XAI-MethylMarker: Explainable AI approach for biomarker discovery for breast cancer subtype classification using methylation data. (1st September 2023)

• Record Type: Journal Article
• Authors: Rajpal, Sheetal; Rajpal, Ankit; Saggar, Arpita; Vaid, Ashok K.; Kumar, Virendra; Agarwal, Manoj; Kumar, Naveen
• Is Part Of: Expert systems with applications. Volume 225(2023)
• Publication Date: 2023-09-01
• Subjects: Explainable AI -- Methylation -- Biomarker -- Deep learning -- Breast cancer subtypes; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2023.120130
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Breast cancer—a heterogeneous disease marked with a high mortality rate, necessitates early diagnosis and treatment. The availability of multi-omic data has revolutionized our understanding of how molecular changes mark the variations in different breast cancer subtypes. Epigenomic changes in the form of DNA methylation differentially impact the expression level of the genes that play a vital role in the onset and spread of these subtypes. So, in this paper, we study the role of these variations in distinguishing between the various breast cancer subtypes. The cardinality of the existing biomarker sets is often too large to be interpreted clinically, and their relevance in classification remains unclear. In this paper, we propose a two-stage XAI-MethylMarker—an explainable AI-based biomarker discovery framework applied to DNA methylation data to arrive at a small set of biomarkers for breast cancer classification. In the first stage, we build a deep-learning network M e t h y l N e t that employs an autoencoder for dimensionality reduction and a feed-forward neural network to classify breast cancer subtypes. In the second stage, we propose a biomarker discovery algorithm, M e t h y l B D A, which employs different explainable techniques for analyzing M e t h y l N e t model and discovers a small set of 52 biomarkers. Using 5-fold cross-validation, we achieved a classification accuracy of 0.8145 ± 0.07 at a 95% confidence interval. To establish the clinical relevance of the discovered biomarkers, we performed a gene set analysis that revealed 14 druggable genes, nine genes linked to prognostic outcomes, and several enriched pathways are known to be significantly associated with distinct breast cancer subtypes.