An explainable-AI framework reveals novel lncRNAs specific for breast cancer subtypes

Area of Science:

• Genomics
• Bioinformatics
• Cancer Biology

Background:

• Long non-coding RNAs (lncRNAs) are increasingly recognized as crucial regulators in cancer development.
• Their utility in large-scale, multi-class cancer subtyping remains under-investigated, often due to data limitations or their auxiliary role in classification.

Purpose of the Study:

• To explore the potential of lncRNAs for multi-class breast cancer (BRCA) subtyping using an explainable artificial intelligence (AI) framework.
• To evaluate the performance of machine learning (ML) models built with lncRNA, mRNA, and miRNA features, individually and in combination.

Main Methods:

• Utilized 7,177 lncRNAs from 1,021 BRCA transcriptomics datasets for subtyping.
• Employed four ML classifiers (Naïve Bayes, Random Forest, ANN, XGBoost) to assess classification performance.
• Applied a sequential feature identification pipeline (ANOVA, Boruta, SHAP) to identify subtype-specific biomarkers.

Main Results:

• XGBoost with lncRNAs alone achieved 89.2% accuracy and 0.99 AUROC.
• Integrating miRNA or mRNA features marginally improved accuracy to 90.8% and 92.2%, respectively; combined features offered no additional benefit.
• Identified 119, 66, 54, and 24 unique subtype-specific lncRNA biomarkers for Luminal A, Luminal B, HER2+, and Basal subtypes, respectively.
• Discovered novel subtype-specific lncRNAs with prognostic significance through survival analysis.

Conclusions:

• lncRNAs possess significant diagnostic and biomarker discovery potential for breast cancer.
• The implemented explainable AI framework offers a systematic approach for evaluating lncRNA-based models for multi-class cancer subtyping.
• This methodology can be adapted for subtyping other cancers using publicly available transcriptomics data.