GO-PCA: An Unsupervised Method to Explore Biological Heterogeneity Based on Gene Expression and Prior Knowledge

Genome-wide expression profiling is a cost-efficient and widely used method to characterize heterogeneous populations of cells, tissues, biopsies, or other biological specimen. The exploratory analysis of such datasets typically relies on generic unsupervised methods, e.g. principal component analysis or hierarchical clustering. However, generic methods fail to exploit the significant amount of knowledge that exists about the molecular functions of genes. Here, I introduce GO-PCA, an unsupervised method that incorporates prior knowledge about gene functions in the form of gene ontology (GO) annotations. GO-PCA aims to discover and represent biological heterogeneity along all major axes of variation in a given dataset, while suppressing heterogeneity due to technical biases. To this end, GO-PCA combines principal component analysis (PCA) with nonparametric GO enrichment analysis, and uses the results to generate expression signatures based on small sets of functionally related genes. I first applied GO-PCA to expression data from diverse lineages of the human hematopoietic system, and obtained a small set of signatures that captured known cell characteristics for most lineages. I then applied the method to expression profiles of glioblastoma (GBM) tumor biopsies, and obtained signatures that were strongly associated with multiple previously described GBM subtypes. Surprisingly, GO-PCA discovered a cell cycle-related signature that exhibited significant differences between the Proneural and the prognostically favorable GBM CpG Island Methylator (G-CIMP) subtypes, suggesting that the G-CIMP subtype is characterized in part by lower mitotic activity. Previous expression-based classifications have failed to separate these subtypes, demonstrating that GO-PCA can detect heterogeneity that is missed by other methods. My results show that GO-PCA is a powerful and versatile expression-based method that facilitates exploration of large-scale expression data, without requiring additional types of experimental data. The low-dimensional representation generated by GO-PCA lends itself to interpretation, hypothesis generation, and further analysis.