machinery. Besides, they provide an abstract representation of the statistical structure of transcriptomic data with flexibility and different degrees of granularity. We showed that deep hierarchical models were capable of learning biologically sensible representations of the data (e.g., the hidden units in the first hidden layer could represent transcription factors) and revealing novel insights regarding the machinery regulating gene expression. We also showed that the model outperformed state-of-the-art methods such as Elastic-Net Linear Regression, Support Vector Machine and Non-Negative Matrix Factorization