Age Prediction by DNA Methylation in Neural Networks

Area of Science:

• Epigenetics
• Computational Biology
• Bioinformatics

Background:

• Aging is influenced by complex factors, including DNA methylation.
• Traditional linear models for DNA methylation aging have limitations.
• Neural networks can capture non-linear relationships but struggle with high-dimensional data and overfitting.

Purpose of the Study:

• To develop a novel neural network approach for more accurate epigenetic age prediction.
• To address the challenges of overfitting and poor generalization in neural networks using large-scale DNA methylation datasets.
• To evaluate the performance of the proposed method against existing statistical and machine learning models.

Main Methods:

• Proposed Correlation Pre-Filtered Neural Network (CPFNN) utilizing Spearman Correlation for feature pre-filtering.
• Compared CPFNN with linear regression models (Horvath's, Hannum's), regularized neural networks (LASSO, Elastic Net), and Dropout Neural Networks.
• Assessed model performance using Mean Absolute Error (MAE).

Main Results:

• CPFNN achieved a Mean Absolute Error (MAE) of 2.7 years, outperforming other models by at least 1 year.
• The study identified a significant association between epigenetic age and Schizophrenia (p=0.024).
• Feature weighting based on correlation with the outcome is crucial for improving prediction accuracy with high-dimensional data.

Conclusions:

• CPFNN offers a superior approach for predicting epigenetic age from DNA methylation data.
• Accurate epigenetic age prediction can be a valuable tool for understanding age-related diseases.
• Pre-filtering features based on their correlation with the outcome is a key strategy for enhancing machine learning model performance in bioinformatics.