Evolution-based design of chorismate mutase enzymes

William P. Russ, Matteo Figliuzzi, Christian Stocker, Pierre Barrat-Charlaix, Michael Socolich, Peter Kast, Donald Hilvert, Rémi Monasson, Simona Cocco, Martin Weigt, Rama RanganathanbioRxiv 2020 · DOI

TL;DR

This paper presents a novel method for designing functional synthetic enzymes by learning evolutionary constraints directly from protein sequence data. By applying this "evolution-inspired" approach to chorismate mutase, they successfully engineered enzymes with natural-like catalytic activity and significant sequence diversity, opening doors for creating entirely new functional proteins.

What Problem Does This Paper Address?

Designing novel proteins with specific functions has long been a grand challenge in biology and engineering. Traditional approaches often rely on detailed atomic structures and physical models of molecular interactions. However, the explosion of protein sequence data and advancements in computational methods offer a new paradigm: learning the "rules" of protein function and folding directly from evolutionary history. The authors aim to demonstrate that sequence-based statistical models, derived from analyzing families of related proteins, are sufficient to specify functional proteins and can be used to generate novel, functional sequences. This addresses the need for a more general, data-driven approach to protein engineering that doesn't solely depend on structural information.

Methods

The core of their method is Direct Coupling Analysis (DCA), a technique originally developed to predict contacts between amino acids in protein structures. Here, DCA is adapted to infer a statistical model from a multiple sequence alignment (MSA) of a protein family. This model captures:

  • Intrinsic amino acid propensities (fields, $h_i(a)$): How likely each amino acid is to appear at a specific position, independent of other positions.
  • Pairwise interactions (couplings, $J_{ij}(a,b)$): How the presence of one amino acid at position $i$ influences the likelihood of another amino acid at position $j$.

These parameters define a probability distribution $P(a_1, \dots, a_L)$ for any given protein sequence. By sampling from this distribution, they generate synthetic sequences. To test the functional capabilities of these sequences, they employ a high-throughput in vivo complementation assay in E. coli. This assay leverages a chorismate mutase (CM)-deficient bacterial strain that requires functional CM for growth on minimal media lacking essential aromatic amino acids. By sequencing the bacterial population before and after selection, they can quantify the "relative enrichment" (r.e.) of each variant, which serves as a measure of its catalytic activity. They also introduce a computational "temperature" parameter ($T$) in their model to control the energy (and thus probability) of sampled sequences, allowing them to explore sequences with varying degrees of similarity to natural ones.

Datasets

The primary dataset used is a multiple sequence alignment (MSA) of 1,259 natural AroQ chorismate mutase (CM) protein domains. This MSA encompasses a broad diversity of bacterial, archaeal, fungal, and plant lineages. They also generated two types of synthetic libraries:

  1. Natural homolog library: Comprising nearly every natural CM homolog in the MSA.
  2. Synthetic variant library: Over 1,900 sequences designed by exploring various parameters of the bmDCA model.

The E. coli strain used for the complementation assay is a CM-deficient strain (KA12/pKIMP-UAUC).

Key Insights and Ideas

  • Evolutionary history encodes functional constraints: The study strongly suggests that analyzing the co-evolutionary patterns within a protein family's sequence alignment is sufficient to capture the essential rules governing its folding and function. This moves beyond simply looking at individual amino acid conservation.
  • DCA as a generative model: The bmDCA model, when used to sample sequences, acts as a "generative model." This means it can produce novel sequences that are statistically similar to natural ones and, importantly, can be functional.
  • "Temperature" as a design knob: The introduction of a computational "temperature" parameter ($T$) allows researchers to tune the generated sequences. Lower temperatures produce sequences closer to the natural distribution, while higher temperatures explore more divergent sequences, offering a way to balance functional similarity with novel sequence space exploration.
  • Sequence distance isn't everything: The paper demonstrates that enzyme function isn't solely determined by how far a sequence is from a natural counterpart or by the conservation of individual sites. The pattern of correlations between amino acids, as captured by the DCA couplings, is fundamentally important.
  • System-specific constraints can be learned: By annotating natural sequences with their functional performance in a specific E. coli assay, they show that additional constraints for system-specific function can be learned, suggesting an iterative design process where models are refined based on experimental feedback.

Key Results

  • The bmDCA model accurately recapitulates empirical first- and second-order statistics of natural sequences and even captures higher-order correlations not used in fitting.
  • Synthetic sequences generated by bmDCA at low "temperatures" (T ≤ 0.66) exhibit functional complementation in the E. coli assay, with some even exceeding the performance of natural sequences.
  • Out of 1,618 synthetic sequences tested, 481 (~30%) rescued growth, with sequence identities to the closest natural counterpart ranging from 42% to 92%.
  • Notably, 46 of these functional synthetic sequences had less than 65% identity to any protein in the MSA, representing at least 33 mutations away from the closest natural homolog.
  • A steep, non-linear relationship exists between bmDCA statistical energy and catalytic function: functional sequences are almost exclusively found below a specific energy threshold (EDCA < 50), which aligns with the energies of natural homologs.
  • By incorporating system-specific functional data, they could predict functional sequences with 83% accuracy.
  • Biochemical characterization of selected natural and designed CMs showed that the designed enzymes possess catalytic parameters closely recapitulating those of natural CMs, confirming they are bona fide synthetic orthologs.

Limitations and Future Work

  • Interpretation of weaker couplings: While direct structural contacts are identified by high-magnitude couplings ($J_{ij}$), the physical interpretation of weaker, non-contacting terms in $J_{ij}$ remains an area for further investigation. These weaker terms may describe collective amino acid evolution within the protein structure.
  • Refining the model: The authors suggest that further refinement of the topology and representation of sequence correlations is needed to fully understand the physics of protein structure and function.
  • Generalizability of the approach: While successful for chorismate mutase, the general applicability of this evolution-based approach to a wider range of protein families and functions needs to be explored.
  • Iterative design for specific phenotypes: The paper lays the groundwork for an iterative design strategy where the bmDCA model is updated with each round of selection to optimize for specific phenotypes and organismal environments. This suggests a future direction of using learned system-specific constraints to guide the design of proteins with tailored activities.
  • Exploring novel functions: The vast sequence space accessible through this method suggests potential for discovering proteins with altered or entirely novel chemical activities.

Summary updated

Abstract

The rational design of enzymes is an important goal for both fundamental and practical reasons. Here, we describe a design process in which we learn the constraints for specifying proteins purely from evolutionary sequence data, build libraries of synthetic genes, and test them for activity in vivo using a quantitative complementation assay. For chorismate mutase, a key enzyme in the biosynthesis of aromatic amino acids, we demonstrate the design of natural-like catalytic function with substantial sequence diversity. Further optimization focuses the generative model towards function in a specific genomic context. The data show that sequence-based statistical models suffice to specify proteins and provide access to an enormous space of synthetic functional sequences. This result provides a foundation for a general process for evolution-based design of artificial proteins. One-sentence summary An evolution-based, data-driven engineering process can build synthetic functional enzymes.
BibTeX
@misc{russ2020evolution,
  author = {William P. Russ and Matteo Figliuzzi and Christian Stocker and Pierre Barrat-Charlaix and Michael Socolich and Peter Kast and Donald Hilvert and Rémi Monasson and Simona Cocco and Martin Weigt and Rama Ranganathan},
  title = {{Evolution-based design of chorismate mutase enzymes}},
  year = {2020},
  doi = {10.1101/2020.04.01.020487},
  howpublished = {bioRxiv},
}
Figure