Similarity identification in gene expression patterns as a new approach in phenotype classification

Abstract

Stratifying healthy and malignant phenotypes and identifying their biological states using high-throughput molecular data has been the focus of many computational approaches during the last decade. Using multivariate changes in expression of genes within biological pathways, as fingerprints of complex phenotypes, we developed a new methodology for Similarity Identification in Gene expressioN (SIGN). In this approach, we use centroid classifier to identify phenotype of each biological sample. To obtain similarity of a given biological sample with classes of phenotypes, we defined a new distance measure, transcriptional similarity coefficient (TSC) which captures similarity of gene expression patterns between a biological pathway in two samples or populations. We showed that TSC, as an interpretable and stable distance measure in SIGN, captures all oncogenic hallmarks for breast cancer even with low sample size, by comparing healthy and patient tumor samples in the largest breast cancer dataset. In this study, we demonstrate that SIGN is a flexible, yet robust approach for classification based on transcriptomics data. Comparing early and late relapses within each molecular subtypes of breast cancer, our method enabled subtype-specific stratification of breast cancer patients into groups with significantly different survival. Moreover, we used SIGN to classify with more than 99% specificity the site of extraction of healthy and tumor samples from the Genotype-Tissue Expression (GTEx) and The Cancer Genome Atlas (TCGA) datasets. We showed that SIGN also enables robust identification of hematopoietic stem cell and progenitors within the hematopoietic hierarchy. We further explored chemical perturbation data in the Connectivity Map (CMAP) database and showed that SIGN was able to classify seven classes of drugs based on their mechanism of action. In conclusion, we showed that SIGN can be used to achieve interpretable and robust transcriptomic-based classification of healthy and malignant samples, as well as drugs based on their known mechanism of action, supporting the generalizability and relevance of the method for the analysis of gene expression profiles.