Computational Pipeline

To assess compositional rearrangements of protein complexes as opposed to their overall abundance changes, a module-wise normalization was performed, as previously described [Ori et al., 2013; Ori et al., 2016]. Proteins belonging to the same complex were normalized by the respective trimmed mean (or interquartile mean) of the complex subunits across all individuals/samples. In case of proteins involved in multiple complexes, the average value from all the corresponding complexes was taken into account. Given the complex-normalized abundances, the variance of each subunit in a given complex was calculated. To compare these variances between different proteomics datasets and approaches, those variances were converted to z-scores per complex (Figure 4). Similarly to testing the consistency between datasets in the above section, we calculated the correlation coefficients (between datasets) for each such a z-score matrix. To compile a reference distribution we permuted each matrix and calculated corresponding correlation coefficients 1000 times, which provides a normal distribution. In a two-sided t-test we then compared the real distribution of correlation values with the ones derived from the random permutations of the dataset. Protein subunits within a complex were considered ‘stable’ or ‘variable’ in case of the associated p-value < 0.05 based on the distribution of z-scores (Table S3). To see whether a given protein is consistently ‘variable’ in a complex throughout all given datasets, the distribution of its z-scores within the complex and across all the datasets were compared to the z-score distribution for all other protein components of the same complex across all datasets (one-sided t-test). A one-sided t-test accounts for the unidirectionality of our hypothesis and gives conservative results. This procedure was done for all proteins in a given complex, and resulting p-values were adjusted using the Benjamini-Hochberg method [Benjamini and Hochberg, 1995].