Large-Scale Machine Learning Analysis Reveals DNA Methylation and Gene Expression Response Signatures for...

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Gemcitabine is a primary chemotherapy for pancreatic adenocarcinoma (PAAD).
• Many PAAD patients exhibit resistance to gemcitabine, necessitating predictive biomarkers.
• Current predictors of gemcitabine response in PAAD are limited and do not leverage machine learning.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting PAAD patient response to gemcitabine.
• To identify molecular features that can distinguish responders from nonresponders to gemcitabine treatment.
• To assess the performance of machine learning models compared to existing biomarkers like hENT1.

Main Methods:

• Collected diverse molecular profiles and clinical data for PAAD patients from the Genomic Data Commons.
• Systematically combined 8 tumor profiles with 16 classification algorithms, generating 128 machine learning models.
• Evaluated each model using multiple 10-fold cross-validations, assessing predictive performance metrics like Matthews correlation coefficient (MCC) and ROC-AUC.

Main Results:

• Seven out of 128 machine learning models demonstrated predictive capability.
• A random forest model using 4 mRNAs achieved an MCC of 0.44 and ROC-AUC of 0.785.
• An XGBoost model utilizing 12 DNA methylation probes achieved an MCC of 0.32 and ROC-AUC of 0.697.
• These machine learning models also showed potential in predicting patient prognosis (overall and progression-free survival).

Conclusions:

• Developed novel, predictive machine learning models for gemcitabine response in PAAD.
• These models outperform traditional markers like hENT1.
• The models are released for prospective validation in independent cohorts of gemcitabine-treated PAAD patients.