Computational Pipeline

To understand to what extent both protein complex abundance and stoichiometry are affected by either sex or diet, a L2-regularized Multiple Linear Regression (Ridge regression with a regularization parameter of 1) was used, as implemented in the scikit Python package (http://scikit-learn.org; Pedregosa et al., 2011). We compared models that predict complex abundance or complex stoichiometries, using as predictors: (i) genetic sex, (ii) diet, and (iii) the combination of genetic sex and diet together. We assessed the quality of each model by the coefficient of determination (R²). This was done for every module considered (complexes and pathways), for abundance, as well as module-normalized data. For pathways we only considered those that were showing a high co-abundance (FDR-corrected p-value < 0.1) as compared to co-abundances derived from a reshuffled dataset.

To estimate prediction performance we used a 10-fold cross-validation scheme. Briefly, we randomly separated the dataset (per complex) into ten groups of equal size, in order to iteratively train a model with nine of them, and to assess the testing performance in the held-out group. For each module the median global R² is reported. The same analysis was conducted with a reshuffled dataset per complex, and the corresponding performance metrics were used in a permutation test approach to assign significance to the true ridge regression coefficients. Specifically, we use the latter distribution to calculate an empirical FDR. Throughout the main text, the global R² performance metric derived for the module or protein is reported as the effect size with its respective FDR-corrected p-value.