Tumor-specific total mRNA expression, or TmS, is a measure of the transcriptional activity of tumor cells compared to the surrounding non-tumor cells that has been proven to be an effective biomarker for patient prognosis and cancer behavior, as well as a tool to quantify intra- and inter-tumor heterogeneity. However, calculating TmS for individual patients would not be a viable strategy for clinical decision-making as it requires processing RNA and DNA sequencing data from the same tumor through intensive computational tools. Therefore, the development of a machine learning model that can predict TmS from raw data could make it more readily applicable in clinical settings. In this study, we developed a machine learning model that can accurately predict TmS across various breast cancer subtypes based on several sources of raw data. Additionally, to evaluate the clinical utility of the predicted TmS values, we conducted survival analyses using the predicted TmS values to divide the sample population into high and low TmS subgroups. We found that the optimal combination of raw data sources for TmS prediction includes gene expression and DNA methylation data; models trained on these data achieved good performance even when data for a small fraction of all genes and methylation sites were included. Furthermore, the survival analysis results reveal the clinical significance of the TmS predictions, as they are consistent with results obtained using real TmS values.
