Computational deconvolution of gene expression by individual host cellular subsets from microarray analyses of complex, parasite-infected whole tissues

Abstract

Analyses of whole organs from parasite-infected animals can reveal the entirety of the host tissue transcriptome, but conventional approaches make it difficult to dissect out the contributions of individual cellular subsets to observed gene expression. Computational deconvolution of gene expression data may be one solution to this problem. We tested this potential solution by deconvoluting whole bladder gene expression microarray data derived from a model of experimental urogenital schistosomiasis. A supervised technique was used to group B-cell and T-cell related genes based on their cell types, with a semi-supervised technique to calculate the proportions of urothelial cells. We demonstrate that the deconvolution technique was able to group genes into their correct cell types with good accuracy. A clustering-based methodology was also used to improve prediction. However, incorrectly predicted genes could not be discriminated using this methodology. The incorrect predictions were primarily IgH- and IgK-related genes. To our knowledge, this is the first application of computational deconvolution to complex, parasite-infected whole tissues. Other computational techniques such as neural networks may need to be used to improve prediction.

Keywords: Bioinformatics; Bladder; Deconvolution; Gene expression; Microarray; Mouse model; Schistosoma haematobium; Schistosomiasis.

Fig. 1

Diagram showing various deconvolution methodologies (adapted from Shen-Orr et al., 2010 and Gaujoux et al., 2013). In partial deconvolution, either cell proportion (B) or cell-specific gene signatures (C) are known together with sample expression profiles. In complete devonvolution, only sample gene expression is known (A). When there is some knowledge of cell proportion and cell-specific gene signatures, a semi-supervised strategy (D) can be used. Each row signifies distinct gene expression profiles, while the different shapes (polygon, hexagon, oval) signify cell types.

Fig. 1

Diagram showing various deconvolution methodologies (adapted from Shen-Orr et al., 2010 and Gaujoux et al., 2013). In partial deconvolution, either cell proportion (B) or cell-specific gene signatures (C) are known together with sample expression profiles. In complete devonvolution, only sample gene expression is known (A). When there is some knowledge of cell proportion and cell-specific gene signatures, a semi-supervised strategy (D) can be used. Each row signifies distinct gene expression profiles, while the different shapes (polygon, hexagon, oval) signify cell types.

Fig. 1

Diagram showing various deconvolution methodologies (adapted from Shen-Orr et al., 2010 and Gaujoux et al., 2013). In partial deconvolution, either cell proportion (B) or cell-specific gene signatures (C) are known together with sample expression profiles. In complete devonvolution, only sample gene expression is known (A). When there is some knowledge of cell proportion and cell-specific gene signatures, a semi-supervised strategy (D) can be used. Each row signifies distinct gene expression profiles, while the different shapes (polygon, hexagon, oval) signify cell types.

Fig. 1

Diagram showing various deconvolution methodologies (adapted from Shen-Orr et al., 2010 and Gaujoux et al., 2013). In partial deconvolution, either cell proportion (B) or cell-specific gene signatures (C) are known together with sample expression profiles. In complete devonvolution, only sample gene expression is known (A). When there is some knowledge of cell proportion and cell-specific gene signatures, a semi-supervised strategy (D) can be used. Each row signifies distinct gene expression profiles, while the different shapes (polygon, hexagon, oval) signify cell types.

Fig. 2

Proportion of B- and T-cells in mouse bladders at various time points following schistosome egg injection. Data derived from previously unpublished flow cytometry data. The sample size was at least three animals in each group.

Fig. 3

Proportions of urothelial cell as calculated by semi-supervised deconvolution techniques. The control group (ctrl) is clustered in a tight region, whereas samples from the experimental group (exp) show more variation, due to variation in B- and T-cell representation.

Fig. 4

Dendrogram showing clustering of incorrectly predicted immunoglobulin-related genes in relation to T-cell related genes. The B-cell genes that were incorrectly identified using the least squares strategy could not be differentiated by a dendrogram-based algorithm, since the genes were misassigned to T-cells. The black-coded genes are T-cell-related genes, with the rest being B-cell related genes.