Introduction
Breast cancer is a highly heterogeneous disease, where patients with similar clinical stages often exhibit vastly different treatment responses and survival outcomes. Traditionally, transcriptomic profiling for molecular subtyping and prognosis has focused on messenger RNA (mRNA). However, the non-coding portion of the genome, specifically microRNAs (miRNAs) and long non-coding RNAs (lncRNAs), has emerged as a critical regulator of gene expression and cancer progression. Despite numerous studies proposing these as superior biomarkers, a systematic evaluation of their added value over established mRNA-based clinical tools has been lacking. This study benchmarks 35 computational methods across 19 datasets to determine the true impact of non-coding RNAs on breast cancer characterisation.
Methodology
The researchers evaluated a comprehensive suite of 35 computational methods, categorized into subtyping and prognosis models.
Computational Methods
The 12 subtyping methods included concatenation-based, similarity-based (e.g., SNF, CIMLR), and model-based approaches (e.g., iCluster). The 23 prognosis methods spanned traditional gene-based signatures (e.g., OncotypeDX, GENE70) and newer models specifically designed for miRNA or lncRNA data. To facilitate this large-scale comparison, the authors developed the CancerSubtypesPrognosis R package, which standardizes data pre-processing, feature selection, and evaluation metrics like the Silhouette score and C-index.
Data Sources
The study utilized 19 independent breast cancer datasets, totaling 5,134 patients. These datasets were sourced from major repositories including The Cancer Genome Atlas (TCGA), METABRIC, and the Gene Expression Omnibus (GEO). The data types varied across cohorts, with some providing matched multi-omic data (mRNA, miRNA, and lncRNA) and others focusing on specific RNA types.
Key Insights and Results
The Value of miRNA in Subtyping
A primary finding of this benchmark is that integrating miRNA data significantly enhances the accuracy of breast cancer subtyping. Methods that combined miRNA and mRNA expression (such as SNF and CNMF) identified patient clusters with more distinct survival patterns than those identified using mRNA alone. Among the subtyping tools, CIMLR (Cancer Iterative Multi-view Learning) demonstrated the highest consistency and internal validity across different data types.
Limitations of lncRNA and Current Prognosis Models
In contrast to miRNAs, lncRNAs did not provide a significant boost to subtyping or prognosis in this study. This may be attributed to the limitations of older microarray technologies, which capture only a fraction of the lncRNA landscape compared to modern RNA-sequencing.
Regarding prognosis, traditional multi-gene signatures remain the most reliable. OncotypeDX achieved the highest mean C-index (0.65), outperforming models based solely on non-coding RNAs. Furthermore, the study highlighted a "concordance crisis": 38.3% of prognosis method pairs showed no effective agreement on individual patient risk assessments (Cohen’s kappa < 0), suggesting that different models often flag different patients as high-risk despite similar overall accuracy.
Conclusion and Future Directions
While non-coding RNAs—particularly miRNAs—show great promise for refining cancer subtyping, they have yet to surpass established mRNA signatures for survival prediction. Future progress will likely depend on:
- Higher Resolution Data: Transitioning from microarrays to RNA-sequencing to better capture the lncRNA landscape.
- Advanced Modeling: Utilizing Deep Learning to account for non-linear interactions between different RNA species.
- Biological Integration: Incorporating known regulatory networks (e.g., miRNA-mRNA interactions) into computational models to improve predictive power and clinical utility.
Summary updated