Sequential analysis of transcript expression patterns improves survival prediction in multiple cancers

BACKGROUND: Long-term survival in numerous cancers often correlates with specific whole transcriptome profiles or the expression patterns of smaller numbers of transcripts. In some instances, these are better predictors of survival than are standard classification methods such as clinical stage or hormone receptor status in breast cancer. Here, we have used the method of “t-distributed stochastic neighbor embedding” (t-SNE) to show that, collectively, the expression patterns of small numbers of functionally-related transcripts from fifteen cancer pathways correlate with long-term survival in the vast majority of tumor types from The Cancer Genome Atlas (TCGA). We then ask whether the sequential application of t-SNE using the transcripts from a second pathway improves predictive capability or whether t-SNE can be used to refine the initial predictive power of whole transcriptome profiling. METHODS: RNAseq data from 10,227 tumors in TCGA were previously analyzed using t-SNE-based clustering of 362 transcripts comprising 15 distinct cancer-related pathways. After showing that certain clusters were associated with differential survival, each relevant cluster was re-analyzed by t-SNE with a second pathway’s transcripts. Alternatively, groups with differential survival identified by whole transcriptome profiling were subject to a second, t-SNE-based analysis. RESULTS: Sequential analyses employing either t-SNE➔t-SNE or whole transcriptome profiling➔t-SNE analyses were in many cases superior to either individual method at predicting long-term survival. We developed a dynamic and intuitive R Shiny web application to explore the t-SNE based transcriptome clustering and survival analysis across all TCGA cancers and all 15 cancer-related pathways in this analysis. This application provides a simple interface to select specific t-SNE clusters and analyze survival predictability using both individual or sequential approaches. The user can recreate the relationships described in this analysis and further explore many different cancer, pathway, and cluster combinations. Non-R users can access the application on the web at https://chpupsom19.shinyapps.io/Survival_Analysis_tsne_umap_TCGA. The application, R scripts performing survival analysis, and t-SNE clustering results of TCGA expression data can be accessed on GitHub enabling users to download and run the application locally with ease (https://github.com/RavulaPitt/Sequential-t-SNE/). CONCLUSIONS: The long-term survival of patients correlated with expression patterns of 362 transcripts from 15 cancer-related pathways. In numerous cases, however, survival could be further improved when the cohorts were re-analyzed using iterative t-SNE clustering or when t-SNE clustering was applied to cohorts initially segregated by whole transcriptome-based hierarchical clustering.