Background/Objectives: DNA methylation is a key epigenetic process that regulates gene expression and is often disrupted in colorectal cancer (CRC). Aberrant methylation of promoter CpG islands can silence tumor suppressor genes and drive tumorigenesis. A subset of CRCs exhibits the CpG Island Methylator Phenotype (CIMP), characterized by widespread hypermethylation and distinct clinical outcomes. Identifying genes whose expression is epigenetically regulated by methylation is important for prioritizing candidate biomarkers and therapeutic targets in CRC. Methods: We developed and compared a series of computational approaches to identify genes whose expression is regulated by DNA methylation in The Cancer Genome Atlas (TCGA) cohort of Colon Adenocarcinoma (COAD) patients. Samples were stratified according to their CpG Island Methylator Phenotype (CIMP) level to capture distinct epigenetic subgroups. The proposed framework integrates methylation and transcriptomic data to systematically detect methylation–expression associations indicative of epigenetic regulation. Results: The best-performing method identified gene sets strongly associated with promoter methylation–expression relationships and enriched for pathways relevant to colorectal cancer progression and patient stratification. To evaluate the robustness and transferability of the approach, it was further validated on independent datasets, including Stomach Adenocarcinoma (STAD), Glioblastoma Multiforme (GBM), and Mesothelioma (MESO), supporting its robustness and potential generalizability across multiple tumor types. Conclusions: Our study highlights the potential of computational pipelines to uncover epigenetically regulated genes in colorectal cancer. The identified candidate genes provide a hypothesis-generating foundation for refining molecular stratification and guiding future studies aimed at epigenetic biomarker discovery and therapeutic hypothesis development.

Objectives: Precision oncology relies on accurate interpretation of tumour-detected gene variants, to guide personalised treatment decisions. However, interpretation of variants in context requires extensive information that is often buried within unstructured biomedical literature and obscured by inconsistent nomenclature, making manual retrieval labour-intensive and prone to omissions. In this study, our aim was to evaluate the potential of large language models to extract variant-treatment-cancer associations at scale from published abstracts. Design: We developed Variantscape , a large-scale, automated pipeline and open-access web tool. It integrates traditional natural language processing methods with state-of-the-art large language models to extract, standardise, and analyse co-associations between genetic variants, cancer types, and therapeutic interventions from published biomedical abstracts. Results: From over 3.7 million abstracts published between January 2014 and October 2026, 886,195 articles containing gene names were eligible for extraction. Among these, 37,895 (4.3%) mentioned a variant, cancer type, and therapeutic agent together, covering 11,985 unique variants across 197 cancer types and 547 therapeutic agents. LLM verification against the abstract text confirmed at least one association in 35,224 papers, giving a literature-derived graph of 10,571 nodes and 44,644 verified edges. Canonical alterations in well-studied cancers (e.g., BRAF V600E in melanoma) were strongly linked to established treatments, while several rare variants also emerged with high-confidence literature support. Conclusions: By applying large language models to biomedical literature, Variantscape enables scalable, context-aware extraction of trilateral variant-treatment-cancer relationships. This approach supports early evidence synthesis/hypothesis generation, highlights underrecognized or rare associations for MTB preparation, and offers a practical resource for accelerating discovery and supporting precision oncology research and translation. Funding: This study was funded by the School of Medicine at the University of St. Gallen in Switzerland under grant number 2300380. Article Summary: Strengths and limitations of this study: ● Variantscape combines traditional natural language processing with large language models to enable scalable, automated extraction of variant-cancer-treatment relationships from over 3.7 million biomedical abstracts, substantially reducing the manual search burden inherent to existing molecular tumour board workflows. ● The pipeline applies large language model-based inference to standardise inconsistent genetic nomenclature across the literature, improving the comparability and reliability of extracted associations that would otherwise be obscured by terminological variation. ● Variantscape is designed as a continuously updatable, open-access resource, meaning it can incorporate newly published evidence without requiring repeated manual intervention. ● The pipeline relies solely on abstract-level text rather than full-text articles, which may limit the depth and granularity of extracted associations, as key methodological details, variant context, and nuanced clinical findings are frequently reported only within the body of published papers. ● Automated extraction using large language models introduces a risk of misattribution of variant-cancer-treatment relationships, and without systematic manual validation of outputs, the precision of identified associations at scale remains difficult to fully characterise.