Data‐driven Protein Engineering

Introduction A protein’s sequence of amino acids encodes its function. This “function” could refer to a protein’s natural biological function, or it could also be any other property including binding affinity toward a particular ligand, thermodynamic stability, or catalytic activity. A detailed understanding of how these functions are encoded would allow us to more accurately reconstruct the tree of life and possibly predict future evolutionary events, diagnose genetic diseases before they manifest symptoms, and design new proteins with useful properties. We know that a protein sequence folds into a three-dimensional structure, and this structure positions specific chemical groups to perform a function; however, we’re missing the quantitative details of this sequence-structure-function mapping. This mapping is extraordinarily complex because it involves thousands of molecular interactions that are dynamically coupled across multiple length and time scales. Computational methods can be used to model the mapping from sequence to structure to function. Tools such as molecular dynamics simulations or Rosetta use atomic representations of protein structures and physics-based energy functions to model structures and functions (1–3). While these models are based on well-founded physical principles, they often fail to capture a protein’s overall global behavior and properties. There are numerous challenges associated with physics-based models including consideration of conformational dynamics, the requirement to make energy function approximations for the sake of computational efficiency, and the fact that, for many complex properties such as enzyme catalysis, the molecular basis is simply unknown (4). In systems composed of thousands of atoms, the propagation of small errors quickly overwhelms any predictive accuracy. Despite tremendous breakthroughs and research progress over the last century, we still lack the key details to reliably predict, simulate, and design protein function. In this chapter, we present the emerging field of data-driven protein engineering. Instead of physically modeling the relationships between protein sequence, structure, and function, data-driven methods use ideas from statistics and machine learning to infer these complex relationships from data. This top-down modeling approach implicitly captures the numerous and possibly unknown factors that shape the mapping from sequence to function. Statistical models have been used to understand the molecular basis of protein function and provide exceptional predictive accuracy for protein design.