A computational method for direct imputation of cell type-specific expression profiles and cellular compositions from bulk-tissue RNA-Seq in brain disorders

Abstract

The importance of cell type-specific gene expression in disease-relevant tissues is increasingly recognized in genetic studies of complex diseases. However, the vast majority of gene expression studies are conducted in bulk tissues, necessitating computational approaches to infer novel biological insights on cell typespecific contribution to diseases. We introduce CellR, a novel computational method that uses external single cell RNA-seq (scRNA-seq) data to infer cell-specific expression profiles from bulk RNA-seq data as well as deconvolving bulk-tissue RNA-Seq data to estimate their cellular compositions. CellR addresses cross-individual gene expression variations by employing genome-wide tissue-wise expression signatures from GTEx to adjust the weights of cell-specific gene markers. It then transforms the deconvolution problem into a linear programming model while taking into account inter/intra cellular correlations. We have developed a multi-variate stochastic search algorithm to estimate the expression level of each gene in each cell type. Extensive analyses on several complex diseases such as schizophrenia, Alzheimer's disease, Huntington disease, and type-2 diabetes validated efficiency of CellR, while revealing how specific cell types contribute to different diseases. We conducted numerical simulations on human cerebellum to generate pseudo-bulk RNA-seq data and demonstrated efficiency of CellR in inferring cell-specific expression profiles directly from bulk data. Moreover, we inferred cell-specific expression levels from bulk RNA-seq data on schizophrenia and computed differentially expressed genes within certain cell-types. Next, using predicted gene expression profile on excitatory neurons, we were able to reproduce our recently published findings on TCF4 being a master regulator in schizophrenia via bulk tissue analysis, but with higher statistical significance, suggesting that cell-type-specific re-analysis of bulk RNA-Seq data can greatly improve understanding of complex brain disorders. In summary, CellR compares favorably (both accuracy and stability of inference) against competing approaches to infer cellular composition from bulk RNA-seq data, and allows direct inference of cell type-specific gene expression, opening new doors to re-analyze vast amounts of gene expression data sets on bulk tissues in many complex diseases.

WormTensor: a clustering method for time-series whole-brain activity data from C. elegans

“…We believe that the advantages of using tensor decomposition outweigh the above disadvantages; this approach not only improves clustering accuracy compared to consensus clustering, as discussed in this work, but also allows for automatic estimation of weights. In data analysis that deals with multiple data sets (e.g., multiple individuals/multi-view/multi-modal/multi-omics data [14,[29][30][31][32] and heterogeneous data fusion [33]), data with large values, a large number of patterns, and large size may dominate in optimization, and multiple data may not be treated equally.…”

“…Hence, weighting is not a trivial issue. Some ad hoc weighting approaches, such as using L2-norm [29], the number of observed elements [34,35], variance [31], the first eigenvalue [30], and correlations with an external standard [32,36], can be used as the weight of each matrix/tensor, but it is unclear which approaches are appropriate. In contrast, in the tensor decomposition algorithm we used, the weights are not pre-set but are instead automatically estimated from the dataset.…”