Group_Variable_Importance

Conda environment

The required packages are installed as a new conda environment including both R and Python dependencies with the following command:

conda env create -f requirements_conda.yml

⚠️ The use of mamba is faster and more stable for packages installation.

R environment

The missing R packages can be found in the "requirements_r.rda" file and can be downloaded using the following commands:

R

load("requirements_r.rda")

for (count in 1:length(installedpackages)) {
    install.packages(installedpackages[count])
}

⚠️ For reticulate, if asked for default python virtual environment, the answer should be no to take the default conda environment into consideration

For the First experiment:

Results computation with R script (compute_simulations):

• Set DEBUG to FALSE.
• N_SIMULATIONS is set to the range (1, 100)
• With N_CPU > 1, the parallel processing is used
• The list of methods contains (marginal, permfit, cpi, cpi_rf, gpfi, gopfi, dgi and goi).
• n_samples is set to 1000 and n_featues is set to 50
• rho_group lists all the correlation strengths in this experiment (0, 0.2, 0.5, 0.8)
• Number of permutations/samples n_perm is set to 100
• The output csv file is found in results/results_csv

Results plotting:

• Preparing csv files with R script plot_simulations_all under [AUC-type1error-power-time_bars]_blocks_100_grps.csv
• The plotting is done under plots/plot_figure_simulations_grps.ipynb with:
  • Figure 1 for the Figure 2 in the main text
  • Power + Time + Prediction scores for the Figure 6 in the supplement
  • Figure 1 Calibration for the Figure 5 in the supplement

For the Second experiment:

• We use compute_simulations_groups.
• The script can be launched with the following command:

  python -u compute_simulations_groups.py --n 1000 --pgrp 100 --nblocks 10 --intra 0.8 --inter 0.8 --conditional 1 --stacking 1 --f 1 --s 100 --njobs 1
  

  • --n stands for the number of samples (Default 1000)
  • --pgrp stands for number of variables per group (Default 100)
  • --nblocks stands for the number of blocks/groups in the data structure (Default 10)
  • --intra stands for the intra correlation inside the groups (Default 0.8)
  • --inter stands for the inter correlation between the groups (Default 0.8)
  • --conditional stands for the use of CPI (1) or PI (0)
  • --stacking stands for the use of stacking (1) or not (0)
  • --f stands for the first point of the range (Default 1)
  • --s stands for the step-size i.e. range size (Default 100)
  • --njobs stands for the serial/parallel implementation under Joblib (Default 1)
• The output csv file is found in results/results_csv under [AUC-type1error-power-time_bars]_blocks_100_groups_CPI_n_1000_p_1000_1::100_folds_2.csv
• The plotting is done under plots/plot_figure_simulations_grps.ipynb with Compare Stacking vs Non Stacking for the Figure 3 in the main text

For the Third experiment:

Scoring results 10-fold cross validated with process_var_groups_outer:

• The data are the public data from UKBB that needs to sign an agreement before using it (Any personal data are already removed)
• The biomarker is set by default to age
• n_jobs stands for serial/parallel computations
• k_fold_bbi stands for the number of folds for the internal cross validation of the method
• k_fold stands for the number of folds for train/test splitting the original data

Significance & Performance 10-fold cross validated with process_var_groups_outer_post:

• The $\underline{representative}$ p-value will be 2*median(p-values) across the 10 folds
• As for the $\underline{performance}$, it is measured on the 10% test set split per fold
• The output csv file is found in results/results_csv under Result_UKBB_age_all_imp_10_outer_2_inner_PERF.csv and Result_UKBB_age_all_imp_10_outer_2_inner_SIGN.csv
• The plotting is done under plots/plot_figure_simulations_grps.ipynb with Figure 3 for the Figure 4 in the main text