Bameth: A Bilstm-Cross Attention Network for Methylation-Based Progression-Level Classification of Colon Adenocarcinoma

Colon adenocarcinoma (COAD) is a prevalent malignancy with high morbidity and mortality, largely due to its asymptomatic onset and late diagnosis. DNA methylation has emerged as a promising biomarker for cancer progression and molecular classification. To capture the stage-based methylation patterns, we propose BAMeth, a novel deep learning framework combining bidirectional Long ShortTerm Memory (BiLSTM) and cross-attention mechanisms. In this study, COAD samples were categorized into four types based on clinical stages and pathological characteristics, representing early (Type I), intermediate (Type II), late (Type III), and normal (Type IV) conditions. BAMeth performs multi-class classification across these stage-based categories using genome-wide methylation profiles. The architecture integrates dimensionality reduction, sequential modeling, and inter-feature dependency learning, enabling extraction of both sequential and contextual dependencies from highdimensional methylation data. Experimental results demonstrate that BAMeth achieves superior performance compared with some existing methods, particularly in terms of accuracy (ACC) by leveraging differentially methylated positions. Functional enrichment analysis of the genes annotated to the key CpG sites further reveals biological relevance in Wnt signaling and ubiquitin-mediated proteolysis pathways, highlighting its interpretability. Overall, BAMeth provides a robust and biologically interpretable framework for stage-based molecular classification of colon adenocarcinoma, offering potential for early detection and progression monitoring in epigenomics-driven oncology research.