elastic_net — ElasticNetClock

Elastic Net regularized linear regression clock.

File: src/elastic_net.py
Test: unittests/test_elastic_net.py

Algorithm

1. Hyperparameter search — GridSearchCV with leave-one-out CV over an alpha × l1_ratio grid. Each fold independently z-scales the training data (no leakage from held-out sample).

2. Final model — refit ElasticNet on all Atlas samples with best (alpha, l1_ratio) at max_iter=100,000.

3. LOSO-CV — leave-one-sample-out predictions on the Atlas training set, each fold with an independent StandardScaler.

4. Prediction — z-scale query counts using Atlas statistics; apply trained model.

Hyperparameter grid

Parameter	Values
alpha	[1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1, 10, 100]
l1_ratio	0.0, 0.1, 0.2, …, 1.0 (11 values)

Computation time

GridSearchCV with LOO-CV and 88 parameter combinations is computationally intensive (~minutes per tissue on a single CPU).

Constructor parameters

Parameter	Default	Description
top_n_var_genes	None	Pre-filter to top-N most variable genes before fitting

Methods

tune_and_train(norm_counts, metadata)

GridSearchCV over alpha × l1_ratio; fits final model on all data.

loso_cv(norm_counts, metadata)

Leave-one-sample-out CV predictions at best hyperparameters. Returns DataFrame with columns [sample_id, age_days, predicted_age].

predict(norm_counts)

Predicts tAge for new samples. Input must be DESeq2-normalized.

save(out_dir)

Saves best hyperparameters and non-zero gene coefficients to out_dir.

Output columns

{tissue}_sexcombined_EN_query_loso.csv:

Column	Description
sample_id	Sample identifier
age_days	True age (Atlas LOSO samples only)
predicted_age	Predicted tAge
source	Atlas or Query

Design choices

No scaler leakage

The original Costa et al. notebook fits StandardScaler on the full Atlas before calling GridSearchCV, leaking held-out statistics. This repo fits the scaler independently per fold inside each CV split.

top_n_var_genes

For Brain tissue, the large number of samples makes GridSearchCV slow. Setting top_n_var_genes=5000 provides a practical speed-up with minimal accuracy loss. The default (None) uses all genes.