Closing the Gap: Epigenetic Age Prediction with ElasticNet and Public Data

Read the Paper (PDF)

Predicted vs. actual age on the held-out EPIC v2 test set (GSE246337, n=500). Left to right: HistGBR (MAE 3.79), ElasticNet (MAE 3.37), DeepStrataAge (MAE 2.45).

---

Overview

This repository contains the code and results for building an epigenetic age clock from publicly available DNA methylation data. We train an ElasticNet model on 11,328 samples from 17 GEO cohorts and evaluate on an independent Illumina EPIC v2 test set, achieving MAE 3.37 years — closing 53% of the performance gap to DeepStrataAge (De Lima Camillo et al., npj Aging, 2026), a 40.7M-parameter DNN trained on 29,167 proprietary samples.

We show that the remaining gap is attributable to data volume, not model architecture: DeepStrataAge's own ElasticNet baseline achieves MAE 2.02 on their data — only 0.13 years behind the full DNN.

Results

Model	Training Data	Test MAE	MedAE	R²	Pearson r
DeepStrataAge (DNN, 40.7M params)	29,167 (proprietary)	2.45	1.79	0.959	0.980
ElasticNet (this work)	11,328 (public)	3.37	2.23	0.937	0.977
HistGBR (this work)	11,328 (public)	3.79	2.77	0.926	0.965

Test set: GSE246337 (500 samples, EPIC v2, age 18–89). Held out during training and hyperparameter tuning. 95% CI for ElasticNet: [3.03, 3.77].

Scaling

Train N	Cohorts	Test MAE	Gap to DSA
3,140	7	4.38	1.93
5,329	9	3.57	1.12
8,722	11	3.52	1.06
11,328	17	3.37	0.91

Method

• Cross-validation: Leave-One-Cohort-Out (LOCO) across 17 cohorts. Random CV inflated results by ~50%.
• Tuning: Optuna TPE with fold-level pruning (60 trials, 40% subsampling). Test set excluded by assertion.
• Features: 12,234 CpGs + sex + platform flag + 5 batch features. Population-mean imputation for cross-platform CpGs.
• Regularization: Optimal l1_ratio shifted from 0.06 (near-Ridge) at 5K samples to 0.45 (true ElasticNet) at 11K — sparsity helps with heterogeneous cohorts.

Data Quality

Three bugs in public GEO datasets caused initial MAE > 24. Fixing them was the largest single improvement:

Issue	Dataset	Effect
Detection p-values interleaved with betas	GSE196696	Mean beta collapsed to 0.13
G-CSF mobilized transplant donors	GSE196696	+22 years epigenetic age shift
Sentrix / GSM ID order mismatch	GSE246337	All age labels shuffled

Training Data

All data is publicly available from GEO.

17 training cohorts (click to expand)

GEO ID	N	Platform	Age	Description
GSE40279	656	450K	19–101	Hannum et al. (2013)
GSE41169	95	450K	18–65	Dutch population
GSE42861	335	450K	20–70	EIRA RA study (controls)
GSE50660	464	450K	38–67	UK Airwave
GSE51032	845	450K	34–72	EPIC-Italy
GSE51057	329	450K	35–70	Menarcheal timing
GSE55763	2,664	450K	24–75	LOLIPOP UK
GSE56105	614	450K	10–75	BSGS (Australia)
GSE72773	310	450K	35–92	Horvath multi-ethnic
GSE72775	335	450K	36–90	Horvath Hispanic
GSE73103	355	450K	14–34	Swedish young adults
GSE87571	729	450K	14–94	Swedish lifespan
GSE106648	139	450K	20–65	MS study (controls)
GSE125105	699	450K	17–87	MPI Psychiatry
GSE132203	795	EPIC v1	18–76	GTP cohort
GSE196696	570	EPIC v1	19–40	CIBMTR
GSE210255	1,394	EPIC v1	21–87	Multi-ethnic

Test: GSE246337 (500 samples, EPIC v2, age 18–89)

Reproduce

Requirements

python >= 3.11
numpy  pandas  scikit-learn  scipy  optuna  torch  matplotlib

Steps

# 1. Download DeepStrataAge weights
#    https://zenodo.org/records/14768481 → data/zenodo/DeepStratv3/

# 2. Download and parse GEO datasets into data/processed/*.parquet
#    (see training data table for accessions)

# 3. Hyperparameter tuning (optional — tuned params included in tuning/)
python tune_optuna.py --model elasticnet --n-trials 60 --subsample 0.4

# 4. Evaluate
python run_definitive_tuned.py

Repository Structure

├── paper/
│   ├── paper.tex                   # Manuscript source (LaTeX)
│   └── paper.pdf                   # Compiled PDF
├── run_definitive_tuned.py         # Main pipeline (tuned params)
├── run_definitive.py               # Pipeline (default params)
├── tune_optuna.py                  # Hyperparameter search
├── src/                            # Utilities (imputation, evaluation, plots)
├── tuning/                         # Best hyperparameters (JSON)
└── results/                        # Final CSVs and figures

Citation

@misc{elouahabi2026epigenetic,
  author       = {El Ouahabi, Maher},
  title        = {Closing the Gap: Epigenetic Age Prediction with ElasticNet and Public Data},
  year         = {2026},
  howpublished = {\url{https://github.com/maher-coder/epigenetic-clock}},
}

References

• De Lima Camillo et al. (2026). DeepStrataAge. npj Aging. DOI: 10.1038/s41514-026-00358-w
• Horvath (2013). DNA methylation age of human tissues and cell types. Genome Biology, 14(10), R115.
• Hannum et al. (2013). Genome-wide methylation profiles reveal quantitative views of human aging rates. Molecular Cell, 49(2), 359–367.

License

MIT