Protein-protein interactions represent an essential aspect of most biological functions. A better understanding of how the sequence of two proteins determines quantitatively the affinity and specificity of these interactions will improve our ability to predict the functional impact of genetic variation as well as design new synthetic proteins of therapeutic or bioengineering interests. Recent advances in artificial intelligence and generative biology have opened the door to such developments. However, training deep learning models to predict affinity from sequence will require vast amounts of data that cover the wide and complex genetic landscape of protein-protein interaction affinity.

Our lab employs a deep mutational scanning approach, deepPCA, that we developed and optimized to quantitatively measure the affinity of millions of variants of protein-protein interactions. Current efforts aim at systematically mutagenizing human protein interaction domain families by quantitatively measure the effects of all possible single amino acid mutants on the whole interaction profile of each family member, generating datasets of millions of binding affinity measurements. Our original approach leverages the combinatorial nature of protein interaction domain families to sample a wide and diverse portion of their genetic landscape. Measuring the impact of mutations at this scale, across all members of the same family on the interaction against all other members, is required to detect the full spectrum of mutation impacts in different sequence and structural contexts. This wealth of data enables us to train thermodynamics-based generative models that capture the sequence-function relationship to predict interaction affinity (deltaG) from sequence alone. These models can then be used to generate new sequences with desired affinity and specificity.

Our research will shed a new light on the genetic architecture of protein function and assess its complexity, a fundamental question with important implications for the predictability of protein function. The experimental and analytical framework developed here will pave the way towards generalized models of protein-protein interaction affinity and contribute to the generative biology revolution.

Contact
Guillaume Diss