Probabilistic analysis of gene expression measurements from heterogeneous tissues

Optimal deconvolution of transcriptional profiling data using quadratic programming with application to complex clinical blood samples

Large-scale molecular profiling technologies have assisted the identification of disease biomarkers and facilitated the basic understanding of cellular processes. However, samples collected from human subjects in clinical trials possess a level of complexity, arising from multiple cell types, that can obfuscate the analysis of data derived from them. Failure to identify, quantify, and incorporate sources of heterogeneity into an analysis can have widespread and detrimental effects on subsequent statistical studies.We describe an approach that builds upon a linear latent variable model, in which expression levels from mixed cell populations are modeled as the weighted average of expression from different cell types. We solve these equations using quadratic programming, which efficiently identifies the globally optimal solution while preserving non-negativity of the fraction of the cells. We applied our method to various existing platforms to estimate proportions of different pure cell or tissue types and gene expression profilings of distinct phenotypes, with a focus on complex samples collected in clinical trials. We tested our methods on several well controlled benchmark data sets with known mixing fractions of pure cell or tissue types and mRNA expression profiling data from samples collected in a clinical trial. Accurate agreement between predicted and actual mixing fractions was observed. In addition, our method was able to predict mixing fractions for more than ten species of circulating cells and to provide accurate estimates for relatively rare cell types (<10% total population). Furthermore, accurate changes in leukocyte trafficking associated with Fingolomid (FTY720) treatment were identified that were consistent with previous results generated by both cell counts and flow cytometry. These data suggest that our method can solve one of the open questions regarding the analysis of complex transcriptional data: namely, how to identify the optimal mixing fractions in a given experiment.

Semi-supervised Nonnegative Matrix Factorization for gene expression deconvolution: a case study

Heterogeneity in sample composition is an inherent issue in many gene expression studies and, in many cases, should be taken into account in the downstream analysis to enable correct interpretation of the underlying biological processes. Typical examples are infectious diseases or immunology-related studies using blood samples, where, for example, the proportions of lymphocyte sub-populations are expected to vary between cases and controls. Nonnegative Matrix Factorization (NMF) is an unsupervised learning technique that has been applied successfully in several fields, notably in bioinformatics where its ability to extract meaningful information from high-dimensional data such as gene expression microarrays has been demonstrated. Very recently, it has been applied to biomarker discovery and gene expression deconvolution in heterogeneous tissue samples. Being essentially unsupervised, standard NMF methods are not guaranteed to find components corresponding to the cell types of interest in the sample, which may jeopardize the correct estimation of cell proportions. We have investigated the use of prior knowledge, in the form of a set of marker genes, to improve gene expression deconvolution with NMF algorithms. We found that this improves the consistency with which both cell type proportions and cell type gene expression signatures are estimated. The proposed method was tested on a microarray dataset consisting of pure cell types mixed in known proportions. Pearson correlation coefficients between true and estimated cell type proportions improved substantially (typically from about 0.5 to approximately 0.8) with the semi-supervised (marker-guided) versions of commonly used NMF algorithms. Furthermore known marker genes associated with each cell type were assigned to the correct cell type more frequently for the guided versions. We conclude that the use of marker genes improves the accuracy of gene expression deconvolution using NMF and suggest modifications to how the marker gene information is used that may lead to further improvements.

Cell subset prediction for blood genomic studies

BACKGROUND

Genome-wide transcriptional profiling of patient blood samples offers a powerful tool to investigate underlying disease mechanisms and personalized treatment decisions. Most studies are based on analysis of total peripheral blood mononuclear cells (PBMCs), a mixed population. In this case, accuracy is inherently limited since cell subset-specific differential expression of gene signatures will be diluted by RNA from other cells. While using specific PBMC subsets for transcriptional profiling would improve our ability to extract knowledge from these data, it is rarely obvious which cell subset(s) will be the most informative.

RESULTS

We have developed a computational method (Subset Prediction from Enrichment Correlation, SPEC) to predict the cellular source for a pre-defined list of genes (i.e. a gene signature) using only data from total PBMCs. SPEC does not rely on the occurrence of cell subset-specific genes in the signature, but rather takes advantage of correlations with subset-specific genes across a set of samples. Validation using multiple experimental datasets demonstrates that SPEC can accurately identify the source of a gene signature as myeloid or lymphoid, as well as differentiate between B cells, T cells, NK cells and monocytes. Using SPEC, we predict that myeloid cells are the source of the interferon-therapy response gene signature associated with HCV patients who are non-responsive to standard therapy.

CONCLUSIONS

SPEC is a powerful technique for blood genomic studies. It can help identify specific cell subsets that are important for understanding disease and therapy response. SPEC is widely applicable since only gene expression profiles from total PBMCs are required, and thus it can easily be used to mine the massive amount of existing microarray or RNA-seq data.