Skip to content

pcr — PCRClock

Principal Component Regression clock using a scikit-learn Pipeline.

File: src/pcr.py
Test: unittests/test_pcr.py

Algorithm

  1. LOSO-CV model selection — fit StandardScaler → PCA(n) → LinearRegression for each candidate n_components. Select the n with highest LOSO-CV R² (ties broken by lowest MAE).

  2. Final model — refit the selected Pipeline on all Atlas samples.

  3. Prediction — transform query counts through the fitted Pipeline.

  4. Feature importance — gene_importance = loadings.T @ coef (equation 7 from the BayesAge2 paper).

Constructor parameters

Parameter Default Description
n_components_range [5, 10, 15, 20] Candidate component counts for LOSO-CV
top_n_var_genes None Pre-filter to top-N most variable genes

Methods

loso_cv(norm_counts, metadata)

LOSO-CV across all n_components_range values using cross_val_predict with LeaveOneOut. Selects and stores the optimal n_components (highest R², ties broken by lowest MAE).

Returns a DataFrame with columns age_days, pred_ncomp_5, pred_ncomp_10, … indexed by sample_id. CV metrics (MAE, R², Pearson r) per n_components are stored in self._cv_metrics.

Note: Mann-Whitney U (Young vs Old) is computed in CalibrationManager.run_pcr(), not here.

fit(norm_counts, metadata, n_components)

Fits the final Pipeline(StandardScaler → PCA → LinearRegression) on all data.

predict(norm_counts)

Transforms query counts through the fitted Pipeline and returns predicted tAge.

get_feature_importance()

Returns per-gene importance as loadings.T @ coef. Higher magnitude = stronger age-predictive signal.

save_loadings(out_dir)

Saves to out_dir:

  • per-component gene loadings
  • top genes per component
  • feature importance scores
  • LOSO-CV R² and MAE metrics

Output columns

PCRClock.loso_cv() returns (Atlas self-evaluation):

Column Description
age_days True age
pred_ncomp_5, pred_ncomp_10, … LOSO-CV predicted tAge per n_components

CalibrationManager.run_pcr() output ({tissue}_sexcombined_PCR_query.csv):

Column Description
sample_id Sample identifier
tAge_n5, tAge_n10, … Predicted tAge at each n_components
age_group Young / Old
condition Experimental condition

Design choice: data-driven model selection

The original notebook selects n_components by visual inspection of Mann-Whitney U results on the query set — involving the query data in model selection. This repo uses LOSO-CV R² on the Atlas training set only, which is a proper cross-validation procedure.