Protein interaction networks represent the backbone of the cell’s functional organization. Their architecture relies on the specificity of these interactions. Mutations that alter specificity may rewire the network and the way it coordinates biological processes, and lead to diseases or functional innovations. Understanding how the sequence of a protein determines the specificity of its interaction profile thus represents an important step towards a model of the genotype-phenotype map.
To address the sequence-function relationship in protein-protein interactions, we conducted a deep mutational scanning of the entire human basic leucine zipper (bZIP) interaction network. Using our established double-deepPCA system, we measured the effect of all single amino acid substitutions in all 54 members, on the affinity for all 54 wild-type members. This dataset of ~2M quantitative protein-protein interaction measurements enabled training a deep neural network that captures the sequence-function relationship in order to quantitatively predict the free energy of binding between two bZIPs from their sequence alone.
I will present the dataset and the deep learning model, as well as preliminary data on two other families, basic helix-loop-helix and SH3 domains, for which we plan to collect 6 and 20M affinity measurements respectively. We foresee that applying this strategy on a variety of domain families of different architecture will enable the development of general models that can predict the affinity of an interaction from the sequence of the two partners.