Machine learning-based single-sample molecular classifier for cancer grading.

Tumor subtyping based on morphological grade is used in cancer treatment and management decision-making and to determine a patient's prognosis. While low- and high-grade tumors are predictive of patient survival for many cancers, tumors of intermediate morphological grades are considered unreliable due to interobserver variability and thus do not have clear prognostic significance. To address this issue, we devised a molecular-based classifier that uses gene expression data from RNA sequencing (RNA-seq) or microarray profiling to predict high- and low-grade risk groups for breast, lung, and renal cancers. For this classifier, we developed a preprocessing procedure that only required expression data from a single sample, without the need for any batch correction or cohort scaling. This classifier, while trained only on RNA sequencing data, achieves highly accurate risk predictions on both RNA-seq and microarray data. First, the molecular grades (mGrades) predicted by this classifier correlated strongly with the pathologist-assigned histological grades and clinical stage. Next, we showed that mGrades were effective in assessing risk levels for G2 samples. Finally, we identified common and unique biological and genetic features in samples of low and high mGrades across breast, lung, and renal cancers. Gene expression patterns as revealed by the classifier can provide useful information for both research and diagnostic purposes.