DIRECTED EVOLUTION REVEALS THE FUNCTIONAL SEQUENCE SPACE OF AN ADENYLATION DOMAIN SPECIFICITY CODE

Abstract

Nonribosomal peptides are important natural products biosynthesized by nonribosomal peptide synthetases (NRPSs). Adenylation (A) domains of NRPSs are highly specific for the substrate they recognize. This recognition is determined by ten residues in the substrate-binding pocket, referred to as the specificity code. This finding led to the proposal that nonribosomal peptides could be altered by specificity code swapping. Unfortunately, this approach has proven, with few exceptions, to be unproductive; changing the specificity code typically results in broadened specificity or poor function. To enhance our understanding of A domain substrate selectivity, we carried out a detailed analysis of the specificity code from the A domain of EntF, an NRPS involved in enterobactin biosynthesis in Escherichia coli. Using directed evolution and a genetic selection, we determined which sites in the code have strict residue requirements and which are tolerant of variation. We showed that the EntF A domain, and other L-Ser-specific A domains, have a functional sequence space for L-Ser recognition, rather than a single code. This functional space is more expansive than the aggregate of all characterized L-Ser-specific A domains: we identified 152 new L-Ser specificity codes. Together, our data provide essential insights into how to overcome the barriers that prevent rational changes to A domain specificity.

INTRODUCTION

Nonribosomal peptides (NRPs) are common natural products of tremendous medical and agricultural importance. Assembly of these natural products by nonribosomal peptide synthetases (NRPSs) involves enzymology that functions analogously to an assembly line. Each amino acid is recognized and incorporated into the NRP by a set of enzymatic domains. Nature has derived the diversity of biological activities of NRPs by changing the number, order, and amino acid specificities of these domains.

The adenylation (A) domain of each module controls amino acid specificity, although the determinants of this specificity are not yet fully clear (1). The crystal structure of an L-Phe-bound A domain revealed 10 residues that interact with the substrate in the binding pocket (2,3). Analysis of the primary sequence of other A domains revealed a ‘specificity code’ which is highly conserved between A domains that recognize the same substrate (3,4). Later, it was shown that the specificity code is located within a subregion of the A domain, referred to hereafter as the recognition subdomain (RS), transfer of which may be responsible for differences in specificity between otherwise close homologs (5, 6). Through the identification of the RS and specificity code, substrate specificities can be inferred from the primary sequence of an A domain (3,7).

The discovery of the specificity code suggested that exchanging the code of one A domain with another would alter substrate selection allowing the production of new NRPs. While this approach seemed promising initially, early successes were likely enabled by the similarities of the A domains, specificity codes, and substrates that were exchanged (3,8–10). Attempts at less conservative changes resulted in variants with substantially reduced catalytic efficiency (11,12). To date, structural biology and bioinformatics have failed to identify a clear path for reliably switching specificity.

Directed evolution has shown promise in overcoming limitations of rational specificity code design. Successive single-residue randomization and an in vitro enzymatic screen were used to identify three single-residue substitutions in the specificity code that result in the activation of a non-native substrate (12). Simultaneous randomization of three specificity code sites was used to alter the structure of the NRPS-derived antibiotic andrimid (13). Directed evolution has also been used with yeast cell surface display to switch substrate recognition (14,15). Additionally, others have used computational methods to guide changes to the substrate-binding pocket to switch amino acid specificity (11,16).

In this study, we used directed evolution of EntF, an NRPS involved in enterobactin (ENT) siderophore biosynthesis in Escherichia coli (Figure 1), to gain a deeper understanding of how the A domain specificity code confers substrate selectivity (17–20). Because siderophores are essential for Fe³⁺ acquisition in iron-limited conditions, this model system is ideal for performing selections or screens to process large DNA libraries. ENT has been used previously in directed evolution screens to identify functional chimeric A domains as well as to understand protein-protein interactions (21–24). Using a genetic selection and saturation mutagenesis to randomize the EntF specificity code, we discovered that there is an expansive functional sequence space, rather than a small number of discrete codes, for recognition of L-Ser. Additionally, we established that this sequence space is likely shared by other L-Ser-specific A domains. We characterized 157 unique EntF variants to assess the tolerance for residue variability at each site in the specificity code and determined that only a few specificity code residues are required for substrate recognition. Our data are consistent with the conclusion that A domains have a strong bias for their native substrate that may complicate targeted specificity code swaps and require directed evolution approaches.

Figure 1. Enterobactin (ENT) formation and the structure, specificity code, and function of the A domain of EntF.

A) Diagram of the ENT biosynthetic pathway. EntE tethers 2,3-dihydroxybenzoic acid (DHB) to EntB. The condensation (C) domain of EntF condenses DHB and L-Ser, previously activated and bound to the thiolation (T) domain of EntF. After one turnover, the DHB-L-Ser monomer is stored on the thioesterase (TE) domain. After three iterations of this process, the final product is cyclized and released. B) The two half-reactions of the EntF A domain. In the first, the A domain activates L-Ser as Seryl-AMP (1) and transfers the seryl group to the T domain (2) in the second. C) Ribbon representation of the EntF (PDB5JA1) A domain with the recognition subdomain, in red, and specificity code residues, highlighted in colors matching D and E. D) Binding pocket residues that form the specificity code and the reaction intermediate-mimic inhibitor, seryl-AVS (25), used for crystallization. E) WebLogo (26) of specificity code sites 1–10 for 82 characterized L-Ser codes (27).

RESULTS AND DISCUSSION

Identification of novel L-Ser specificity codes.

In Nature, the specificity code for an amino acid may fall into several distinct groups, e.g., L-Leu has 4 different code groups, (3,27,28). However, whether there is sequence flexibility, i.e. tolerance for variation, in the specificity code of a single A domain is not known. This knowledge gap has contributed to the lack of successful specificity code changes in efforts to reprogram NRPS enzymology. To address A domain specificity code flexibility, we targeted residues two through nine of the 10-site EntF specificity code for site-saturation mutagenesis (Figure 1D; Figure S1A). Asp649 and Lys952, in EntF notation, and hereafter referred to in specificity code notation as sites 1 and 10, were not randomized because they interact with the amino and carboxyl groups of the amino acid substrates and are highly conserved. For the remaining eight sites, we designed libraries 1 (L1) and 2 (L2) to randomize sites 2, 3, 4, 6, and 7 and sites 3, 4, 6, 7, and 9, respectively (Figure S1A). Based on sequence alignments of 82 characterized L-Ser-specific codes (Figure 1E; Table S1), site 5 appeared to be the most variable while site 8 is in almost always Val or Ile. Both sites are considered “wobble” sites for L-Ser-specific codes and are the only two sites without a ~90% predominant residue (27). Thus, to maximize our chances of identifying sequence variation, sites 5 and 8 (dark green in Figure 1C, 1D, and 1E) were excluded from the initial libraries. We did not target residues outside the original ten-residue code (3) since these specificity code residues are the most likely to have direct interactions with the substrate and intermediates during catalysis, therefore having a higher probability of impacting substrate recognition. Additionally, L-Ser is the only amino acid competent to form the trilactone core of ENT thereby restriction our analysis to specificity codes that are competent for recognizing this amino acid at some level.

The entF specificity code mutant libraries were electroporated into an E. coli ΔentF strain and clones enabling ENT production were selected on iron-limiting media containing the iron chelator ethylenediamine-di(o-hydroxyphenylacetic acid), EDDHA. As little as 1 μM EDDHA inhibited growth of a ΔentF strain, thus, to capture EntF specificity code variants, including those significantly reduced in function, we used 1 μM EDDHA in all selections. After confirmation that the growth phenotype was conferred by the plasmid, the RS-encoding region of each ENT producer strain was sequenced (Table S2). The results from L1 and L2 suggested that His4, Ser6, and Asp9 (yellow, purple, and blue, respectively, in Figure 1C, 1D, and 1E) are nearly invariant, with one Ser4 exception (Figure 2).

Figure 2. Residue usage by code site in the EntF variants and characterized L-Ser-specific A domains.

Pie charts describing, by site in the specificity code, the amino acid usage across all libraries. The wild-type EntF residues are shown in the top row followed by the 82 characterized L-Ser-specific A domain residues, and, subsequently, each of the four libraries. Circles with diagonal lines designate non-mutagenized sites. The color of each sector corresponds to the key at the bottom; both are organized alphabetically, clockwise and left to right. For each library, n designates the number of isolated strains containing DNA-unique entF mutants, provided they had no non-code residue substitutions. All Libraries n= 200, 216, 50, 168, 50, 216, 168, and 17 for sites 2–9, respectively.

To explore the possibility that a change at site 5 or site 8 is required for variation at site 4, 6, or 9, we designed library 3 to randomize sites 4, 5, 6, 8, and 9 (Figure S1A). Selections from L3 yielded only a single EntF variant, EntF 3-58. EntF 3-58 has His4, Ser6, and Asp9, reinforcing the requirement for these residues. We discovered that site 8 also has limited potential for variability; thus, this library contains only one highly variable site, dramatically reducing the proportion of viable EntF variants. In contrast, library 4 was designed to exclude sites 4, 6, and 9 and focus on the remaining, more flexible sites 2, 3, 5, 7, and 8. Consequently, L4 yielded the largest number of functional EntF variants. Considering all libraries, we identified 225 DNA-unique functional clones corresponding to 157 unique EntF-specificity code variants that included 26 residues not previously observed at their particular sites when compared to characterized L-Ser-specific A domains (27, Table S1).

Our dataset had a site-residue consensus for His4-Ser6-Asp9, with one Ser4 exception. Likewise, in Nature, among all characterized A domains, His4-Ser6-Asp9 is characteristic of L-Ser specificity; approximately 97% of A domains that have His4-Ser6-Asp9 are L-Ser specific. However, 11 other three-site-residue combinations, e.g. Val2-Phe5-Ser6, are just as, or more, characteristic of L-Ser specificity among A domains in Nature (27). Yet, for EntF, we experimentally found that only His4-Ser6-Asp9 are strictly required. The importance of His, Ser, and Asp is supported by a recent structure of EntF, crystallized with the catalytic-intermediate mimic, serine adenosine vinylsulfonamide (Ser-AVS), showing these residues in close proximity to the seryl moiety of the substrate (29). At site 8 we also observed residue usage similarities between the EntF variants and the characterized serine-specific codes: both contain Val and Ile in a roughly 55:45 proportion (Figure 2).

The residue usage between the EntF variants and the characterized L-Ser-specific A domains is less similar at sites 5 and 7 (dark green and light green, respectively, in Figure 1C, 1D, and 1E). In both datasets, site 5 is tolerant of a wide range of residues, yet more so in the EntF variants which have 15 allowed residues compared to seven for the codes found in Nature. Likewise, site 7 is more diverse in the EntF variants with 12 observed residues for the EntF variants compared to five in the natural codes (Figure 2). The tolerance for amino acid variability at these two sites is consistent with the observation that both are distal to the substrate, particularly site 5 (Figure 1D). However, based only on the characterized codes, the tolerance for diversity at site 7 would appear approximately as strict as for sites 2, 3, 4, 6, and 9; thus, the variability observed among the EntF variants at site 7 was unexpected.

At sites 2 and 3 (light green in Figure 1C, 1D, and 1E), the EntF variants deviate significantly from the characterized A domains. In Nature, Val2 and Trp3 predominate, with diversity similar to sites 4, 6, 7, and 9. However, in the EntF variants, site 3 is highly variable with 12 allowed residues, and site 2 has an approximately even distribution of Val- and Pro-containing variants. The latter is highly unexpected because Pro is observed only once (at site 3) in the codes of naturally occurring L-Ser-specific A domains and is the second least-used code residue among all characterized A domains, after Arg (27). Furthermore, in the EntF structure, site 2 is proximal to the substrate; thus, tolerance of Pro2 was surprising.

Characterization of in vivo function and substrate specificity of EntF variants.

To discriminate between the isolated ENT-producing strains in vivo, we screened each for growth in M9 liquid minimal media with EDDHA concentrations ranging from 50 μM to 850 μM. The maximum tolerated concentration of EDDHA for each strain is shown in Table S2 and is referred to as an MIC_EDDHA. Variants with more code differences from wild-type typically resulted in a lower MIC_EDDHA. However, this was not always the case as many strains had a low MIC_EDDHA despite only one or two residue differences, and several strains had a high MIC_EDDHA despite three to four differences. Thus, we found a broad range of in vivo ENT production, approximated by MIC_EDDHA, among the isolated strains (Table S2).

To test whether the observed differences in MIC_EDDHA were due to broadened A domain specificity, we analyzed a subset of EntF variants spanning all libraries and all MIC_EDDHA tiers using the ATP/PP_i exchange assay (Table S3). A broadening of specificity would reduce ENT production due to the activation or aminoacylation of other amino acids not competent for ENT formation. Even the least active EntF variants were specific for L-Ser (Figure S2), though this may be due to our minimum threshold of growth at 50 μM EDDHA for detailed characterization. It is possible that variants initially isolated on iron-limited media with only 1μM EDDHA, but then failed to grow at 50 μM EDDHA, had broader substrate specificity. Future characterization of these variants may provide insights into how to overcome the inherent L-Ser specificity of the EntF A domain. Thus, the differences in MIC_EDDHA observed for the characterized ENT-producer strains is not due to broadened A domain specificity for non-serine substrates. Additionally, the levels of the co-purified MbtH-like (MLP) protein, YbdZ, important for in vivo ENT production and in vitro A domain activity (20), were very similar among the purified EntF variants, determined by immunoblotting done as previously reported (30). Therefore, the amino acid substitutions did not disrupt the EntF/YbdZ interactions that influence A domain function (20,30,31).

Next, we examined whether the EntF variants were impacted kinetically for L-Ser activation. ATP/PP_i exchange assays were used to determine the apparent K_m for L-Ser-binding and apparent V_max of a subset of variants (Table S3, Figure S3). With two exceptions, the variants ranged from 2% to 21% of the catalytic efficiency (V_max/K_m) of wild-type EntF with predominantly a K_m effect. Overall, we found no correlation between either apparent K_m or apparent V_max and in vivo ENT production as measured by MIC_EDDHA. Two variants, 4-136 and 3-58, had higher catalytic efficiency than wild-type EntF. Interestingly, a strain expressing 3-58 only grew to 600 μM EDDHA, while one carrying 4-136 grew at the highest tested EDDHA concentration, as did the wild-type. The difference in MIC_EDDHA between strains carrying 3-58 and 4-136, as well as the lack of correlation between kinetic parameters and MIC_EDDHA, suggest that some other aspect of ENT biosynthesis beyond L-Ser recognition is impacted by the specificity code substitutions.

Molecular modeling of substrate binding.

One such aspect is the binding of the Ser-AMP intermediate that is formed prior to transfer to the 4’-phosphopantetheinyl group on the thiolation domain (Figure 1B). To address this question, we used the recently solved crystal structure of EntF with a non-reactive Ser-AMP-phosphopantetheinyl intermediate mimic, Seryl-AVS (29), to model binding of Ser-AMP and L-Ser. Modeling of L-Ser in the binding pocket of the in vitro-characterized EntF variants revealed a correlation between the K_m and in silico CHARMM-based interaction energy scores, referred to as IPRO energy scores, for L-Ser, which account for non-covalent forces of interaction (van der Waals, electrostatics, and solvation). Furthermore, the K_m values correlated more strongly with the IPRO energy scores for L-Ser than for Ser-AMP, the catalytic intermediate (Figure 3A). This was expected since the K_m measures the binding of L-Ser, not Ser-AMP. However, the MIC_EDDHA correlated much more strongly with the IPRO energy scores for Ser-AMP than for L-Ser (Figure 3B). This correlation was also observed when considering all the variants (Figure 3C). These data suggest that EntF-Ser-AMP binding is important for the function of EntF and that specificity code substitutions may impact this binding. For example, variant 4-19 has the highest IPRO energy score of all EntF variants and a high associated MIC_EDDHA of 600 μM, despite the third lowest V_max/K_m among the in vitro-characterized enzymes. Similarly, the relatively low Ser-AMP IPRO energy score of 3-58 could be the reason why its associated MIC_EDDHA is lower than both the wild-type and 4-136, despite similar kinetic parameters (Table S3).

Figure 3. in silico IPRO energy scores of EntF variants for L-Ser correlate with K_m, binding for Ser-AMP correlates with MIC_EDDHA.

Correlation between the IPRO energy score with L-Ser or Ser-AMP and A) K_m and B) the associated MIC_EDDHA of in vitro-characterized EntF variants (Table S3). A more negative value for IPRO energy score indicates stronger binding. C) Correlation between IPRO energy score with Ser-AMP by each EntF variant (blue) and the MIC_EDDHA of the corresponding entF mutant strain (green). IPRO energy scores and MIC_EDDHA are normalized to wild-type.

Using the modeled EntF-Ser-AMP-bound complexes, we observed that most of the high-functioning variants (by MIC_EDDHA) have strong electrostatic interactions with the Ser-AMP intermediate. On the other hand, the low-functioning variants either exhibit weaker interactions or lack similar interactions with Ser-AMP altogether. Additionally, the low-functioning variants have more intra-enzyme electrostatic and hydrophobic interactions, typically distal to Ser-AMP (Figure 4A and 4B).

Figure 4. Substrate-binding pocket interactions differ between high- and low-functioning variants.

Diagrams of the substrate-binding pockets of (A) WT EntF, EntF variant 4-71, and (B) EntF variant 106 highlighting differences between their energy-minimized structures, based on PDB 5JA1, in the electrostatic interactions of the code residues and Ser-AMP (dashed line; side view – top row) and the intra-enzyme interactions (green line; bottom view – bottom row). Each diagram also contains specificity code residues 2–9 (light pink with blue font), the Ser-AMP intermediate (yellow), and residues differing from WT (black font).

Characterization of residue usage in variant specificity codes.

To test whether the diversity of codes was influenced by bias in library construction, we analyzed ~100 unselected clones from each library and determined that significant nucleotide usage biases occurred with a trend toward overrepresentation of Cs in sites 2-7 and Gs in sites 8 and 9. This is reflected in a significant skew in the predicted amino acid distributions for all sites except 7 and 9 (Table S4) with, at most, two-fold up or down changes relative to NNK proportions. These biases did not prevent underrepresented residues from emerging from the selection, with some of the most frequently observed residues at sites 3, 5, and 7, being underrepresented.

Using a chi-squared test, we compared the amino acid input frequencies in the libraries (based on the determined nucleotide usage bias) to the outputs of the selection to determine whether selection occurred at each site and to determine whether the relative proportions of the observed residues matched the input. This comparison showed that selective pressure was exerted on each site, including the highly variable sites 3 and 5 (Figure 5A) and that in all cases the observed residue proportions deviate from expectation, indicating favor or disfavor by the selection (Figure 5B). We saw enrichment of Val2, Trp3, Val5, Leu7, and Ile8, which, except for Val5 and Ile8, are the wild-type residues. WT residues Ser5, His7, Phe7, and Val8, however, were all unenriched.

Figure 5. Preferential residue usage is observed at all sites.

A) Comparison of amino acid residue proportions at sites 3 and 5 between the input and selection output. Each sector color corresponds to an amino acid residue according to the key at the bottom and sectors are sorted, clockwise from the top, by contribution to the chi-squared statistic. B) Comparison of the residue proportions at each mutagenized site in the L4 dataset to the input. The color of each sector corresponds to the key at bottom, both of which are organized alphabetically, clockwise and left to right, respectively.

Potential explanations for enrichment include effects on in vivo production of ENT (MIC_EDDHA) or co-variation between sites. Using ANOVA, we examined whether certain residues were associated with higher or lower MIC_EDDHA on average. Across all libraries, only one site-residue combination, Trp3, had an average associated MIC_EDDHA significantly different than any other. It is likely that most codes function better with a bulky hydrophobic residue at the bottom of the substrate-binding pocket (Figure 4B). Notably, variants with Pro, among the rarest residues in all specificity codes (27), at site 2, were not outperformed by variants with Val2, the WT residue.

We analyzed all pairwise combinations of sites for deviation from an even distribution of the residues observed at one site amongst those observed at another, using a chi-squared test. This analysis revealed a skewed distribution, suggesting covariation, between sites 2 and 3, 2 and 5, and 3 and 8. (Table S5). Covariance between sites 2 and 3 and sites 3 and 8 can be rationalized due to their proximity in the binding pocket, however, covariance between sites 2 and 5 is more surprising (Figure 1D). We observed no correlation between covariance and MIC_EDDHA or residue enrichment. For example, both Val2 and Trp3 are enriched, however, the Val2/Trp3 pair is underrepresented in the covariance. Patterns of residue frequency may simply reflect the extent to which different residues allow for possibilities at other sites, i.e., how many functional specificity code “solutions” exist given a particular set of residues at other sites.

Variant EntF codes function in non-EntF protein contexts.

To determine whether the specificity codes we identified occur in other A domains, we searched 146,187 sequence-unique A domains from GenBank using SANDPUMA (27). 11,026 were predicted by multiple methods to be specific for L-Ser, providing a total of 23 unique L-Ser specificity codes. Among these, five different specificity codes match those found in the EntF variants. Two of these five codes, DVWHLSLIDK (3-58) and DVWHLSLVDK (4-213), are found in A domains that have been characterized and confirmed to be L-Ser specific (32–34). Three of these five codes, matching 4-136, K16A, and 4-54, had not been previously characterized in any A domain. The remaining 152 EntF variant codes, despite being biologically functional, do not match the specificity code of any A domain sequence in GenBank. To investigate whether this large number of unobserved codes is only relevant to EntF, we phylogenetically compared the A domains with codes matching the EntF variants to those of characterized L-Ser-specific A domains. We found that A domains with codes matching the EntF variants are not confined to the clades most closely related to EntF (Figure S4), suggesting that we were able to identify diverse L-Ser specificity codes.

We were interested in determining whether any uncharacterized A domains that have a specificity code that matches one of our EntF variant codes is specific for L-Ser in their native context. The EntF specificity code variant DVWHYSLVDK (4-136) is found in the A domain of DltA from Paenibacillus donghaensis. The A domain of DltA is 50.3% identical to EntF across the RS region and 42.4% identical overall. We overproduced and purified the A-PCP from DltA in E. coli, assayed it by ATP/PP_i exchange, and determined that it activated only L-Ser (Figure 6). Thus, the EntF variants greatly expand the number of characterized unique L-Ser specificity codes, adding up to 155 to the previously known 23 (27), and can be used to confirm in silico A domain specificity predictions.

Figure 6. A variant EntF specificity code, differing at all five mutagenized sites, functions in a non-EntF context.

Comparison of the in vitro activity of purified proteins EntF WT, EntF variant 4-16, DltA WT, and DltA variant 4-16 with pooled amino acid substrates (A) and the individual amino acids contained in pool 1 (B) as determined by ATP-PP_i exchange assay. The color key in panel A applies to both panels. Pool 1: SATPG, 2: VILMC, 3: NDEQF, 4: YWHKR.

The five naturally-occurring codes that match the EntF variants differ from wild-type EntF at one or two sites. To test whether a less similar code could function in a non-EntF context, we changed the code of the DltA A domain to match the EntF variant 4-16, which differs from both EntF and DltA at five sites. We assayed the DltA 4-16 variant by ATP/PP_i exchange and found that it is specific for L-Ser (Figure 6). The activity of DltA is 27.2% that of EntF 4-136, while the activity of DltA 4-16 is 19.8% of EntF 4-16. This similarity suggests that the reduction in activity is primarily due to the differences between the two proteins rather than the 4-16 code in DltA. Thus, all of the EntF variant codes have the potential to activate L-Ser outside of an EntF context and the functional sequence space of EntF may also extend to other L-Ser-specific A domains.

In conclusion, our findings show that the EntF specificity code, and possibly any other, has the potential for variation greatly exceeding that which occurs in Nature. Despite the presumably relaxed selective pressure in a laboratory setting, the identification of a sequence space was surprising since all EntF proteins found to date in bacteria have the same specificity code. Even a recently identified group of EntF homologs from yeast, diverged over 60 million years, with just 56% identity with E. coli EntF (35), differ by only one code residue.

The variable tolerance for residue diversity among the specificity code sites along with the broad functional sequence space for L-Ser present several interesting possibilities for directed code swaps. First, our data suggest that specificity changes could be accomplished with minimal perturbations to the code by targeting key residues that confer specificity, e.g. His4, Ser6, and Asp9, in the case of EntF. Other residues, at more variable sites, could be left unchanged to preserve intra-protein interactions or be adjusted to best fit the target substrate. On the other hand, the sequence space of the EntF specificity code is consistent with the conclusion that some level of L-Ser activation, rather than of the desired substrate, would persist despite rational code swaps. Instead, to recognize a non-native substrate, a selection or screen would be necessary to overcome the improbability of choosing a code that is, first, outside of the sequence space for the native substrate, and, second, inside the sequence space for the non-native substrate. In summary, our data provide essential insights and suggest strategies that can be leveraged to overcome the barriers preventing rational changes to A domain substrate specificity.

METHODS

Plasmids and bacterial strains.

Bacterial strains, plasmids, and primers are listed in Table S6. BW27749 ΔentF E. coli strain was constructed using pMAK705-entF as previously described (36). Plasmid pACYC184entF-ES was constructed with entF from E. coli MG1655 and EagI and SacI restriction sites flanking the RS-encoding region (Figure S1). Plasmids pCR-BluntII-TOPOentF-RS-F1–5 were constructed with the RS-encoding gene fragments, each cloned using the TOPO method (Invitrogen). Plasmid pACYC184entF-ES-RS_sub was constructed by replacing the RS-encoding region of pACYC184entF-ES with placeholder E. coli DNA. Plasmids pACYC184entF-ES-L1–4 were constructed by amplification of the RS-encoding fragments in pCR-BluntII-TOPOentF-RS-F1–5 using NNK mutagenic primers, overlap-extension PCR, and ligation into pACYC184entF-ES-RS_sub (Figure S1). Selection-isolated plasmids were designated as “pACYC184entF-ES-variant#” corresponding to the strain number (Table S2). Plasmids pET28bentF-ES-variant#, pET28bentF-ES-wildtype, pET28bdltA-wildtype, and pET28bdltA-4-16 were constructed using polymerase incomplete primer extension, PIPE (37). Plasmid pACYC-duet-1 containing the E. coli MLP-encoding gene, ybdZ, was previously constructed (20).

EntF library creation.

The RS-encoding fragments in pCR-BluntII-TOPOentF-RS-F1–5 were amplified using primers containing NNK codons corresponding to the residues targeted for mutagenesis. The mutagenized RS-encoding fragments were combined by overlap extension PCR and ligated into pACYC184entF-ES-RS_sub (Figure S1C). Ligations were electroporated into NEB 10β cells. Transformants were pooled and pACYC184entF-ES-L1–4 plasmids (Figure S1C), consisting of 2.5, 3.5, 2.5, and 7.5 million transformants, respectively, were recovered.

Selection and isolation of ENT producers.

Plasmids pACYC184entF-ES-L1–4 were electroporated into BW27749 ΔentF cells and incubated on M9 minimal media noble agar plates with 0.4% (v/v) glycerol, 1 μM EDDHA (Complete Green Company), and antibiotic (chloramphenicol [34 μg mL⁻¹] or streptomycin [100 μg mL⁻¹]) (Figure S1D). Chloramphenicol was used for pACYC184entF-ES-L1–2, pACYC184entF-ES-L3–4 were switched to streptomycin to eliminate a chloramphenicol-resistant pACYC184entF-wildtype contaminant. For technical reasons, libraries were not saturated. Most notably, the frequency of observation of wildtype contamination was much higher than that of successful transformants for library 3, which had the most restrictive combination of sites targeted for mutagenesis. This wildtype contamination was traced to the reversion of the recA1 allele in commercially purchased competent cells, followed by recombination with the wildtype entF in the chromosome of these cells during library construction. ENT producer colonies were streaked for isolation on M9 plates with 0.4% (v/v) glycerol, 1 μM EDDHA, and antibiotic. From each, four colonies were incubated in a 96-well plate containing M9 with 0.4% glycerol (v/v), 50 μM EDDHA, and antibiotic. Plasmid preps were done from colonies that grew. Plasmids were screened by digest to eliminate pACYC184entF-wildtype (non-EagI or SacI). Each correct construct was re-transformed into BW27749 ΔentF and selected in liquid M9 with 50 μM EDDHA. Plasmid DNA was recovered, and the RS-encoding region was sequenced (Figure S1D).

Phenotypic characterization.

MIC_EDDHA assays were done in technical triplicate. BW27749 ΔentF strains were grown overnight in LB, sub-cultured into LB, grown until an OD₆₀₀ of ~0.5, normalized to an OD₆₀₀ of 0.5, and 0.5 µL was inoculated into 200 μL of M9 with 0.4% glycerol (v/v), antibiotic, and 250, 350, 400, 450, 500, or 600 μM EDDHA. Cultures were inoculated through an Excel Scientific AeraSeal sterile membrane, covered with a second membrane, and incubated at 37 °C and 250 RPM for 45 hrs. If the OD₆₀₀ rose to ≥ 0.2, from a calculated starting OD₆₀₀ of 0.01, the strain was considered to have grown. Strains that grew at 600 μM EDDHA were tested at 600, 650, 700, 750, 800, and 850 μM EDDHA. Strains that did not grow at 250 μM EDDHA were tested at 50, 100, 150, and 200 μM EDDHA.

Sequence analysis of entF mutants.

Each library was transformed into DH5α and the RS-encoding region of ~100 transformants per library was sequenced. A chi-squared test was used to check for statistically significant skews away from perfect randomization of the NNK sites in terms of the usage of nucleotides, codons, and the corresponding amino acids from. The nucleotide usage frequencies, combined across libraries but not sites, were used to make an adjusted genetic code. This genetic code was used to formulate the expectations for comparison to the selection output. For the remaining statistical analyses, the dataset of 225 DNA-unique entF mutants was considered. To determine if selection influenced the residues allowed at each site, the amino acid residue usage at each site from this dataset was compared to the adjusted genetic code, further adjusted for the exclusion of the stop codon, using a chi-squared test. To determine if selection influenced the relative proportions of the amino acids observed in the output, codes containing residues used fewer than five times at a given site were removed. This dataset was compared to the adjusted genetic code, adjusted for the exclusion of any unobserved residues, in addition to the stop codon, using a chi-squared test.

The amino acid diversity at each mutagenized specificity code site was compared between libraries and to a dataset of 82 characterized, L-Ser-specific A domains (27, Figure 2, and Table S1). SANDPUMA (27) was used to search the 146,187 sequence-unique A domains available in GenBank and to identify 11,026 A domains predicted to be L-Ser-specific using Active Site Motif (ASM), Support Vector Machine (SVM), and profile Hidden Markov Model (pHMM) methods. Codes extracted from this set were searched for the codes of the EntF variants (Table S2). To identify any covariance between specificity code sites, the DNA-unique L4 dataset, with codes containing rare residues or non-code mutations removed, was used. For pairwise combinations of sites, this dataset was compared to the expectation of a proportional distribution of the possible substitutions at one site amongst those at each other site, using a chi-squared test. To detect the effect of specific residues at each site on the MIC_EDDHA of the associated strain, ANOVA was used in the Excel XLSTAT package with default parameters on the DNA-unique L4 dataset with codes containing rare residues or non-code mutations removed.

Construction of phylogenetic tree.

MAFFT v7.310 (38) was used for multiple sequence alignment of the 82 characterized L-Ser-specific A domains (27, Table S1), several with specificity for L-Ser analogs or β-Ala (included as an outgroup), as well as 70 found in GenBank with predicted specificity codes matching several of the variants using default parameters. The alignment was manually trimmed and realigned, and a phylogenetic tree was constructed with FastTreev2.1.9 (39) with default parameters, rooted on AAG02364.1 and edited using FigTree v.1.4.2 (http://tree.bio.ed.ac.uk/software/figtree/).

Overproduction and purification of EntF and DltA variants.

EntF variants (Table S3) were co-overproduced in E. coli BL21(DE3) with a C-terminal hexahistidine tag with the MLP, YbdZ. DltA proteins were overproduced in E. coli BL21(DE3) ybdZ::acc(3)IV with a C-terminal hexahistidine tag. All protein purifications were done as previously described (20,31,40). Protein concentration was determined by BCA assay (Pierce).

Radiolabeled ATP/PP_i assays of the EntF and DltA variants.

ATP/PP_i exchange assays were performed as previously described (31,40). Variants (Table S3) were assayed in duplicate for substrate activation against four pools of five amino acids, Pool 1: SATPG, Pool 2: VLIMC, Pool 3: NDEQF, Pool 4: YWHKR, and then with individual amino acids from Pool 1. The apparent K_m and apparent V_max of the variants were determined in the linear range for product formation and 8 min reaction time (Figure S3), and calculated using non-linear regression analysis (GraphPad Prism ver. 6.0h).

IPRO energy score calculations.

Computational models of wild-type EntF and 157 variants were constructed in complex with the L-Ser or Ser-AMP (substrates). The reported structure of EntF (29, PDB 5JA1) with serine adenosine vinylsulfonamide (Ser-AVS) was used as the model. Energy-minimized structures of the EntF variants were generated using the Mutator module of the Iterative Protein Redesign and Optimization Suite of programs, IPRO (41). The complexes were energy minimized using CHARMM-force field (42). The CHARMM-based interaction energy scores (or IPRO energy scores) between a variant and substrate was computed as a sum of pairwise additive, non-bonded energy terms accounting for (a) van der Waal’s, (b) electrostatics, and (c) implicit solvation using the Generalized-Born implicit solvation method (43). This is conceptually akin to RosettaLigand (44) which reports Rosetta scores for the non-covalent forces between enzyme and ligand at the binding pocket as an in silico analogue of substrate affinity. Following side-chain conformation alterations done in Mutator, the location of substrates was re-adjusted using improved rigid-body docking (45) by randomly perturbing substrates along and around the X, Y, and Z axes using a Gaussian distribution centered at zero, with standard deviations of 0.2, 0.2, and 2.0 Å, respectively. 500 iterations are performed with subsequent interaction energy (binding score) re-calculation.