Machine learning workflows to estimate class probabilities for precision cancer diagnostics on DNA methylation...

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Genomics

Background:

• DNA methylation profiling is a key technology for molecular tumor classification.
• Standardized methods for well-calibrated probability estimates in complex cancer diagnostics are needed.

Purpose of the Study:

• To evaluate and compare machine learning (ML) workflows for DNA methylation-based cancer diagnostics.
• To identify optimal ML strategies for unbiased class probability (CP) estimation.

Main Methods:

• Evaluated random forests (RFs), elastic net (ELNET), support vector machines (SVMs), and boosted trees with post-processing calibration.
• Utilized ridge-penalized multinomial logistic regression (MR) and Platt scaling for calibration.
• Compared workflows on a 2,801-sample brain tumor cohort and The Cancer Genome Atlas data using 5x5-fold cross-validation.

Main Results:

• ELNET performed best as a stand-alone classifier with superior calibration.
• The optimal two-stage workflow involved MR-calibrated SVM (linear kernel), followed by ridge-calibrated RF.
• MR proved to be the most effective calibration method across classifiers.

Conclusions:

• Developed protocols guide the selection and tuning of ML workflows for reliable CP estimates in precision diagnostics.
• The findings support the use of DNA methylation data and optimized ML for advanced cancer classification.
• Open-source R scripts are available to facilitate implementation for researchers.