Subclassification of Breast Cancer through Comprehensive Multi-omics Data to Benefit Distinct Survival Outcomes

Abstract Breast cancer is a heterogeneous complex of disease consisting of several subtypes which display different biological and clinical behaviors. Traditional classification methods regarding histological types, tumor size and histological grade have limitations in classification of breast cancer subtypes. In the past decade, the development of high-throughput technologies in omics allowed us to discover breast cancer’s molecular subtypes and biomarkers. In this study, different feature selection techniques consist of t-test, least absolute shrinkage and selection operator (LASSO) and Boruta were applied to select the most significant features from the omics data, including exon expression RNA-seq, methylation450k and gene expression RNA-seq collected from The Cancer Genome Atlas (TCGA). The significant features being selected were used as inputs for machine learning classifiers for breast cancer classification. We evaluated five different classification models including random forest (RF), support vector machine (SVM), naive Bayes (NB), k-Nearest Neighbor (KNN), and multivariate adaptive regression splines (MARS). The results show that the performances of the combined model based on exon expression and gene expression RNA-seq were better than other classification methods in terms of accuracy. Interestingly, our partitioning of luminal samples by SVM, NB and MARS would have a clinical advantage over the PAM50 partition of the luminal samples though their predictions only moderately agree with PAM50 calls. Based on the methylation data, the screened four methylation-driven genes (PRDM16, ADPRHL1, SPRY1, and TBCC) were identified to associate with tumorigenesis. Moreover, survival analysis showed DNA methylation profiles improved prediction of survival outcomes for luminal A and luminal B subtypes. Our results suggested that RNA-seq, exon expression and methylation hold complementary information for classifying breast cancer subtypes. Analysis of samples using methylation data demonstrates that the sample partitions predicted by SVM, NB and MARS show a higher relation with methylation patterns. We delineated genomic epigenomic characteristic for breast cancer subtypes as well as their specific survival. Compared with the PAM50 standard, our improved and refined feature selection and classification provide diagnostic markers and may contribute to the precision of diagnosis and thus, to more personalized treatment.