Protein prediction models support widespread post-transcriptional regulation of protein abundance by interacting partners

Author summary The abundance of mRNA is often measured as a surrogate variable of protein levels, but how well the mRNA level of different genes correlate with their protein across samples remains incompletely understood. Here we trained machine learning models over large RNA sequencing and mass spectrometry data from up to 8 cancer types in the CPTAC data sets to evaluate how well protein level variances across samples can be predicted from their transcripts. Despite voluminous data, up to one-third of genes shows poor mRNA-protein correlation suggesting their protein abundance is not primarily regulated from cognate transcripts. The inclusion of mRNA level information from protein interaction partners into the prediction models substantially improved prediction performance for a subset of genes, suggesting their protein abundance may be primarily regulated post-transcriptionally through protein-protein interactions. Notably, these proteins involve not only subunits of large multi-protein complexes such as the ribosome as previously suspected, but many proteins that form stable interactions with one or few other partners, including the propionyl-CoA carboxylase, mitochondrial calcium uniporter, calcineurin, and others. The results add to emerging evidence of independent regulation of protein levels from their cognate transcripts and suggest avenues to improve the interpretation of transcriptomics data.