Guidelines & Consensus
meth-SemiCancer2 Enhances DNA Methylation Cancer Subtyping Across Cohorts
Choi & ChaeBMC Bioinformatics
Precise molecular subtyping is critical to guide targeted oncology therapeutics, early diagnosis, and prognosis, yet translating DNA methylation profiles into reproducible clinical subtypes is frequently impeded by high dimensionality, batch effects, and a paucity of labeled datasets. Technical noise across distinct patient cohorts can obscure underlying biological variation, generating spurious subtype boundaries and impeding reproducibility.
To address cross-cohort batch variations and feature representation challenges, researchers engineered meth-SemiCancer2. The computational framework incorporates a multi-phase design: initial pretraining on labeled source datasets to identify subtype signatures, adversarial training to harmonize distribution discrepancies between source and unlabeled target cohorts, contrastive learning to isolate domain-invariant features, and semi-supervised fine-tuning with subtype alignment to iteratively refine pseudo-labels.
Evaluations spanned seven cancer types: breast cancer, colon adenocarcinoma, prostate adenocarcinoma, glioblastoma, renal cell carcinoma, thyroid carcinoma, and low-grade glioma. Through 10-fold cross-validation on source data and validation on independent target cohorts, meth-SemiCancer2 demonstrated superior predictive accuracy and robust cluster delineation compared to conventional machine learning classifiers, its predecessor meth-SemiCancer, benchmark domain adaptation methods, and existing DNA methylation batch-correction pipelines. The model also maintained robust classification performance under data-constrained conditions, confirming its utility when only fractional source or target profiles are available.
Key Takeaways
- Across seven evaluated malignancies—breast cancer, colon adenocarcinoma, prostate adenocarcinoma, glioblastoma, renal cell carcinoma, thyroid carcinoma, and low-grade glioma—meth-SemiCancer2 consistently outperformed predecessor meth-SemiCancer, conventional machine learning classifiers, two state-of-the-art batch correction tools, and alternative deep learning models in subtype prediction accuracy.
- In cross-cohort evaluations, the framework generalized effectively to unlabeled target datasets to yield distinct subtype clusters and sustained stable classification performance even when restricted to fractional subsets of source or target data.
Read the original sourceBMC Bioinformatics
