Mining and Statistical Modeling of Natural and Variant Class IIa Bacteriocins Elucidate Activity and Selectivity Profiles across Species

Class IIa bacteriocin antimicrobial peptides (AMPs), an alternative to traditional small-molecule antibiotics, are capable of selective activity toward various Gram-positive bacteria, limiting negative side effects associated with broad-spectrum activity. This selective activity is achieved through targeting of the mannose phosphotransferase system (manPTS) of a subset of Gram-positive bacteria, although factors affecting this mechanism are not entirely understood. Peptides identified from genomic data, as well as variants of previously characterized AMPs, can offer insight into how peptide sequence affects activity and selectivity. The experimental methods presented here identify promising potent and selective bacteriocins for further evaluation, highlight the potential of simple computational modeling for prediction of AMP performance, and demonstrate that factors beyond manPTS sequence affect bacterial susceptibility to class IIa bacteriocins.

ABSTRACT

Class IIa bacteriocin antimicrobial peptides (AMPs) are a compelling alternative to current antimicrobials because of potential specific activity toward antibiotic-resistant bacteria, including vancomycin-resistant enterococci. Engineering of these molecules would be enhanced by a better understanding of AMP sequence-activity relationships to improve efficacy in vivo and limit effects of off-target activity. Toward this goal, we experimentally evaluated 210 natural and variant class IIa bacteriocins for antimicrobial activity against six strains of enterococci. Inhibitory activity was ridge regressed to AMP sequence to predict performance, achieving an area under the curve of 0.70 and demonstrating the potential of statistical models for identifying and designing AMPs. Active AMPs were individually produced and evaluated against eight enterococcus strains and four Listeria strains to elucidate trends in susceptibility. It was determined that the mannose phosphotransferase system (manPTS) sequence is informative of susceptibility to class IIa bacteriocins, yet other factors, such as membrane composition, also contribute strongly to susceptibility. A broadly potent bacteriocin variant (lactocin DT1) from a Lactobacillus ruminis genome was identified as the only variant with inhibitory activity toward all tested strains, while a novel enterocin variant (DT2) from an Enterococcus faecium genome demonstrated specificity toward Listeria strains. Eight AMPs were evaluated for proteolytic stability to trypsin, chymotrypsin, and pepsin, and three C-terminal disulfide-containing variants, including divercin V41, were identified as compelling for future in vivo studies, given their high potency and proteolytic stability.

IMPORTANCE Class IIa bacteriocin antimicrobial peptides (AMPs), an alternative to traditional small-molecule antibiotics, are capable of selective activity toward various Gram-positive bacteria, limiting negative side effects associated with broad-spectrum activity. This selective activity is achieved through targeting of the mannose phosphotransferase system (manPTS) of a subset of Gram-positive bacteria, although factors affecting this mechanism are not entirely understood. Peptides identified from genomic data, as well as variants of previously characterized AMPs, can offer insight into how peptide sequence affects activity and selectivity. The experimental methods presented here identify promising potent and selective bacteriocins for further evaluation, highlight the potential of simple computational modeling for prediction of AMP performance, and demonstrate that factors beyond manPTS sequence affect bacterial susceptibility to class IIa bacteriocins.

INTRODUCTION

Antibiotic resistance is a growing threat to global public health. Reports estimate that over 2 million people in the United States suffer from antibiotic-resistant infections annually, resulting in 23,000 deaths and up to $20 billion in excess costs (1, 2). It is projected that 10 million annual deaths globally will be caused by antimicrobial resistance by the year 2050, demonstrating the worsening state of this problem (2). Improper use of antibiotics and the lack of discovery of new antimicrobial therapies are the two main factors driving this development. Widespread use of antibiotics, including medically relevant antibiotics, in animals for food production creates an unnecessary selective pressure for the development of antibiotic resistance (3, 4). Similarly, broad-spectrum antibiotics, antibiotics which have antimicrobial activity toward a wide array of pathogenic and commensal bacteria alike, are overprescribed and overused, further contributing to the development of resistance. These mechanisms of antibiotic resistance can then spread to clinically relevant, pathogenic bacterial populations through horizontal gene transfer (5). Further exacerbating this situation, few new antibiotics have been discovered due to limited financial incentives and the high risk associated with their development for pharmaceutical companies (6, 7). Given these circumstances, there is a dire need for novel classes of selective antimicrobial therapies.

Antimicrobial peptides (AMPs) are one compelling alternative to traditional, small-molecule antibiotics, either as individually produced free peptides (8–10) or through engineered cellular delivery at the site of infection (11–14). AMPs offer a variety of different mechanisms of antimicrobial activity, with a range of selectivity, potency, and cytotoxicity levels. While the potency of these molecules demonstrates their ability to replace or support current antibiotics, the high selectivity of these molecules is particularly important, both to minimize unnecessary selective pressure on the broader bacterial populations and to limit negative health effects associated with the broad-spectrum activity of small-molecule antibiotics. For example, it has been demonstrated that broad-spectrum antibiotics vastly reshape the microbiota (15, 16), allowing for opportunistic infections, such as Clostridium difficile, to develop, while use of more selective antimicrobial therapies, such as the two-peptide bacteriocin thuricin CD, results in much lower perturbation of the gut microbiota (17, 18). While natural evolution has developed families of AMPs with selective activity to various bacterial species, further work is necessary to optimize this activity toward clinically relevant bacterial strains while minimizing off-target activity to broader bacterial populations. Previous work has been conducted toward this aim with several AMPs, such as the screening of random variants of the lantibiotic nisin for improved activity and selectivity to multiple Gram-positive bacterial strains (19), the generation of synthetic AMPs containing species-specific targeting moieties (20), and the evaluation of single- and multimutant microcin J25 variants for reduced activity to commensal Escherichia coli strains while retaining high activity to pathogenic Salmonella (21), among others.

Class IIa bacteriocins are one promising family of AMPs characterized by small size (30 to 50 amino acids), a highly conserved N-terminal alpha-helical domain with a more variable C-terminal domain, and potent, selective activity to various strains of Gram-positive bacteria (22–25). Class IIa bacteriocins are known to interact directly with the mannose phosphotransferase systems (manPTSs) of various Gram-positive bacteria. Specifically, class IIa bacteriocins interact with the EIIC and EIID components of manPTSs to induce pore formation in the bacterial membrane and ion leakage, ultimately resulting in cell death (26–28). Previous work with the class IIa bacteriocin pediocin PA-1 has demonstrated that these AMPs are capable of effectively targeting Listeria in vivo with limited impact on the commensal bacteria (10, 29). Further work optimizing class IIa bacteriocins has greatly expanded our understanding of how the sequence of an AMP affects its activity. Multiple studies have focused on mutating natural class IIa bacteriocins to identify residues essential for activity (30), assess the beneficial impact of charged residues on target-cell binding and activity (31), and identify variants with improved activity or stability (32–35). Additionally, it has been well demonstrated that susceptible bacteria readily develop resistance to class IIa bacteriocins through downregulation of target manPTSs and other mechanisms, such as shifts in metabolic activity or cell membrane adaptation (36–38). However, it is hypothesized that combinatorial treatments of class IIa bacteriocins with AMPs or antibiotics that act through different mechanisms will aid in limiting resistance (39–41).

While class IIa bacteriocins show promise as novel antimicrobial therapies, much work remains to develop efficacious molecules for in vivo treatment. Proteolytic stability is a key limitation for peptide therapies in vivo (42, 43), and the ability to engineer peptides that retain antimicrobial activity with increased proteolytic stability is lacking. Toward this goal, further work exploring AMP sequence-activity relationships is necessary to use in parallel with the understanding of protein sequence-stability relationships. Furthermore, factors affecting the susceptibility of bacterial strains to particular AMPs must be better understood to aid the engineering of AMPs with selective targeting of harmful bacteria to limit negative health effects caused by broad-spectrum activity and to slow the development of resistance. Expansive evaluation of class IIa bacteriocins would inform these elements broadly and identify new candidate therapeutics. Thus, in this study, we evaluated a library of natural and variant class IIa bacteriocins for antimicrobial activity across an array of bacterial strains. We used ridge regression to predict AMP performance from sequence and further analyzed the susceptibility of strains of enterococci and Listeria to class IIa bacteriocins to elucidate the extent of the impact the manPTS sequence has on susceptibility.

RESULTS

AMP library design contains broad coverage of class IIa bacteriocin sequence space.

We sought to evaluate class IIa bacteriocin AMPs to expand our understanding of sequence-function relationships and to identify promising lead therapeutics. We hypothesized that evaluation of a broad array of sequence diversity would allow us to elucidate sequence-activity relationships across different bacterial strains and identify AMPs with desirable selectivity profiles. To achieve the desired breadth, we constructed an AMP library containing 150 previously identified class IIa bacteriocins from the UniProt database (44) as well as random and rationally designed variants of six previously studied bacteriocins. We included variants that have been previously characterized to serve as positive controls while exploring entirely novel variants identified from genomic information. The high prevalence of randomly mutated class IIa variants (900 random variants out of 1,150 total variants within the library) enabled us to explore the class IIa bacteriocin tolerance to random mutagenesis. Rationally designed AMPs were used to evaluate the utility of domain swapping for identifying compelling class IIa bacteriocins and our ability to design active and proteolytically stable variants using PeptideCutter software to identify protease cleavage sites (45). Domain swapping has been previously used in literature to identify functional class IIa bacteriocins, so we sought to further explore this method (33, 46). Additionally, proteolytic stability is one known limitation of class IIa bacteriocins for in vivo applications (47, 48), so rational design of highly stable lead molecules would be a useful advance.

Six previously characterized class IIa bacteriocins (Table 1) were selected as the basis for generation of random and rational variants. Here, we denote these six AMPs seed sequences as they seeded the random and rational AMP libraries. These six bacteriocins were chosen because all of them have been previously studied multiple times (30, 33, 49–52), allowing for comparisons with other studies, and have been shown to have relatively high potencies, suggesting that these are good starting points for engineering. While the amino acid sequences used for enterocin A, enterocin NKR-5-3C, enterocin P, and divercin V41 were the natural AMP sequences, inactive variants of pediocin PA-1 (D17N) and sakacin P (K11E) (pediocin PA-1_D17N and sakacin P_K11E, respectively) were used as seed sequences (30, 33). The inclusion of pediocin PA-1_D17N and sakacin P_K11E allowed both the inclusion of known negative controls within the library screen and exploration of whether any gain-of-activity variants could be found.

TABLE 1.

Alignment of seed and consensus sequences and identification of N-terminal, interior, and C-terminal regions^a

^aDots represent agreement with the consensus sequence.

To generate class IIa chimeras, the six seed sequences were aligned, the N-terminal, internal, and C-terminal domains were defined, and the domains were swapped between all six seed sequences. All unique N-terminal domains of seed sequences were swapped between the internal and C-terminal domains of all active seed sequences to generate 20 chimeras. To explore possible interactions between the ATR motif at positions 1 to 3 in the N-terminal domain and the Y7S mutation unique to enterocin P, chimeras were constructed with the enterocin P N-terminal domain and each interior and C-terminal domain with a Y7S mutation. Thirty chimeras were generated by swapping all unique N-terminal and C-terminal domains with the consensus (53–55) sequence of all class IIa bacteriocins identified in the UniProt database. Random variants were generated containing 2, 4, or 6 simultaneous mutations in seed sequences (150 variants for each seed sequence, for a total of 900 random variants).

The genes encoding the described natural, rational, and random class IIa bacteriocin variants were synthesized as a pool of oligonucleotides. The genes were amplified, cloned into the pNZC expression vector with a Usp45 signal peptide (56, 57), and transformed into Lactococcus lactis for expression. The chloride-inducible pNZC expression vector was chosen to act as a constitutive expression vector, given the high chloride concentration in brain heart infusion (BHI) medium, and for convenient use in L. lactis (56). L. lactis was chosen for AMP expression because of its status as a model lactic acid bacterium and its demonstrated ability to deliver peptides in vivo in a contained system (58–60). To sample the library, individual, random colonies were grown in each well of a total of 13 deep 96-well plates. Whole-cell PCR was conducted on each well to append plate, row, and column indices to identify AMP sequences using high-throughput sequencing. A total of 940 wells (75%) could be confidently identified, with a total of 309 unique AMP sequences, 166 of which were in the initial oligonucleotide pool (53%). This low frequency of sequences from the initial oligonucleotide pool (166/1,130, or 15% of oligonucleotide pool sequences) most likely stemmed from initial bias in the oligonucleotide pool, which may have been amplified during PCR steps. The remaining 143 unique sequences consisted of point mutants which resulted in sequences highly similar to sequences in the oligonucleotide pool (44 sequences), insertion/deletion mutants (55 sequences), point mutants which resulted in premature stop codons (39 sequences), and sequences containing DNA construction errors (5 sequences). While this frequency of erroneous sequences is higher than expected, a total of 210 identified sequences (see Table S2 in the supplemental material) were within expected class IIa sequence space and were used for further analysis. While alternative gene synthesis or pool sampling approaches could enable even deeper coverage of the proposed population, the set of 210 represents a broad set of IIa variants.

AMP library exhibits broad activity towards enterococcus and varied tolerance to random mutagenesis.

To evaluate the isolated class IIa bacteriocins, we performed agar diffusion assays to categorize variants as highly active, moderately active, or inactive, followed by more precise characterization of all active variants. Six enterococcus strains, four Enterococcus faecalis and two Enterococcus faecium strains, were used as indicators for reasonable breadth and to provide an opportunity to assess strain and species selectivity. The selected strains were chosen as a relevant sampling of pathogenic and nonpathogenic strains, many of which have known antibiotic resistance.

Sixteen of 20 (80%) natural class IIa bacteriocins displayed some level of inhibitory activity toward at least one strain (Table S2). The high rate of activity is expected, given that several of these variants have been previously shown to be active. However, eight of the tested AMPs have not been previously characterized to our knowledge, and five of these were found to be active. Thus, our library design effectively achieved one of its goals of expanding the active IIa repertoire. As for the four natural IIa bacteriocins observed as inactive against all six strains, one was the negative-control variant seed sequence of pediocin PA-1_D17N while two others were fragmented variants of enterocin A. These data provide a partial constraint on their known activities in regard to partial selectivity or ineffective expression from the L. lactis host while also identifying limitations of genomic annotation to discover active AMPs. Four of the six parental seed sequences were observed in library screening, with enterocin A and sakacin P_K11E being unobserved.

Thirteen of 183 (7%) random variants were active (Table S2). Of the 103 random variants whose parental sequences were observed as active, 11% retained some level of inhibitory activity toward at least one indicator strain (Fig. 1A). The active frequency was substantially higher for double mutants (10/30, or 33%) than for quadruple or hexamutants (2% or 0%, respectively). These results confirm relatively high tolerance to two mutations but low tolerance to accumulated mutation (61, 62). One hexamutant, sakacin P_{K11E Y2V G4D N5V N24C I25G N27E}, exhibited activity to one strain; however, because the parental seed sequence sakacin P_K11E was unobserved, this AMP was not included in the data shown in Fig. 1A, and it is unclear whether the parental sequence had any activity.

FIG 1.

Summary of observed AMP activity from library screen displays broad activity to all tested strains. (A) Activity of observed random mutants suggests that bacteriocin IIa AMPs have a relatively low tolerance to mutagenesis. (B) Most AMPs display an increasing potency across all indicator strains, with limited selectivity. We define a selectivity index here as the activity score toward one strain divided by the average activity score across the other five strains. (C) Average AMP activity ordered by rank across all six indicator strains. Line thickness indicates the number of observations for a particular AMP. (D) Comparison of AMP activities between E. faecalis and E. faecium indicator strains. The sum of activity scores for E. faecalis and E. faecium strains is shown on the y and x axes, respectively. The red line represents an equal fraction of activity toward both sets of strains. The gray zone represents a zone of 50% selectivity between E. faecalis and E. faecium strains. Six AMPs (numbered 1 to 6) fall outside the 50% selectivity zone: enterocin DT1 (1) and enterocin DT3 (2) showed selectivity to E. faecium while lactocin DT2 (3), enterocin NKR-5-3C_{S12A E18D} (4), enterocin NKR-5-3C_{S12T T22A} (5), and enterocin NKR-5-3C (6) had selectivity toward E. faecalis.

Incomplete library sampling hindered our ability to extensively evaluate the impact of chimeric design or proteolytic stability design. Only six domain-swapped variants and one enterocin A stability variant were observed in library evaluation. Three domain-swapped variants retained activity toward multiple strains, which is suggestive of tolerance, yet the limited sampling of chimeric options precludes a robust conclusion. Interestingly, all three inactive domain-swapped variants included the consensus interior region, suggesting that critical residues may have been lost in design of the consensus region. The one observed enterocin A stability variant was inactive across all six observations.

To identify selective AMPs and trends in class IIa bacteriocin activity, the activity scores of all AMPs (0 for inactive, 0.5 for inhibition less than that of enterocin P, and 1.0 for activity comparable to or greater than that of enterocin P) were averaged over all observations and analyzed across all six indicator strains (Fig. 1). Most active AMPs exhibited broad-spectrum activity toward multiple indicator strains with limited selectivity (Fig. 1B and C; Table S2). Twenty-three of 32 AMPs exhibited an activity score of $\ge$0.2 against at least 4 of 6 strains, while 17 AMPs had an activity score of $\ge$0.5 against at least 4 strains. Yet multiple AMPs exhibited strong specificity: five AMPs were observed with activity toward only one strain, including enterocin P_{N7V Y10E} and enterocin NKR-5-3C_{G20V I31M}, with activity scores of 0.5 and 0.43, respectively. In comparison to the parental enterocin P and enterocin NKR-5-3C activity scores, these mutations drastically reduced observed activity across all six indicator strains, with activity toward only one strain being retained. In contrast to these selective AMPs, only five AMPs exhibited a full breadth of activity, with scores of $>$0.5 across all six indicator strains, including enterocin P and two enterocin P variants: G6D K15G and K15R V31S (Fig. 1C; Table S2).

Across species, most AMPs displayed comparable levels of activity toward both E. faecalis and E. faecium strains (Fig. 1D). The strongest exception was enterocin DT3, an AMP from Enterococcus pallens, which demonstrated higher activity toward the two E. faecium indicator strains than the four E. faecalis strains (Fig. 1D). One additional variant exhibited an appreciable E. faecium preference, while four AMPs exhibited an E. faecalis preference.

Class IIa bacteriocins show limited specificity between species.

To more thoroughly determine AMP activity and specificity profiles, all 32 active AMPs were individually produced, and their total activities were quantified against eight strains of enterococci and four strains of Listeria monocytogenes. L. monocytogenes strains were included to evaluate specificity trends between enterococci and Listeria, and an additional two E. faecium strains were added to have an equal sampling of E. faecium, E. faecalis, and L. monocytogenes strains. We chose to produce the unmodified AMPs in L. lactis cultures and conduct ammonium sulfate (AS) precipitation to generate more concentrated samples rather than use purification tags, which may affect structure or activity of small class IIa AMPs. Given this method of production, we quantified total activity as the minimum inhibitory dilution (MID), which is the lowest dilution of resuspended AS precipitation solution that inhibited growth. This metric of total activity is the product of an AMP’s ability to be produced by L. lactis and inhibit growth, which is directly meaningful to intended applications with in situ cellular delivery. To ensure that this was comparable across all AMPs, L. lactis growth conditions and AS precipitation resuspension volumes were kept constant. MIDs were determined using serial dilutions of resuspended AS precipitation products in an agar diffusion experiment, and the lowest dilution which led to formation of any zone of inhibition was determined to be the MID.

Several tested AMPs showed comparable levels of activity to nearly all tested indicator strains, suggesting a lack of specificity. Seventeen of 32 AMPs (53%) displayed inhibitory activity toward at least 10 of the 12 indicator strains. Two AMPs were especially potent, with MIDs of $\le$0.1 against 10 of the 12 indicator strains, correlating to a 10-fold dilution of resuspended AS precipitation product inhibiting bacterial growth. Only one AMP, lactocin DT1 from Lactobacillus ruminis, had activity toward all 12 indicator strains, albeit with limited activity against E. faecium strain NRRL B-2354 and L. monocytogenes strain V7. Hiracin JM79 had an MID of $\le$0.04, displaying high potency toward all strains except L. monocytogenes strain V7, toward which it was inactive. Interestingly, E. faecium strain NRRL B-2354 and L. monocytogenes strain V7 were tolerant to nearly all class IIa bacteriocins tested. This is particularly interesting for E. faecium strain NRRL B-2354 because it shares identical and nearly identical manPTS EIIC and EIID genes with E. faecium strains 8E9 and 6E6, respectively, both of which were highly susceptible to class IIa bacteriocins (Fig. 2; see also Fig. 4).

FIG 2.

Summary of MID data identifies AMPs capable of selective activity toward Listeria. MID values represent the lowest fraction of AMP AS precipitate solution that results in growth inhibition. The gradient of blue to red represents the most to least potent MID values. AMP names highlighted in gray are identified as selective. Dots in the alignments show sequence agreement with consensus sequence.

FIG 4.

Alignment of manPTS EIIC (A) and EIID (B) sequences. The genes for E. faecium 8E9 were amplified using primers designed from the E. faecium 6E6 manPTS sequences (see Table S3 in the supplemental material) and Sanger sequenced. All other manPTS genes were identified from genomic information. Dots in the alignments show sequence agreement with consensus sequence.

While no AMPs displayed selective targeting of E. faecium strains, several AMPs displayed low activity to E. faecium while retaining high activity to several E. faecalis and/or L. monocytogenes strains (Fig. 2). Three AMPs displayed activity to seven of eight E. faecalis and L. monocytogenes strains, with limited activity to only one E. faecium strain. Six AMPs displayed at least a 4-fold increase in potency to the three susceptible L. monocytogenes strains relative to the levels of activity against any other strains tested, while a natural enterocin variant, enterocin DT2, displayed at least a 25-fold increase in potency to the three susceptible L. monocytogenes strains compared to the levels of all other strains tested. Of note, only two gain-of-activity variants were observed: enterocin P_{G6D, K15G} and enterocin P_LM2. To identify class IIa bacteriocin sequence features that affect specificity, we analyzed these selective variants for sequence motifs in the C-terminal domain as previous literature suggests that the C-terminal domain may act as a targeting domain for specific manPTS genes (51, 63). Interestingly, only two active AMPs contained the C-terminal sequence motif GGFGGR, and both of these had higher activities toward the three susceptible L. monocytogenes strains than to the seven susceptible enterococcus strains. In comparison, 18 of 32 active AMPs contained either the GGA(I/V)PGKC or GLAGMGH C-terminal sequence motifs, and these AMPs commonly had comparable MIDs across many strains of different species (Fig. 2). These trends support the hypothesis that C-terminal sequence motifs of class IIa bacteriocins play a strong role in determining their ranges of activity toward different species. However, it remains unclear if this trend is due to stronger/weaker interactions between these AMP sequence motifs and the manPTSs, the cell membrane, or another factor specific to various bacterial species. Ultimately, most class IIa bacteriocins displayed comparable potency levels across most strains tested, with limited selectivity (Fig. 3A). However, when activity to species is considered, some AMP variants display equivalently increasing potency and specificity (Fig. 3B). The line defined by y = x in Fig. 3B is the limit of maximal selectivity, given that the lowest observed activity was an MID of 1 (MID values of >1 were treated as 1 for selectivity calculations to eliminate infinite values). Several AMPs exhibit high potency while falling near the y = x line, exhibiting nearly the maximum measurable selectivity. These AMP variants suggest that class IIa bacteriocin activity can be tailored to different species.

FIG 3.

AMPs display minor selectivity at a species level but limited selectivity at a strain level. (A) Maximum AMP potency toward any strain plotted against maximum selectivity toward any strain. All AMPs fall above the line y = x, showing a broadly increasing potency toward several strains with limited selectivity. (B) Maximum average potency toward any species plotted against maximum selectivity toward any species. Some AMPs fall near the line y = x, displaying increasing potency with selectivity toward a given species.

Ridge regression of AMP sequence-activity data shows that manPTS sequence is as predictive as strain identity.

Beyond using this broad analysis of class IIa bacteriocin sequence space to identify compelling AMPs, we aimed to advance the understanding of sequence-activity relationships. We hypothesized that a generalized linear model could identify such relationships to predict AMP performance. To test this hypothesis, the activity of all 210 observed AMPs was ridge regressed to AMP sequence in four separate models. Ridge regression was used to minimize overfitting and because initial ridge regression models were shown to offer improved prediction over lasso regression models. The four different models were chosen to test which set of information best trains a ridge regression model and were assessed via a receiver operating characteristic (ROC) curve. In the first model, AMP sequences were regressed to their binary activity toward all indicator strains; i.e., if an AMP had activity toward at least one indicator strain, it was classified as active. This model performed only moderately above random, with an area under the ROC curve (AUC) of 0.60 ± 0.02 (Fig. S4). The second set of models (one for each strain) independently evaluated activity for each strain; a nearly equivalent average AUC of 0.59 ± 0.01 indicated that independent strain-specific modeling does not aid predictive power. We then evaluated a model in which strain identity was encoded as an input along with sequence, and all data were jointly modeled. This approach elevated the AUC to 0.70 ± 0.04 (P = 0.04 versus model 2), which reveals that strain-specific information is useful in distinguishing AMP activity toward the different strains tested. The fourth model replaced strain identity by manPTS sequence encoding, which did not further elevate model performance (AUC = 0.69 ± 0.04). Thus, specific manPTS sequence information provided value equivalent to strain identity. The equivalent strength of manPTS sequence information is consistent with its hypothesized dominant role in dictating susceptibility to class IIa bacteriocins. While moderate predictive performance was achieved by models 3 and 4, no obvious trends in charge, hydrophilicity, or polarity of beneficial amino acids at certain positions were identified.

manPTS sequence does not fully define susceptibility to class IIa bacteriocins.

While most class IIa bacteriocins displayed limited selectivity, new molecules were characterized that displayed selective activity to Listeria strains, suggesting that class IIa bacteriocin activity can be tailored to certain species. Therefore, we sought to identify any trends in susceptibility to class IIa bacteriocins between species in the hope of elucidating the underlying mechanisms by which selectivity is achieved. We initially compared manPTS sequences between enterococcus strains and L. monocytogenes strain ATCC 51775, given that class IIa bacteriocins are known to interact with manPTS EIIC and EIID domains (26, 27). The manPTS genes from L. monocytogenes strain ATCC 51775 were used for this analysis as this is the only tested Listeria strain with available genomic sequence data. More significant differences in manPTS sequences between L. monocytogenes and enterococci than between E. faecium and E. faecalis were observed, which may contribute to increases in Listeria susceptibility to class IIa AMPs (Fig. 4). However, given the manPTS sequences were used from only one Listeria species, more data points are necessary to confirm these trends.

To further assess the specificity of AMPs across bacterial strains and species, the MID values for all active AMPs were pairwise compared across all tested strains (Fig. 5). Linear trends appear when the MIDs between strains of the same species were compared (Fig. 5; Fig. S5). Given that nearly all strains within a particular species share nearly identical manPTS sequences (Fig. 4), this is not surprising. However, for some strains with identical manPTS sequences, susceptibility could still differ by an order of magnitude. Examples of this are the MIDs of the AMP enterocin NKR-5-3C_LM4 to the four E. faecalis strains, for which values ranged from 0.028 to 0.4 despite identical manPTS sequences. More telling is the comparison of activities of several AMPs to E. faecium strains 8E9 and NRRL B-2354, in which strain 8E9 was found to be highly susceptible to class IIa bacteriocins while strain NRRL B-2354 appeared tolerant to nearly all class IIa bacteriocins even though the strains have identical manPTS sequences.

FIG 5.

Activity of all AMPs plotted between each individual strain shows susceptibility trends between species. The MIDs of each AMP in bacteriocin units were plotted between all individual strains. A bacteriocin unit is defined here as the inverse of the lowest active fraction of AMP AS precipitate solution. One-to-one linear trends appear between most strains of the same species, shown in the subgroups near the diagonal. The inset shows an example plot of activity of all 32 AMPs toward strains 6 (x axis) and 8 (y axis). Given that both strains 6 and 8 are E. faecalis strains, a strong trend similar to y = x appears. Plots for strains 4 and 12 fall nearly vertical or horizontal due to the low susceptibility observed for these strains. Strains are identified by number as follows: 1, E. faecium 8E9; 2, E. faecium 6E6; 3, E. faecium 7A; 4, E. faecium NRRL B2354; 5, E. faecalis V583; 6, E. faecalis CH116; 7, E. faecalis Pan7; 8, E. faecalis Com1; 9, L. monocytogenes ATCC 51775; 10, L. monocytogenes M-03-1213B-1; 11, L. monocytogenes CDC 7762; 12, L. monocytogenes V7. Figure S5 in the supplemental material shows a circle plot displaying correlation coefficients and slopes of linear fit of all subplots.

When susceptibilities between different species were compared, the tested E. faecalis strains appeared to have a larger correlation in levels of susceptibility with L. monocytogenes strains than with E. faecium strains (Fig. S5). This trend is observed even though the L. monocytogenes strain ATCC 51775 manPTS EIIC and EIID sequences have only 88% and 80% sequence similarity with E. faecalis manPTS EIIC and EIID sequences, respectively (Fig. 4; Fig. S6). Conversely, the minor correlation in susceptibility to class IIa bacteriocins observed between E. faecium strains 1 to 3 and E. faecalis strains 5 to 8, identified in the legend of Fig. 5, would be expected to be higher given their manPTS EIIC and EIID sequence similarities of 93% and 92%, respectively. E. faecium and L. monocytogenes strains show very little correlation in susceptibility to class IIa bacteriocins and have 89% and 78% similarity in manPTS EIIC and EIID sequences, respectively. These trends clearly demonstrate that factors other than manPTS sequence significantly impact susceptibility to class IIa bacteriocins.

C-terminal disulfide-containing class IIa bacteriocins are compelling for in vivo application.

Toward identifying compelling molecules for further study and evaluation in vivo, we evaluated the proteolytic stability of eight of the most potent AMPs against trypsin, chymotrypsin, and pepsin (Fig. 6). While class IIa bacteriocins are known to be very thermally stable, proteolytic stability is a common limitation of peptide therapies for in vivo applications (42, 43). Trypsin, chymotrypsin, and pepsin were selected because they are highly prevalent in the human digestive system (64). The eight AMPs were incubated briefly with various concentrations of the given protease, heated to inactivate the protease, and tested for loss of activity compared to that of an untreated control in an agar diffusion assay. The eight AMPs selected were chosen as a sampling of potent natural and variant AMPs with high potency toward L. monocytogenes strain ATCC 51775. This strain was selected for this assay as it was observed to be the most susceptible to class IIa bacteriocins during MID quantification.

FIG 6.

k_cat′/K_m values of trypsin, chymotrypsin, and pepsin against tested AMPs. Footnotes: a, MID values are those listed in Fig. 2; b, all protease sites were identified with PeptideCutter (45), all k_cat′/K_m values are in units of mM_enzyme⁻¹ mM_AMP⁻¹ s⁻¹, and values for gray boxes could not be quantified due to noise in the assay results. The gradient of blue to red represents the most to least stable k_cat′/K_m values.

Trypsin k_cat′/K_m values could be robustly determined only for six of the eight tested AMPs due to experimental noise in the assay. Of these, four AMPs saw no loss in activity across the tested enzyme concentrations (≤256 μM); thus, the normalized proteolytic efficiency was <0.009 mM_enzyme⁻¹ mM_AMP⁻¹ s⁻¹. The LM2 variant of enterocin NKR-5-3C, which introduces an additional trypsin site via A1T T2K mutation, exhibits appreciable susceptibility. Hiracin JM79, which lacks a C-terminal disulfide, is even more susceptible. For chymotrypsin and pepsin, the parental enterocin NKR-5-3C had a greater susceptibility than both N-terminal chimeric variants tested (LM2 and LM4). All three variants had the same number of identified chymotrypsin cleavage sites, so this would not explain the observed differences. Additionally, the parental peptide and LM2 variants have identical chymotrypsin cleavage sites, so it is unclear what sequence features enabled higher proteolytic tolerance. The two other enterocin NKR-5-3C variants, S12T T22A and S12A E18D, showed proteolytic susceptibility similar to the parental level. Across all three proteases, divercin V41, which contains two disulfide bridges, had the highest proteolytic stability. Among variants, both N-terminal chimeric enterocin NKR-5-3C variants displayed the highest stabilities across the tested proteases. Thus, all three AMPs are compelling molecules for further evaluation in vivo, given their high potencies and stabilities.

DISCUSSION

The results shown here expand our database of characterized AMPs for analysis, identify compelling molecules for further study and optimization, demonstrate the utility of statistical modeling in AMP engineering efforts, and highlight the significance of factors beyond manPTS receptors in determining bacterial susceptibility to class IIa AMPs. Although only a small fraction of our initial library was isolated, 210 unique class IIa bacteriocins were evaluated for inhibitory activity to six enterococcus strains in a L. lactis secretion agar diffusion assay. While several of these AMPs were previously characterized, many more were novel variants identified from genomic data or variants containing multiple amino acid mutations from a parental sequence. The activity data from random variants suggest that random mutagenesis of class IIa bacteriocins should limit the number of simultaneous amino acid mutations in situations where throughput is limited to two or three residues in any variant to increase the likelihood of finding functional variants. Additionally, the enhanced selectivity achieved by double mutants of enterocin P (N7V Y10E) and enterocin NKR-5-3C (G20V I31M) may suggest that a more efficient route for generating AMPs with selective activity will be to introduce selectively deleterious amino acid mutations into highly potent AMPs to abolish activity to off-target strains. As for chimeras, our incomplete sampling of chimeric class IIa variants hinders firm conclusions; however, the high success rate, albeit among only six observations, and previous literature utilizing chimeric class IIa bacteriocins suggest (i) that chimeric class IIa bacteriocins are promising stepping stones to generate improved-activity variants (33, 65, 66) and (ii) that class IIa bacteriocins may be segmented into domains which serve distinct functions, as has been previously hypothesized.

While screening was in part conducted to identify highly potent class IIa bacteriocins, a key focus of this study was the investigation of selective activity of these AMPs. Ridge regression determined that inclusion of strain identity information was beneficial for predicting AMP performance, suggesting that differences in bacterial susceptibility could be predicted. The equivalency of predictive benefit from using manPTS sequences rather than strain identity information is consistent with a strong role for manPTS in determining bacterial susceptibility to class IIa bacteriocins, as was previously hypothesized. Yet further analysis of bacterial susceptibility to class IIa bacteriocins showed that strains with identical manPTS sequences could have significantly different susceptibilities to the same AMP. While some correlation was observed between manPTS sequence and susceptibility to class IIa AMPs, there are clear outliers which do not agree (see Fig. S6 in the supplemental material). This suggests that other factors, such as bacterial membrane composition or metabolism, play a more significant role in determining bacterial susceptibility to various AMPs than previously thought and should be further investigated in the future.

Activity and stability characterization of individual AMPs in the current study aimed at identifying compelling molecules to further pursue for development as in vivo therapeutics. We sought to identify AMPs capable of selectively targeting the individual species of E. faecium, E. faecalis, or L. monocytogenes based on the hypothesis that such AMPs may limit negative effects associated with off-target activity to commensal bacteria of other species. Yet most active AMPs displayed broad activity profiles toward all tested strains due to our reduced sampling of our AMP library and the broad activity of many previously characterized class IIa AMPs which seeded our library. We suspect that given the narrow range of commensal bacteria which possess the necessary manPTS genes encoding the target receptors for class IIa bacteriocins, the potent AMPs identified here are still compelling as in vivo therapies which limit activity to most commensal bacteria. Notably, however, class IIa bacteriocins with selective activity toward Listeria were also identified, which could further reduce activity to commensal enterococci in vivo. For future work developing proteolytically stable class IIa bacteriocins for in vivo efficacy, evaluations should focus on those containing C-terminal disulfides which reduce accessibility of common protease cleavage sites, and we have identified divercin V41 as one compelling molecule with increased potency and stability over that of other tested class IIa bacteriocins.

MATERIALS AND METHODS

Bacterial culture and strains.

Lactococcus lactis NZ9000 cells were grown in brain heart infusion (BHI) medium (W.W. Grainger, Inc.), which contained 1.6% (vol/vol) agar in the case of solid-phase growth. Cultures were grown under stationary conditions at 30°C. E. coli cells were grown in lysogeny broth (LB; Fisher BioReagents) which contained 1.6% (vol/vol) agar in the case of solid-phase growth. All E. coli cultures were grown at 37°C, with shaking of liquid cultures at 250 rpm. When specified, LB was supplemented with 5 μg/ml chloramphenicol. All enterococcus and Listeria strains were grown in liquid BHI medium at 37°C with shaking at 250 rpm. All bacterial strains and plasmids used are listed in Table S1 in the supplemental material.

Class IIa bacteriocin library design.

The class II bacteriocin protein family full alignment from the Pfam database (67) was used to seed a search in Jackhmmer (68) to identify class IIa bacteriocin sequences in the UniProt database (44). There was no taxonomic restriction, and iterations were performed with the hit threshold set to an E value of 0.01 until convergence to identify the maximum number of homologous sequences. The output of this search was parsed to eliminate duplicate protein sequences as well as sequences shorter than 30 amino acids or longer than 50 amino acids. A total of 150 remaining sequences were included in the library of class IIa bacteriocins. Sequence identifiers for natural AMPs are the UniProt entry identification numbers (Table S2). Natural AMPs which have been previously characterized are also identified in the text and Table S2 with their given names. All 150 class IIa bacteriocins from the UniProt database were included in library 1.

Enterocin A, enterocin P, enterocin NKR-5-3C, divercin V41, sakacin P_K11E, and pediocin PA-1_D17N were selected as seed sequences for the rational and random library designs; the first four AMPs were previously shown in literature to have moderate to high antimicrobial activity and, thus, were believed to be strong starting points for generating improved variants (33, 49, 50, 52, 69) whereas sakacin P_K11E and pediocin PA-1_D17N were chosen as inactive variants of the parental sakacin P and pediocin PA-1 for rational and random mutant library design to serve as negative controls and to evaluate the likelihood of finding mutations capable of restoring antimicrobial activity.

AMP variants were rationally designed to test several hypotheses regarding class IIa bacteriocin structure and function. AMP chimeras were designed by swapping all unique N-terminal domains of all active seed sequences. A consensus (53–55) interior domain was designed from all 150 class IIa bacteriocins identified from UniProt, and chimeras were constructed with this interior domain and all unique N-terminal and C-terminal domains. Enterocin A variants were designed for improved proteolytic stability. PeptideCutter software (Swiss Institute of Bioinformatics) was used to identify sites in enterocin A with a high susceptibility to trypsin, chymotrypsin, and pepsin (45). The most frequent susceptible residue identified was tyrosine, so all variants were designed such that all tyrosine residues, individually and combinatorically, were mutated to serines. All rational and chimera variants were included in library 1.

A library of 900 random multimutants was generated from the six seed sequences to evaluate the class IIa bacteriocin tolerance to random mutagenesis. The library was composed of 50 mutants containing two, four, or six simultaneous random mutations in the 33-amino-acid interior region of each of the six seed sequences. As an example, to generate the group of variants containing two mutations in enterocin A, two positions from 1 to 33 were randomly chosen and mutated to a random amino acid that was not the parental residue. This process was repeated to generate 50 variants containing two amino acid mutations, 50 variants containing four amino acid mutations, and 50 variants containing six amino acid mutations, for a total of 150 enterocin A random variants. The process was then repeated for the remaining five seed sequences, for a total of 900 random variants. For the inactive sakacin P_K11E and pediocin PA-1_D17N variants, positions 11 and 17, respectively, were not allowed to be mutated during this process. The group of 900 random multimutants was defined as library 2.

Oligonucleotide library construction.

AMP libraries 1 and 2 were synthesized as oligonucleotides by Twist Biosciences. The libraries were designed such that each library could be amplified through PCR independent of the other library using specific DNA primers. Following a 16-cycle PCR for initial amplification of each library, DNA encoding the Usp45 signal peptide was ligated upstream of the AMP as this signal peptide has been shown to enable high secretion efficiency in L. lactis (57). This construct was then amplified and assembled (HiFi; New England Biolabs [NEB]) into the chloride-inducible pNZC expression vector (56). Following assembly, the final DNA constructs were transformed into E. coli (5-alpha competent cells; NEB), and cells were grown overnight on LB-agar plates with 5 μg/ml chloramphenicol. Following overnight growth, all cells were collected from agar plates, and the DNA was extracted and stored at –20°C until further use. All primers used for library construction are included in Table S3.

Ninety-six-well stock plate preparation.

Final DNA constructs for libraries 1 and 2 were transformed into electrocompetent L. lactis cells prepared according to a published procedure (70). Following transformation, cells required roughly 48 h of growth at 30°C for colonies to be visible on agar plates. Individual colonies were plucked from agar plates and inoculated into wells of deep 96-well plates containing 1 ml of BHI medium. All plates contained two control wells that were inoculated with freezer stocks of L. lactis expressing enterocin P (positive control) and empty pNZC vector (negative control), respectively. Plates were incubated for 18 to 24 h at 30°C. Following growth to saturation, 100 μl of culture from each well was added to 100 μl of 60% glycerol in sterile 96-well plates to create a 30% glycerol stock plate. The plates were covered and stored at –80°C until further use. Five plates were prepared containing library 1 constructs, and 10 plates were prepared containing library 2 constructs.

Illumina sequencing and well identification.

Following 30% glycerol plate preparation, whole-cell PCR was conducted on 1 μl of culture from each well with primers that appended row, column, and plate indices adjacent to the coding region of each construct to identify AMP sequences through high-throughput sequencing. Nextera index adaptors N501 to N508 and N701 to N712 were used as row and column indices, respectively. Combinations of unique 5-bp and 4-bp sequences on the 5′ and 3′ ends of the coding region, respectively, were used as plate indices. PCRs were conducted in multiple 96-well PCR plates to allow for independent amplification of all wells using Q5 High-Fidelity DNA polymerase (New England Biolabs). PCR products for all wells of a plate were mixed, purified (QIAquick PCR purification kit; Qiagen, Hilden, Germany) to reduce sample volume, and gel extracted. DNA pools were then sequenced on two runs on an Illumina iSeq 100 system to identify AMP constructs present in each well of all glycerol stock plates. Sequencing on the Illumina iSeq 100 was conducted at the University of Minnesota Genomics Center. All primers used for DNA sequencing are included in Table S3.

Illumina iSeq sequencing generated approximately 7 million reads specific to libraries 1 and 2. Sequences were processed using Usearch by filtering for a maximum error rate of 0.001 and denoised using the unoise3 command to correct single-base pair errors which may have occurred during sequencing (71, 72). A second processing step was conducted to identify the individual AMPs in each well of the glycerol stock plates.

For each sequencing run, the distribution of reads of AMP constructs per well was calculated as some wells were poorly amplified during whole-cell PCR and, thus, were poorly represented in the sequencing data. Thresholds of 90 reads and 10 reads for the first and second sequencing runs, respectively, were used to distinguish true AMP identification from noise (Fig. S1). Unique sequences with reads above this threshold were isolated and analyzed via a custom MATLAB script. Sequencing primers annealed to the plasmid backbone before the signal peptide, so reads of plasmid containing no gene were ∼40 nucleotides. Reads containing true AMPs would include this backbone region, the Usp45 signal peptide, and the AMP of interest, so a 75-nucleotide cutoff was selected to distinguish empty vector from true constructs. All wells where the most frequent unique sequence was $\le$75 nucleotides and accounted for $\ge$60% of total reads were discarded as these wells contained empty vector. Wells with two or more unique sequences with lengths of $\ge$75 nucleotides but not accounting for $\ge$50% of total reads in the well were also discarded as multiconstruct wells. A well was identified as a single AMP construct if either of the following criteria were met:

• i.If the most frequent unique sequence was $\ge$75 nucleotides and accounted for $\ge$50% of the total well reads, the well was identified as the most frequent unique sequence; or
• ii.If the most frequent unique sequence was $\le$75 nucleotides (empty vector) and the second most frequent unique sequence was $\ge$75 nucleotides and accounted for 40% of the total well reads, the well was identified as the second most frequent sequence.

Following identification of wells containing single constructs, DNA sequences were trimmed to remove the Usp45 signal peptide and were translated into amino acid sequences for further analysis. AMP sequences of $\le$10 amino acids were excluded from analysis, but all other sequences were retained, including out-of-library constructs. Following the individual AMP identification process, whole-cell PCR was conducted on 40 random wells, and the products were Sanger sequenced to validate the sequence identification process.

Agar diffusion activity assays to measure growth inhibition.

Agar diffusion activity assays were conducted to test the inhibitory activity of every well of each glycerol stock plate against E. faecium 8E9, E. faecium 6E6, E. faecalis V583, E. faecalis CH116, E. faecalis Pan7, and E. faecalis Com1 (Table S1). One milliliter of BHI medium was added to each well of sterile, deep 96-well plates. Each well was then inoculated with cells from the corresponding well of a glycerol stock plate. Deep 96-well plates were incubated under stationary conditions for 18 to 24 h at 30°C.

For making bacterial agar plates, 5 ml of BHI medium was inoculated with cells from 30% glycerol stocks of enterococcus cultures to create starter cultures. Enterococcus starter cultures were incubated for 18 to 24 h at 37°C with shaking at 250 rpm. After overnight growth, a BHI-agar mixture was prepared by adding 1.6% (wt/vol) agar to BHI medium and autoclaving. The mixture was allowed to cool to ∼45°C, and then enterococcus culture was added at 0.05% (vol/vol) and mixed by inversion. Approximately 17 ml of the mixture was spread to a thin layer on fresh petri dishes and allowed to cool for 30 to 60 min at room temperature. Once solidified, 3 μl of overnight culture from each deep 96-well plate was deposited onto each pathogen plate, allowed to dry, and incubated under stationary conditions for 18 to 20 h at 37°C. Due to the size of agar plates, one 96-well plate had culture from wells split between three agar plates, as shown in Fig. S4. Following overnight growth, colonies were washed off each plate with 5 ml of sterile phosphate-buffered saline (PBS) for improved resolution of halo formation and size. Halo formation and size were then recorded relative to halo formation from the enterocin P positive-control wells to give AMP activity scores. Non-halo-forming wells were scored as 0, halos which were noticeably smaller than those of the positive-control wells were scored as 0.5, and halos which were comparable to or larger than the positive control were scored as 1 (Fig. S2). The scoring of halo size relative to an enterocin P positive control, which was included on all plates, allowed for interplate comparisons.

Individual AMP production and ammonium sulfate precipitation.

Unmodified AMPs were produced in L. lactis cultures and ammonium sulfate (AS) precipitated to generate more concentrated samples rather than through use of purification tags, which may affect structure or activity of small class IIa bacteriocins. Additionally, this method allowed the activity testing of many class IIa bacteriocins at low cost compared to that of chemical AMP synthesis. To produce individual AMP solutions, 40 ml of BHI medium was inoculated from glycerol stocks of AMP-producing or negative control, empty pNZC-containing L. lactis and incubated under stationary conditions overnight at 30°C. The culture was centrifuged at 3,500 × g for 5 min, the supernatant was discarded, and the pellet was resuspended in an equal volume of fresh BHI medium. The culture was then incubated under stationary conditions for 4 h at 30°C. Following incubation, the culture was centrifuged at 3,500 × g for 5 min, and the supernatant was sterile filtered. AS was added to the supernatant at 45% (wt/vol) to achieve a 70% saturated AS solution, which was rotated for 18 h at 4°C. The AS solution was centrifuged for 10 min at 11,000 × g, and the pellet was resuspended in 1 ml of ultrapure MilliQ water and heat sterilized at 98°C for 10 min. The resulting AS precipitation solutions were stored at –20°C until further use.

Agar diffusion activity assays to determine total AMP inhibitory activity.

Agar diffusion assays were conducted to determine the total inhibitory activity of AMP AS precipitate solutions against eight enterococcus strains and four Listeria strains (Table S1). Agar plates containing the indicator cultures were created as described previously. Fresh aliquots of AMP and pNZC precipitate solutions were thawed overnight at 4°C prior to being used. After thawing, a 3-fold dilution series of AMP precipitate solution in pNZC precipitate solution was prepared in a deep 96-well plate, and 5 μl of each dilution was plated on indicator agar plates in triplicate. Agar plates were incubated under stationary conditions for 18 to 24 h at 37°C, after which halo formation was identified and recorded to determine AMP inhibitory activity. To quantify total inhibitory activity, we defined the minimum inhibitory dilution (MID) as the lowest dilution of resuspended AS precipitation solution that inhibited growth. This unitless metric of total activity is the product of an AMP’s ability to be produced by L. lactis and inhibit growth, which is directly meaningful to intended applications with in situ cellular delivery. Following determination of the lowest fraction of AMP solution that resulted in halo formation for all replicates, the MID was statistically calculated, given that the true MID is between the observed MID and the next lowest tested AMP concentration, as described previously for statistically determining MICs (73, 74).

Proteolytic stability assay.

AMP AS precipitate solutions were treated with various concentrations of trypsin, chymotrypsin, and pepsin (T1426, C1429, and P6887, respectively; Sigma-Aldrich) to determine the AMP proteolytic susceptibility to relevant proteases. Trypsin and chymotrypsin dilutions were prepared in 0.1 mM HCl solutions and incubated with AMP AS precipitate solutions at 25°C. Pepsin dilutions were prepared in 3 mM HCl solutions and incubated with AMP AS precipitate solutions at 37°C. AMP AS precipitate solutions were incubated at the specified temperatures for 5 min as 1:1:1 mixtures with the protease solutions at various concentrations and a pH buffer to achieve a final solution pH within the activity range for the specified proteases. Trypsin and chymotrypsin samples were mixed with 1 part of 51 mM NaOH–80 mM glycine at pH 10 to achieve a final mixture pH of ∼7.5. Pepsin samples were mixed with 1 part of 20 mM HCl–100 mM KCl at pH 2 to achieve a final mixture pH of ∼3.5. The samples were heated to 98°C for 20 min to inactivate the proteases, and remaining AMP activity was determined using the agar diffusion assays as described previously using L. monocytogenes ATCC 51775 as the indicator strain. Samples containing no protease were also included as a positive control.

To quantify activity of all tested samples, ImageJ software was used to analyze images of the agar diffusion plates. The average brightness of each halo, excluding the interior point made by the pipette tip, was measured using ImageJ. Brightness values were locally normalized to the mean brightness of four points at the corners of each halo (Fig. S3). The inverse of locally normalized brightness values was then globally fit to the equation below, which is derived from the Michaelis-Menten equation assuming a low substrate concentration (supplemental derivation [see equations S1 and S2 in the supplemental material]), using the fitnlm function on MATLAB. The inverse of normalized brightness values was used as darker halos with lower brightness values correlate with high activity. Global fitting of triplicate data was used to calculate standard error in the parameter estimates.

$$B=B_{\text{min}}+\left(B_{\text{max}}-B_{\text{min}}\right)\text{exp}\left(-\frac{k_{\text{cat}}\prime {\left[E\right]}_0\Delta t}{K_m}\right)$$

In the equation above, B represents brightness, [E]₀ is the total protease enzyme concentration, and t is time. Reported k_cat′/K_m values are normalized to the initial substrate concentration and, thus, include a term [mM substrate]⁻¹.

Bacteriocin IIa AMP and ManPTS sequence analysis and modeling.

manPTS EIIC and EIID sequences were extracted from genomic sequences of strains E. faecium 6E6 and E. faecalis V583, CH116, Com1, and Pan7 strains from NCBI (75). Primers were designed from the sequence of the E. faecium 6E6 manPTS (Table S3) and used to amplify the gene from E. faecium 8E9 using colony PCR. The gene fragment was then Sanger sequenced. Independently, class IIa bacteriocin and manPTS sequences were aligned using the multialign function in MATLAB with default settings. All 210 bacteriocin IIa sequences identified during evaluation of libraries 1 and 2 were included. For sequence modeling, sequences were one-hot encoded, and all positions that were conserved across all observations were eliminated to minimize the size of the final one-hot-encoded matrix.

Sequence-activity data were modeled using the lassoglm function in MATLAB with 5-fold cross-validation and lambda values ranging from 10⁻³ to 10³ and assuming that activity scores were drawn from a binary distribution. Alpha values were set to 0.01 to use ridge regression due to overfitting concerns. Activity data were either the binary result of whether an AMP had activity toward at least one strain (210 data points) or the binary result of whether an AMP had activity to a particular indicator strain (210 AMPs and 6 indicator strains for a total of 1,260 data points). The data sets were randomly split into 10 equal partitions. Models were iteratively trained on nine partitions and evaluated on the 10th until all partitions had been evaluated (10-fold cross validation). Reported root mean square error (RMSE) and positive prediction values are based on the predictions for all 10 evaluation data sets for a given model. Models were trained and evaluated on the following data sets:

• i.The reduced one-hot-encoded AMP sequences fit to their binary activity toward all indicator strains; i.e., if an AMP is active toward at least one indicator strain, it is deemed active (210 data points)
• ii.The reduced one-hot-encoded AMP sequences fit to their binary activity toward each individual indicator strain (6 models, each with 210 data points)
• iii.The reduced one-hot-encoded AMP sequences and a 6-column, strain-indicator matrix fit to their binary activity toward each individual strain (1,260 data points); and
• iv.The reduced one-hot-encoded AMP and manPTS sequences fit to their binary activity toward each individual strain (1,260 data points).

manPTS sequence similarities between species were calculated between the manPTS EIIC and EIID sequences from E. faecium NRRL B2354, E. faecalis V583, and L. monocytogenes ATCC 51775. E. faecium NRRL B2354 and E. faecalis V583 manPTS genes are identical or nearly identical to all other available manPTS genes from strains of the same species used in this work. Given that L. monocytogenes ATCC 51775 is the only Listeria strain with available sequence data for manPTS genes, it was assumed that this is representative of all four strains tested, given the high homology observed in E. faecium and E. faecalis.

Data availability.

All sequences are available in the supplemental material. Plasmids will be provided upon request (Benjamin Hackel [hackel@umn.edu], University of Minnesota).