Mapping determinants of protein specificity with high-throughput mutagenesis and probabilistic modeling

Abstract

Proteins evolve by iteratively varying their sequences to traverse fitness landscapes for different functions, driving the emergence of novelty at larger scales. Understanding the design principles of proteins requires a quantitative understanding of the map between a given protein sequence and an array of different fitness landscapes—for example, for different substrates or ligands. Scientists have studied this question using patterns found in natural sequences and by collecting their own maps of how variation in protein sequence affects different cellular functions. In this thesis, I will begin by broadly reviewing the problem of understanding and predicting protein sequence-function landscapes and then introduce the particular system most of interest to this thesis: membrane transporter proteins. Transporters’ sequences must encode functional rules for highly tuned specificities across chemical space, as well as for carefully energetically balanced transitions between structurally distinct conformations. First, in Chapter 2, I briefly discuss my work on the evolutionary history of one such family of transporters, the Natural resistance associated macrophage transporters (or Nramps). In Chapter 3, I then use high-throughput mutagenesis of a library guided by structure and evolution to systematically dissect the determinants of metal import and specificity in a model homolog of this family, DraNramp. I find that a key set of core residues in the first and second shells is essential to allow for import of the typically excluded substrate of Mg2+. A wide range of surrounding residues throughout the protein’s core additionally act as hotspots for modulating both epistatic interactions between mutations within one fitness landscape and specificity modulation between fitness landscapes. I then propose a theoretical model in which residues modulating the protein’s conformational equilibrium could underlie both effects. In Chapter 4, I build a new database of results from multiplexed functional assays of specificity, including my own, to ask if and how probabilistic models trained on natural sequence information can be used to guide our search for protein variants that alter substrate specificity. I find that many popular machine learning-based approaches systematically bias their predictions against variants that alter specificity by conditioning on local sequence context. To address this, I propose a simple weighted difference between models that can guide the sampling of sequence libraries to enrich for variants with altered specificity. Finally, in Chapter 5, I discuss work done collaboratively with Jacob Licht and Rachelle Gaudet in which we systematically analyzed the commonalities and differences between conformational transitions in the superfamily of membrane transporters containing DraNramp, identifying a common core “rocking” mechanism with additional protein-specific variations, which we categorize. In sum, I have done several projects analyzing protein specificity and evolution from several angles: experimental mutagenesis, machine learning, and structural analysis. These results demonstrate advances in our ability to predict and understand the determinants of protein specificity but additionally suggest that a properly predictive understanding of the design rules of protein specificity remains out of reach. In Chapter 6, I discuss prospects for this field.