cfMethylPre: deep transfer learning enhances cancer detection based on circulating cell-free DNA methylation...

Area of Science:

• Biotechnology
• Genomics
• Computational Biology

Background:

• Cancer diagnosis relies on early detection for improved patient outcomes.
• Circulating cell-free DNA (cfDNA) methylation is a promising noninvasive biomarker.
• Current methods struggle with high-dimensional methylation data, small sample sizes, and interpretability.

Purpose of the Study:

• To develop a novel deep transfer learning framework, cfMethylPre, for accurate cancer detection using cfDNA methylation.
• To leverage large language model embeddings for enhanced feature representation in methylation data.
• To identify novel cancer-associated genes through interpretable model analysis.

Main Methods:

• Developed cfMethylPre, a deep transfer learning framework integrating DNA sequence embeddings and methylation profiles.
• Pretrained the model on 2801 bulk DNA methylation samples across 82 cancer types and normal controls.
• Fine-tuned the model on cfDNA methylation data for cancer detection.

Main Results:

• Achieved superior predictive accuracy with a weighted Matthews Correlation Coefficient of 0.926 and a weighted F1-score of 0.942.
• Identified three novel breast cancer genes (PCDHA10, PRICKLE2, PRTG) with inhibitory effects on cell proliferation and migration.
• Demonstrated the model's interpretability and biological validation of identified genes.

Conclusions:

• cfMethylPre offers a powerful and interpretable tool for noninvasive cancer diagnostics.
• The framework enhances feature representation and predictive accuracy for cfDNA methylation analysis.
• Identified novel genes provide new targets for breast cancer research and precision oncology.