Mirages in continuous directed enzyme evolution: a cautionary case study with plantized bacterial THI4 enzymes

Abstract

Continuous directed evolution (CDE) improves the characteristics of a target enzyme by hypermutating the enzyme gene in vivo, coupling enzyme activity to growth of a microbial platform, and selecting for growth rate (Molina et al., 2022). Directed evolution can be interfaced with genome editing to expand the gene pool available for plant breeding; this powerful combination (DE–GE) has been neatly termed ‘a Green (r)Evolution’ (Gionfriddo et al., 2019). THI4 enzymes, which make the thiazole moiety of thiamin, are good testbed targets for plant CDE technology. Plant THI4s are energy-inefficient suicide enzymes that could potentially be replaced by efficient, non-suicide bacterial THI4s to increase biomass yield by as much as 4% (Joshi et al., 2021). However, bacterial THI4s are O₂-sensitive and otherwise ill-adapted to plants (Joshi et al., 2021). We therefore previously ran CDE campaigns in the yeast OrthoRep system to ‘plantize’ bacterial THI4s, that is, to improve function in an aerobic, plant-like milieu (Figure 1a) (García-García et al., 2022). Two notably successful campaigns were for the THI4 from Mucinivorans hirudinis (MhTHI4); these campaigns culminated when populations acquired single V124A or Y122C mutations that improved growth to near the wild-type rate (Van Gelder et al., 2023). Such culmination can be overcome by increasing the selection pressure (Molina et al., 2022).

Figure 1

Open in figure viewerPowerPoint

OrthoRep campaigns to plantize a bacterial THI4 and their outcomes. (a) The OrthoRep system. The target enzyme (MhTHI4 V124A), plus or minus a poly(A) tail, is encoded on the cytoplasmic p1 plasmid that also carries a LEU2 marker. p1 is hypermutated by a p1-specific, error-prone DNA polymerase (TP-DNAP1_611 or TP-DNAP1_633) encoded on a nuclear plasmid. The BY4741 platform strain carries a thi4Δ deletion to couple growth to the activity of the THI4 on p1. (b) The combinations of expression-reduction regimes with cold turkey (CT) or gradual (G) selection. (c) The non-synonymous (blue) and synonymous (green) mutations that had swept populations by the end of campaigns. When populations failed to grow early in campaigns, surviving populations were split into subpopulations (A, B, C) and propagated independently. The five mutant sequences whose non-synonymous mutations were tested are indicated in orange. (d) Growth of the five evolved populations from which these mutant sequences came compared to the lack of growth supported by these sequences when purged of synonymous mutations and cloned into fresh p1 and fresh cells (with a 72A tail and TP-DNAP1_611). The native MhTHI4, the parent V124A mutant and yeast THI4 (ScTHI4) were included as benchmarks as well as an empty vector (EV) control. Data are mean ± SE for the last three passages of the campaigns for evolved subpopulations and for 12 replicate cultures for the tests of non-synonymous mutations.

In the present work, we increased selection pressure on the MhTHI4 target by reducing its expression, thus reducing enzyme activity and cell growth rate and renewing the scope for improvement. In OrthoRep, expression can be reduced by shortening the genetically encoded poly(A) tail of the target's mRNA or reducing the copy number of the plasmid (p1) bearing the target gene (Ravikumar et al., 2018; Zhong et al., 2018) (Figure 1a). These manoeuvres led to improved, that is, faster-growing, populations harbouring MhTHI4s with new non-synonymous mutations, plus synonymous ones. Surprisingly, testing indicated that the synonymous mutations were probably largely or wholly responsible for the observed improvements in growth rate. As such ‘mirages’ seem highly likely to appear in other OrthoRep CDE projects, we document them here as a cautionary case study.

Our previous campaigns with MhTHI4 used a commercially recoded gene, of which the V124A mutant (Van Gelder et al., 2023) was the starting point for the present campaigns (Appendix S1). As before, the platform strain was BY4741 thi4Δ, which requires a thiazole precursor (HET) or thiamin for growth. The poly(A) tail was shortened from ~72A to 24A or 0A to reduce expression twofold or tenfold, respectively (Zhong et al., 2018). To reduce p1 copy number, the error-prone DNA polymerase was changed from TP-DNAP1_611 to TP-DNAP1_633, which also raises the mutation rate (García-García et al., 2022). Each expression-reduction regime was confirmed to work as expected (Figure S1) and then combined with ‘cold turkey’ selection (culture without added thiamin or HET) or ‘gradual’ selection (initial supplementation with limiting thiamin or HET, i.e. tapered to zero) (García-García et al., 2022). Subculturing was every 4–6 days. Nine independent populations of the V124A mutant were engineered for each expression-reduction regime and subjected to cold turkey or gradual selection, giving 54 initial subpopulations (Figure 1b). Subpopulations that survived without supplementation were split in three (denoted A, B, C) and propagated independently. Campaigns were ended when growth rate plateaued. Sequencing bulk DNA from subpopulations identified mutations of interest, that is, those that had fully replaced a wild-type base (‘swept’ the subpopulation).

The cold turkey strategy proved effective; after 33 passages, it yielded two subpopulations (V124A with no A tail and V124A with TP-DNAP_633) whose splits all reached OD₆₀₀ 1.5–5.0 by the end of each passage. All told, 11 non-synonymous mutations were found in the selected subpopulations along with nine synonymous mutations (and two 10B2 promoter mutations, most likely neutral because 10B2 is already optimized; Zhong et al., 2018) (Figure 1c). Notable non-synonymous mutations were a reversion of V124A and its displacement by Y122C; this switch implies that Y122C is functionally superior.

Five sequences with non-conservative non-synonymous mutations (Figure 1c) were advanced for further testing and resynthesized to purge synonymous mutations. The purged sequences were then cloned into fresh p1 plasmid and platform cells, and cell growth was tested in the conditions used to evolve the mutant sequences in which these unpurged sequences supported growth but the V124A mutant used as a starting point did not. Note that these culture conditions involve greater aeration (i.e. higher O₂ levels) than those used to select the V124A mutant (Van Gelder et al., 2023), and that this reduces complementing activity. Unlike their unpurged counterparts, the purged sequences did not support growth, whereas the yeast THI4 positive control did so, as expected (Figure 1d). The failure of the non-synonymous mutations alone to confer improved growth indicates that the accompanying synonymous mutations were also necessary—or even sufficient—for the observed growth phenotype. An alternative explanation based on acquired beneficial nuclear mutations is a priori unlikely (i) because the genomic mutation rate is ~100 000-fold lower than that of the OrthoRep target gene (Molina et al., 2022) and (ii) because we have seen no previous cases of this, let alone simultaneously in separate populations (García-García et al., 2022; Van Gelder et al., 2023).

That non-synonymous mutations can improve the performance of a commercially codon-optimized bacterial enzyme in yeast is not surprising (Lanza et al., 2014). The codon optimization algorithm strongly favoured abundant yeast codons, which is not always the most effective scheme (Lanza et al., 2014), and indeed the synonymous mutations obtained all led to less-abundant codons, for example, Gly GGT→GGC, Asp GAT→GAC. What is surprising is that the effects of synonymous mutations completely dominated the campaigns and that no new non-synonymous mutations with substantial benefits were recovered. The non-synonymous mutations were thus presumably near-neutral passenger mutations that were already present in genes in which beneficial synonymous mutations arose. Consistent with this possibility, the non-synonymous mutations F14L and I198T (Figure 1c) were previously classed as neutral or mildly deleterious (Van Gelder et al., 2023). Also, all but one of the 11 non-synonymous mutations were at non-conserved or weakly conserved positions, and six of them occur naturally (Table S1). Note that OrthoRep's extremely high mutation rate is designed to make many mutations in the target gene—this is a feature, not a bug—(Molina et al., 2022), so that patterns like those in Figure 1c are sometimes to be expected.

A tactical conclusion from the failure to obtain mutant MhTHI4s substantially better than V124A or Y122C is that the mutational space accessible by TP-DNAP1_611 and TP-DNAP1_633, which make transition mutations but few transversions (García-García et al., 2022), has now been largely explored. Next-generation error-prone OrthoRep DNA polymerases that make many more transversions—and hence a far wider range of amino acid changes (Rix et al., 2024)—are likely to allow further progress.

A strategic conclusion for OrthoRep as an aerobic, eukaryotic platform to plantize non-plant enzymes is that users' default assumption should be that synonymous mutations confer growth improvement as effectively as non-synonymous ones—although just by increasing enzyme's expression level instead of changing its properties, that is, by subverting the aim of the CDE campaign and creating an improvement mirage. To maximize efficiency, promising non-synonymous mutations should therefore be tested without any accompanying synonymous mutations at an early stage, as in Figure 1d. This strategic conclusion applies equally to using OrthoRep as a platform to improve plant enzymes because, whether the target enzyme gene is a native plant DNA sequence or a plant gene recoded for yeast expression, codon use could well be suboptimal, that is, improvable by ‘yeastizing’ mutations.

To summarize: the codon bias issue illustrated here, like yeast's preference for distinct amino acids at particular positions in a given protein (Van Gelder et al., 2024), must be monitored vigilantly when using OrthoRep as a platform to improve enzymes for use in plants. Codon and amino acid bias do not, however, compromise OrthoRep's evolutionary power because yeastized features of the DNA or protein sequences they lead to can be swiftly recognized and rejected. Forewarned is forearmed.