A Machine Learning Approach to Heterologous Membrane Protein Overexpression

Abstract

Membrane protein production is difficult; their biogenesis does not stop with translation but also requires translocation and integration into a lipid bilayer. These additional steps hamper their heterologous expression which significantly impedes biophysical and structural studies. Detailed and anecdotal evidence in literature suggests that a variety nucleotide and amino-acid sequence level determinants may potentially support or hinder their biogenesis, e.g. mRNA pausing elements, codon adaptation, transmembrane segment hydrophobicity, “positive inside rule.” By training a preference-ranking Support Vector Machine, we have developed a statistical model that predicts a relative likelihood of a membrane protein's successful expression using quantitative experimental data of overexpression. This model is rigorously validated against expression outcomes from small-scale laboratory experiments (e.g. expression tests that routinely precede structural studies) published in the literature as well as large-scale expression trials from a Protein Structure Initiative consortium facility. We show remarkable agreement between the predicted and experimental expression outcomes and propose our model, trained and cross-validated on the entire corpus of data, as a tool for the membrane protein biophysics community to streamline the process of overexpressing a target for study. Given the framework of our model, it can be trivially re-trained as additional experimental outcomes are gathered from past work or created from experiments. Furthermore, the relative weights gathered from parameters of the statistical model may help further characterize translocation mechanisms and suggest intriguing areas for further biophysical and computational experiments.