Abstract 1554: Development of a deconvolution algorithm for tissue-based gene expression data

Abstract Tissue data provide substantially more information than cell-line data, and offer new opportunities to study cancer biology and evolution in its actual microenvironment, when multiple tissue samples of the same cancer type are analyzed together. However, it is very challenging to do information discovery from tissue data because of their compositional complexity - each dataset represents a mixture of gene-expression data from multiple cell types. Hence, meaningful tissue-data analyses require to first sort out the detailed contributions to the observed tissue-level data by different cell types. However, the computational challenge in solving the tissue data deconvolution problem stems from the reality: each cell type has a very large number of complex relations among its expressed genes and pathways, which are preserved under different conditions. To make deconvolution results meaningful, the co-expressions among functionally closely related genes, must be captured and enforced in a deconvolution problem formulation. We have a fundamentally novel formulation of the tissue-data deconvolution problem, which is pathway instead of gene based. It preserves co-expressions of genes of the same pathways through capturing and using an expression signature among such genes, and allows differential expressions of pathways by giving them varying weights in different tissues. In addition, this pathway-based model substantially reduces the number of free variables compared to gene-based models, making our problem formulation efficiently solvable. The unique ideas of our approach are: (i) estimate the proportion of each cell type within a mixture based on expression patterns of cell-type specific genes; (ii) identify co-expressed gene clusters among genes encoding each pathway (as defined by REACTOME) in each cell type based on cell line data; (iii) derive a condition-invariant expression signature for each of the ~2,000 REACTOME pathways in each cell type; (iv) demonstrate that each set of cell line gene-expression data can be uniquely represented as a weighted sum of such signatures; and (v) formulate the deconvolution problem as to estimate the weight of each pathway in each cell type that minimizes the Frobenius norm of deviance matrix of estimated and observed gene expression matrices. We demonstrated its effectiveness on simulated data, i.e., in silico mixtures of gene-expression data from different cell types with varying proportions of each cell type. We anticipate that the successful development and deployment of the planned deconvolution method will enable and inspire a wide range of new ways to study tissue-based expression data and uncover the very rich information hidden in cancer tissues about the fundamental biology of cancer. Citation Format: Sha Cao, Chi Zhang, Ying Xu. Development of a deconvolution algorithm for tissue-based gene expression data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 1554. doi:10.1158/1538-7445.AM2017-1554