DNA methylation biomarkers-based pan-cancer classifier: predictive modeling for cancer classification.

Area of Science:

• Molecular diagnostics
• Bioinformatics
• Oncology

Background:

• Machine learning (ML) in diagnostics faces challenges with inflated performance assessments.
• Methodological issues hinder the reliable implementation of ML in personalized medicine.
• A robust framework is needed to address these challenges in omics-driven diagnostics.

Purpose of the Study:

• To develop and validate a pan-cancer classification framework using DNA methylation data.
• To address methodological challenges in omics data-powered ML for diagnostics.
• To create a reliable tool for oncological diagnosis.

Main Methods:

• Curated large-scale DNA methylation datasets (10,756 primary, 2,306 validation samples).
• Employed a custom biomarker selection strategy based on effect size, considering variance and class imbalance.
• Utilized nested cross-validation for ML model training, tuning, and evaluation.
• Integrated Local Outlier Factor for anomaly detection and filtering in samples.
• Validated the framework using methylation profiles for 3,905 central nervous system (CNS) tumors.

Main Results:

• Simple ML models, specifically logistic regression, outperformed complex algorithms like deep neural networks.
• A logistic regression classifier achieved a balanced accuracy (BACC) of 0.90 for 54 cancer/healthy tissue types using 1208 CpG sites.
• The CNS tumor classifier, also logistic regression-based, reached a BACC of 0.94 across 59 CNS tumor subtypes.
• Anomaly filtering enhanced performance across all tested sample categories.

Conclusions:

• DNA methylation profiling combined with controlled ML practices enables robust oncological diagnostic solutions.
• The developed framework can significantly increase the efficacy of cancer diagnosis.
• Inference pipelines are publicly accessible via a web platform (https://opp.pum.edu.pl/).