Li · IEEE/ACM transactions on computational biology and bioinformatics 2022 · machine learning model development and validation study · n=?

Age Prediction by DNA Methylation in Neural Networks.

Cited 13 times in the scientific literature.

Level 4 - case-series / case-control

Level 4 by design analogy (cross-sectional computational model development and validation using retrospective genomic data).

PubMed 34048347 · doi:10.1109/TCBB.2021.3084596 · record verified 2026-08-30

What was done

The authors developed the Correlation Pre-Filtered Neural Network (CPFNN), a framework using Spearman correlation to filter high-dimensional CpG features prior to neural network training. They compared CPFNN's age prediction accuracy against standard linear epigenetic clocks (Horvath's and Hannum's formulas), regularized neural networks (LASSO and elastic net), and dropout neural networks. They also evaluated associations between predicted epigenetic age and disease states (Schizophrenia and Down Syndrome).

What was found

CPFNN predicted chronological age with a Mean Absolute Error (MAE) of 2.7 years, outperforming the benchmark linear and regularized models by at least 1 year in MAE. Epigenetic age was significantly associated with Schizophrenia (p = 0.024); the specific numerical p-value for Down Syndrome was omitted in the abstract text.

Why it matters

Pre-filtering high-dimensional methylation features addresses overfitting in deep learning architectures, improving non-linear biological age prediction over standard linear formulas.

Limits

The abstract does not state the sample size (n), data sources, tissue types, or demographic characteristics, and omits the reported p-value for Down Syndrome. Generalizability to external independent cohorts cannot be determined from the abstract alone.

Cited by