Fig. 5

Phenotype-guided topic modeling of bulk RNA-seq data of breast cancer. a Predicted top genes for the phenotype-guided topics for basal and ER + breast cancer subtypes. GTM-decon was trained on the sparsified TCGA-BRCA bulk RNA-seq data with the basal and ER + cancer subtypes as the guide. Five topics were used per subtype and therefore 10 topics in total. The heatmap illustrates the topic probabilities of the top 20 genes from each topic. As a comparison, the genes were also labeled as up- or downregulated if they were deemed differentially expressed by the DESeq2 analysis. b Classification of basal and ER + subtypes based on phenotype-guided topic scores. The 5 topics for the same subtype were summed to obtain the overall score for basal and ER + subtype. Subjects in the rows were sorted by their basal topic scores. c GSEA analysis of the basal and ER + subtype topics. Significantly enriched MSigDb HALLMARK pathways were identified for each topic and displayed as barplots. The heights of the bar indicate the − log10 adjusted p-values and the colors indicate enriched pathways. d Predicted top genes for the phenotype-guided topics for histological subtypes. Same as in panel a but for ductal and lobular subtypes. e Classification of histological subtypes. Same as in panel b but for ductal and lobular subtypes. f Evaluation of the subtype classification accuracy on the test breast tumor samples. We trained the phenotype-guided GTM-decon separately on 80% of the sparsified TCGA-BRCA tumor samples using basal/ER + and histological types as the guides and evaluated its phenotype prediction accuracy on the 20% held sparsified samples. As a comparison, we also trained and evaluated logistic regression and random forest on the same training and test split, respectively. The classification accuracy on the test set by each method were displayed in the barplots