Discovery of Antimicrobials By Me x : Massively Parallelized Growth Assays

The number of newly approved antimicrobial compounds has been steadily decreasing over the past 50 years emphasizing the need for novel antimicrobial substances. Here we present Me x , a method for the high-throughput discovery of novel antimicrobials, that relies on E.coli self-screening to determine the bioactivity of more than ten thousand naturally occurring peptides. Analysis of thousands of E. coli growth curves using next-generation sequencing enables the identication of more than 1,000 previously unknown antimicrobial peptides. Additionally, by incorporating the kinetics of growth inhibition, a rst indication of the mode of action is obtained, which has implications for the ultimate usefulness of the peptides in question. The most promising peptides of the screen are chemically synthesized and their activity is determined in standardized susceptibility assays. Ten out of 15 investigated peptides eciently eradicate bacteria at a minimal inhibitory concentration in the lower µm or upper nm range. This work represents a step-change in the high-throughput discovery of functionally diverse antimicrobials.


Introduction
Natural compounds are fundamental for drug discovery as they provide the biological relevance and structural diversity required to identify drug-like pharmacophores. [1] Owing to their high structural complexity and their ability to penetrate tissues and membranes, peptides are becoming increasingly important for many therapeutic areas. [2] Especially antimicrobial therapies have a very strong demand for novel compounds due to rising antimicrobial resistance. [3] Although about 3,000 antimicrobial peptides have already been discovered, advances in genome sequencing and mining provide an everincreasing number of peptides with elusive functions. [4,5] Large peptide libraries can be screened for antimicrobial activity using bacteria self-screening protocols. Here, peptides are expressed from their encoding DNA template and then accumulate either in the cytosol, the periplasm or at the bacterial surface. [6] If antimicrobial, their expression negatively impacts the proliferation rate or survival of the expressing cell. Sequencing of the peptide-encoding DNA of such impaired cells allows for the identi cation of antimicrobials in large pools of uncharacterized peptides. However, previous self-screening approaches failed to deliver large fractions of highly active peptides, or were unsuited for the screening of big libraries. [7][8][9] Thus, novel high-throughput screening methods are urgently needed.
We gathered naturally-encoded peptides from peptide and genomic sequence databases and assayed them for antimicrobial activity using massively parallelized growth assays (Me x ). Combined, the method delivered a rich collection of functionally diverse and highly active antimicrobial peptides.

Results
We rst designed a library of naturally-encoded peptides. For this, we collected the amino acid sequences of 3,063 peptides with already experimentally validated activity ("Parents" from here on) from the antimicrobial peptide database (APD) (Figure 1a). [4] Notably, Parents differed considerably with respect to the host from which they were derived, length, physiochemical properties, chemical modi cations, 3D-structure, and sequence (Figure 1a/b). Next, we applied tblastn on the translated nucleotide databases accessible through the NCBI using the amino acid sequence of the Parents as queries. [10] This search yielded 36,898 amino acid sequences with a similarity of ≥ 21.1% to the Parents ("Similars" from here on). Unlike the Parents, only very few of the Similars have been synthesized or experimentally tested. However, owing to their natural origin and similarity to the Parents, a fraction of the Similars is likely to display antimicrobial properties. [11,12] For technical reasons (Methods), we applied a cut-off of 42 amino acids in peptide chain length and selected Similars with at least 62.2% sequence similarity. In this way, a library of 2,122 Parents and 10,300 Similars (Figure 1b) was obtained. Examination of the nal library indicated net charges from 10 to +15 and hydrophobicity of -3.5 to 2.9 (GRAVY scale; Figure S1). Additionally, we were able to allocate the origin of 7,497 of these peptides to the kingdom animalia, 74 to fungi, 678 to bacteria, and 2,485 to plantae ( Figure S2).
For Me x , we converted the peptides into corresponding oligonucleotides ( Figure S3), retrieved the latter as a pool after chemical synthesis on a microarray, and ligated the sequences into a plasmid on which their expression was controlled by the tightly regulated P BAD promotor (Figure 1c). We then transformed the model organism, E. coli TOP10, with the peptide-encoding DNA library. Using next-generation sequencing (NGS), we only identi ed 10,663 different peptideencoding DNA sequences (listed by ID in File S1) in E. coli indicating sequence bias in the initial oligonucleotide pool ( Figure S4). .
To assess the antimicrobial activity of the DNA-encoded peptides, we performed Me x and generated growth curves for each of the 10,663 peptide-expressing E. coli strains. To do so, we inoculated three liquid cultures each with 500 million transformed cells, and induced peptide synthesis after four cell doublings ( Figure S5). Because the synthesis of an antimicrobial peptide should inhibit the growth of the expressing host, the propagation rate of the peptide-encoding DNA will also be reduced. Hence, we harvested bacteria at the time of induction as well as 1.5 h, 3.0 h, and 4.5 h post-induction and used NGS to count reads for each peptide-encoding DNA. To derive growth curves (Figure 1d), we calculated the abundance of each strain (ID) using the respective NGS read counts and multiplied these with the measured cell concentration of the entire liquid cultures (OD) thereby obtaining an approximation of the strain-speci c concentrations (OD ID ) at each sampling point. Comparing OD ID of all peptide-expressing strains after 4.5 h, we found that intracellular expression of 1,240 peptides (11.6%) signi cantly inhibited the growth of their host ("Me x -actives" from here on; Wald's test, p-value (p)<0.05, adjusted for multiple testing (adj.); Figure S6). The remaining peptides did not show growth inhibition in Me x , likely because they are not antimicrobial at all or require chemical modi cations not introduced in E. coli, could not access their (e.g. extracellular) target, or did not reach inhibitory concentrations due to limited mRNA or peptide stability.
Next, we con rmed that the intracellularly synthesized peptides also inhibited growth if the strains were grown individually. For this, we selected 110 peptideexpressing strains experiencing different levels of growth inhibition in Me x and measured their growth in microtiter plate wells ( Figure S7a/b). As the growth curves recorded in Me x and microtiter plates were comparable ( Figure S7c), we concluded that the complex dynamic of the Me x -culture did not bias the results. Screening 10,663 peptides at once allowed us to address several research questions. Firstly, we sought to con rm that our approach of exploiting sequence similarities to known antimicrobial peptides indeed allowed us to identify antimicrobials. In fact, 1,035 out of 1,240 Me x -actives (83%) were Similars, i.e. peptides whose functions were not reported on the APD. A closer look revealed that for 310 inactive Parents we found at least one active Similar. As an example, Parent Apo5 APOC1 667 APD (nomenclature: name of Parent on APD ID Origin ), itself inactive, yielded 27 Similars of which one showed eight amino acid differences to the Parent and displayed antimicrobial activity ( Figure S8). We argue that the amino acids by which the inactive parent and the active similar differed were of high importance for activity and necessary for evading the abovementioned reasons for failed growth inhibition in Me x . Furthermore, 47 Parents spawned an overrepresentation of active Similars (Fisher's exact test, adj. p<0.05; Figure S9). Examples include Myticin-B (21/31), which spawned 31 Similars, of which 21 were active, and PepG1 (11/11). This indicates that the respective peptide sequences have considerable plasticity and can accommodate multiple amino acid exchanges without losing activity. We argue that these peptides might be well suited for additional modi cations performed for instance in the course of lead optimization programs. [13] Secondly, we evaluated the phylogeny of the hosts from which the inhibitory peptides were derived. For this, all peptides of the library were grouped taxonomically based on their natural host. We then calculated the fraction (%) of Me x -actives within the ranks Kingdom and Class (Fisher's exact test; Figure   S10). Me x -actives were signi cantly underrepresented (p<0.05) among bacteria (8.5%), amphibians (7.7%), and mammals (10.3%) but overrepresented (p<0.05) in insects (13.4%), birds (25%), ray-nned shes (15.6%) and bivalves (31.8%). Since insects contain by far the most species in the animal kingdom, this indicates a huge and so far undiscovered pool of antimicrobials in insects.
Thirdly, as cationic and hydrophobic peptides generally display antimicrobial activity, we wondered whether growth inhibition in Me x was biased by the physiochemical properties of peptides. [14] However, linear regression analysis indicated no correlation of growth inhibition with hydrophobicity (correlation = 0.04) and charge (correlation = -0.01; Figure S11a). Furthermore, among the 47 Parents with overrepresented active Similars, there was no clear relationship between charge or hydrophobicity and growth inhibition ( Figure S11b). We thus conclude that growth inhibition in Me x -actives is driven by the speci c antimicrobial activity of a peptide either damaging the cytoplasmic membrane or binding and inhibiting other cellular components.
To investigate peptides further, we characterized the 50 most growth inhibitory peptides as indicated by their rank in the Me x screening (rank 1-50; 38 similars, 12 parents) (Figure S12a-c). Initial tests were performed with two biosensor constructs, containing the cspA and recA promoters, which upon activation are indicative of translation impairment and DNA damage, respectively. [15] The results indicated translational impairment for 11 and DNA damage for 12 peptideexpressing strains (one-sided t-test, adj. p<0.05; Figure 2b; Figure S13), which suggests that these peptides target intracellular macromolecules. In fact, many peptides traverse the membrane(s) of bacteria without permeabilization and kill cells by binding or blocking intracellular macromolecules. [16,17] For example, Metalnikowin IIA 8984 APD , Metalnikowin III 9011 APD , known ribosomal inhibitors, and Pyrrhocoricin 7122 NCBI , whose parent is also a ribosomal inhibitor, caused the strongest indication for translational impairment in our assay. [18] Next, we measured membrane damage by quantifying propidium iodide (PI) uptake. Expression of 11 peptides resulted in membrane damage, with the strongest damages observed for Delta Lysin I similars whose Parent is a membrane pore inducing bacteriocin from Staphylococcus ( Figure S13, Table S1). Interestingly, 11 out of 12 peptides that caused membrane damage signi cantly inhibited growth already after 1.5 h in Me x (Wald's test, p<0.05) (Figure 2a; Figure S14) while for 25 peptides, and especially for those with putative intracellular targets, growth inhibition started only after 4.5 h (Figure 2a; Figure S14). Noteworthy, delay of the growth inhibition onset has been reported to be indicative of peptides that interact with an intracellular target. [19] We concluded that this effect could be observed in Me x and hence reanalyzed all growth curves recorded for the Me x -actives. Growth was signi cantly inhibited after 1.5 h for 806 peptides (65%) suggesting membrane damage (File S1) but only after 4.5 h in the case of the remaining 434 peptides, suggesting interaction with an intracellular target.
Next, we chemically synthesized 15 out of the 20 peptides that were found to be most growth inhibitory in Me x and determined their minimal inhibitory concentrations (MIC), their membrane damaging capabilities, and hemolytic activity when added to cells as a synthesized chemical (Table 1; Figure S12).
Firstly, no MIC was obtained for ve peptides; however, as four of these ve were either Parents or derived from Parents known to be inactive against E. coli (Table S1), we believe that these peptides exerted activity in the cytosol if synthesized intracellularly but could not reach their target (e.g. the cytoplasmic membrane) when added as a chemical substance to the growth medium. Remarkably, 10 of the 15 peptides for which MICs were recorded, very e ciently inhibited the growth of E. coli (MICs: 0.4 -20 µm; mean = 3.7 µm; median = 1 µm), a concentration range that could qualify as a starting point for drug development. [20] We selected the most active Similar, HFIAP-1 4545 NCBI, and measured the activity against other clinically relevant Gram-negative and positive bacteria. Similar HFIAP-1 4545 NCBI inhibited growth of these strains (MICs: 0.4 -5.6 µm; Table S2), which suggests a broad activity spectrum even though Me x screening was performed in another host. These results indicated that even though we screened the peptide library synthesized cytosolically, Me x -active peptides also strongly inhibited growth when added to cells externally and that the employed E. coli lab strain was suited well for the identi cation of peptides also active against other species.
Secondly, to evaluate the degree rely on damaging of membranes damage for the exertion of antimicrobial, we measured the uptake of PI, when adding the peptides 4 log 2 concentration steps above and below their MIC. As complete lysis of both outer and inner membrane resulted in false negatives in PI measurement (see the decrease in PI signal when increasing the concentration of Melittin in Figure S15), we also quanti ed the point at which both membranes were lysed by measuring the release of intracellularly expressed GFP. Only Ascaphin-6 9286 APD , Enterocin RJ-11 3780 APD, Oxyopinine 2b 9690 APD , and HFIAP-1 4545 NCBI , showed strong membrane damage (>25% PI-positive cells below MIC concentration) in a range of the membrane damaging peptide Melittin (Table 1; Figure S15). For the remaining peptides, and especially for those with reported intracellular targets (Fig. 2b) and a delayed growth inhibitory effect in Me x (Fig. 2a), no membrane damage could be detected at all, or occurred considerably above the MIC ( Figure S15). Hence, other mechanisms, such as blocking of protein translation as reported for the parent of Pyrrhocoricin 7122 NCBI , must play a role in bacterial killing. [18] These results con rmed that the previously recorded intracellular characterization of the peptides can be a good indication for the activity of chemically produces peptides discovered in the Me x assay.
Lastly, as hemolysis is a reliable and sensitive indicator for cytotoxicity assessment, we measured the hemolytic activity of all MIC active peptides. [21,22] Toxicity towards human erythrocytes (>5% hemolysis compared to the Triton-X100 control) at the MIC was only observed for the membrane damaging peptide Enterocin RJ-11 3780 , a known hemolytic staphylococcal toxin, and for the control Melittin (Table 1). [23] Additionally, the membrane damaging peptides Oxyopinine 2b 9690 APD , a known hemolytic spider toxin, and HFIAP-1 4545 NCBI displayed hemolytic activity at higher concentrations ( Figure S16; Table 1). [24] All other peptides did not damage erythrocytes at the tested concentration. This suggests that most of the active peptides found in the Me x screen were not cytotoxic thereby further corroborating the potential of the isolated specimen for drug development. Table 1 Summary of antimicrobial activity assays of the 20 most active peptides in Me x . Peptides for which a MIC could be determined are highlighted in green. Int characterization is derived from the experiments summarized in Figure 2b. Mean MIC-values are recorded (n=3) in microtiter plate assays using chemically s peptides against the screening strain E. coli TOP10; p.f. = peptide puri cation failed. Membrane damage is reported at the peptide concentration, relative to t which at least 25% of cells acquired externally added PI. (n=2; Figure S16). Hemolysis as a percentage is related to full lysis after treatment with 2.5% Triton- Figure S16).

Discussion
We applied Me x for the highly parallelized discovery and characterization of more than 10,000 structurally diverse, and naturally-encoded peptides ( Figure   1a/b; Figure S1; File S1). Chemical production and microtiter plate based screening of a library of such high diversity and size, and containing peptides longer than 40 amino acids would have been very cost-and time-consuming. However, Me x takes a shortcut by using in silico optimized and pooled oligonucleotides as templates for ribosomal peptide synthesis ( Figure S3) and intracellular activity assessment via monitoring the growth inhibitory effects (Figure 1c/d).
Growth curves recorded by Me x via NGS were comparable to those obtained for a few tested strains if grown compartmentalized in microtiter plates ( Figure  S7). This indicated that pooling of the peptide-expressing strains did not bias the experimental outcome. Moreover, the high hit rate (10 out of 20 peptides) obtained for Me x -active peptides if synthesized as chemicals and tested in MIC assays under stringent CSLI-assay conditions (Table 1) corroborates the robustness of Me x . In addition, screening of naturally-encoded sequences delivered a large fraction of highly active peptides, by far exceeding the performance of other approaches selecting randomly or semi-randomly designed peptide libraries. [7][8][9] As naturally occurring peptides are preselected for biological activity, including antimicrobial activity, our results con rmed the advantage of screening sets derived from genomic databases. [1,12,25] When extrapolating from our hit-rates (50% of Me x -actives were active in MIC assays using chemically synthesized peptides) to the entire library (1,035 active similars and 205 active parents), we found more than 500 previously unreported, active antimicrobial peptides, derived from organisms of various biological classes ( Figure S6; Figure S10).
Moreover, by analyzing growth curves of the peptide-expressing strains (displayed as OD ID , Figure 2a), observing intracellular stress response mechanisms (Figure 2b), and investigating the membrane damage of a subset of peptides (Table 1, Figure S15), we estimated that about one-third of the active peptides interact with intracellular targets to exert antimicrobial activity (File S1). These results hence cast a fresh look on the eld of antimicrobial peptides as only very few examples (<50) of intracellularly active peptides are known. [26] We hypothesize that nature designed the peptides such that passage of bacterial membranes and binding to macromolecules is a frequently built-in feature. As the transition from the discovery pipeline to the patient is often hampered by the poor speci city of membrane damaging peptides in vivo, Me x can be a valuable tool for the high-throughput discovery of peptides that do not rely on membrane damage for bacterial killing. [27,28] Taken together, Me x enables rapid discovery and classi cation of naturally-occurring and functionally diverse antimicrobial peptides. However, we argue that Me x can also be used for de novo design or optimization of natural peptides by directed evolution approaches and that, the principal technology can eventually also be used for screening in drug-resistant (e.g. Pseudomonas aeruginosa or Acinetobacter baumannii). Ultimately, Me x will hence allow paving the way towards the discovery of next-generation antibiotics for human applications.

Chemicals and reagents
Unless otherwise stated, all chemicals, reagents, and primers were obtained from Sigma Aldrich (Buchs, CH). Restriction enzymes and their buffers were obtained from New England Biolabs (Ipswich, USA). Synthetic genes were obtained from Integrated DNA Technologies (Leuven, BE) or Twist Bioscience (San In silico generation of peptide library We collected all peptide sequences (called "parents") available on the APD in May 2017 (https://aps.unmc.edu/). [4] These sequences were used as input queries to nd sequence-similar peptide sequences in the NCBI non-redundant nucleotide collection (nr/nt), a collection that holds sequences from GenBank, European Molecular Biology Laboratory (EMBL), DNA Databank of Japan (DDBJ), and Reference Sequence database (RefSeq), as well as translated protein information from the protein database (PDB). [10] By applying tblastn, 170,300 additional peptide sequences (called similars) were found. [29] Because we were limited to 12,412 different peptides with a maximum length of 42 amino acids (the chosen platform for the synthesis of the peptide-encoding oligonucleotides allowed 12'412 different sequences with a maximal length of 170 bases), we discarded similars with sequence similarity to the respective parent of less than 62.2%. The following parameters were used for the tblastn search: maximum sequences = 100; matrix = BLOSUM62; gap cost = 11.1; word size = 6; active low complexity lter; adjustment = conditional compositional score matrix adjustment.

Sequence distance among parents and similar
To visualize sequence diversity among parents, we created a sequence-based phylogenetic tree. We performed pairwise global alignment of all parent sequences using the Needleman-Wunsch algorithm, as implemented in the R Bioconductor package 'Biostrings' (https://bioconductor.org/packages/release/bioc/html/Biostrings.html). The BLOSUM62 substitution matrix was used to compute the alignment scores, which were converted into pairwise distances following the method Scoredist. [30] Based on the pairwise distances between parents, we used hierarchical clustering with average linkage to compute a dendrogram of sequences re ecting their similarities. parents and their tblastn-derived similars were consolidated into groups, which were named after the parent from the APD (https://aps.unmc.edu/). In the sequence-based phylogenetic tree, each similar was stacked on top of its parent at the tip of the dendrogram. A similar may appear multiple times if it was found multiple times in the tblastn search using different parents.
Peptide-encoding DNA architecture The corresponding oligonucleotide sequences of the peptide library were synthesized using microarray technology supplied from CustomArray Inc. (now GeneString, Piscataway, USA). The chosen platform allowed 12'412 different oligonucleotides with a maximal length of 170 bases. A generic oligonucleotide design employing four functional units was created ( Figure S3): A coding unit, a ller unit, and two universal units for ampli cation. This process was automated for each sequence by using an in-house written script in R. The coding unit contained the reverse translation of the peptide amino acid sequence into a codon-optimized DNA for E. coli. We always chose the most abundant codon for each amino acid. In cases in which restriction sites had been introduced that could potentially interfere with subsequent manipulations, the crucial codon was replaced by the second most abundant one for this amino acid. The ller sequence was added to compensate for the various lengths of peptide genes (shortest coding sequence = 15 nucleotides, longest coding sequence = 126 nucleotides) and adjust the total of ller and coding unit to 129 nucleotides for all members of the library. To do so, we rst added a UAA stop codon to the end of the coding sequence and then added downstream a semi-random sequence, ensuring a GC content of 40% for the ller sequence and limiting the number of identical nucleotides following each other to three. By adding this ller sequence we maximized sequence disparity at the DNA level (many coding sequences are homologs) thereby potentially increasing both synthesis and, later, sequencing quality. Two ampli cation units, of 23 and 18 bases, respectively, were appended upstream and downstream of the coding sequence and ller unit and contained the ribosomal binding site and restriction sites for the enzymes PstI and HindIII. Two amplify the peptide-encoding DNA, primer 1: CTGCACAAAGCTTACGTG, complementary to the upstream ampli cation unit, and primer 2: CACGTAAGCTTTGTGCAG, reverse complementary to the downstream ampli cation unit were used. The nal 170 bases long oligonucleotide sequences as synthesized are listed by ID in File S2 (erroneous sequences were discarded).

Peptide-encoding DNA cloning
The chemically synthesized and single-stranded oligonucleotides were separated from their array and we received them as a pool. This pool was aliquoted in 10 mM Tris-Cl, 1 mM EDTA, pH 8 and deep-frozen at 80°C. The pool was ampli ed by polymerase chain reaction (PCR) in a 50 µl reaction using 5 ng of the template and 10 µm HPLC-puri ed primer 1 and primer 2, complementary to the ampli cation sites, and 25 µl of Phusion® High-Fidelity PCR Master Mix with HF buffer. The ampli cation was performed using 25 cycles of 98°C for 15 s, 55°C for 20 s, and 72°C for 5 s. The now double-stranded peptide-encoding DNA sequences were puri ed using a DNA puri cation kit. DNA concentration was measured using a NanoDrop 2000 Spectrophotometer (Thermo Fisher Scienti c) and 500 ng of the puri ed product was used for a restriction digest using enzymes HindIII-HF and PstI-HF in Cutsmart buffer. The digested product was again puri ed using a DNA puri cation kit and ligated to plasmid pBAD (Thermo Fisher Scienti c) digested with the same enzymes. [31] This plasmid harbored the tightly controllable P BAD promoter for peptide gene expression, a pBR322 replication of origin, and a resistance gene encoding for beta-lactamase. For ligation, pBAD was puri ed using a 1% agarose gel and a DNA gel recovery kit after digestion. Next, T4 ligase (800 units) was used to ligate 100 ng of cut pBAD vector and 10 ng peptide-encoding DNA sequences in T4 ligase buffer (molar ratio of 7:1 insert:vector). The ligation mix was incubated for 14 h at 16°C. The ligation product was dialyzed in deionized water using lters (MilliporeSigma, Burlington, USA) and 1 µl of the mix was used to transform 20 µl of CloneCatcher™ Gold DH5G Electrocompetent E. coli (Genlantis, Burlington, USA) cells using electroporation. Recovered cells were plated and incubated overnight on LB agar plates supplemented with carbenicillin. Afterward, ~500,000 colonies were washed off the plates using LB medium, and the plasmids containing the peptideencoding DNA sequences were extracted from 2.5*10 9 cells using a plasmid isolation kit. An aliquot of 5 ng of these plasmids was used to transform E. coli TOP10 cells using the protocol from the transformation above. A total of 1'000'000 colonies were recovered from the plates after overnight incubation by washing with LB medium, the suspension was diluted to OD = 1 with LB-medium, glycerol was added to a nal concentration of 20% (v/v), and aliquots of 500 million cells were stored at 80°C.

Growth experiment
Three aliquots of E. coli TOP10 harboring the peptide-encoding DNA sequences on the pBAD plasmid were thawed and added to three 1 l ba ed shake asks containing 100 ml of LB medium + 100 µg ml −1 carbenicillin. The cultures were grown for roughly 7.5 h at 37°C. When the OD reached 0.2, the cultures were supplemented with l-arabinose to a nal concentration of 0.3% (w/v) to induce peptide expression. Cell samples were taken from each biological replicate at the point of induction and 1.5 h, 3 h, and 4.5 h post-induction. The plasmids were extracted from all samples using a plasmid isolation kit.

NGS
For the generation of Me x growth curves, peptide-encoding DNA sequences on plasmids, collected from the three replicates across four time points during the growth experiment, were sequenced by NGS. Additionally, the abundance of peptide-encoding DNA sequences in the original oligonucleotide pool and after transformation of the assay strain E. coli TOP10 was assessed by NGS as well. Peptide-encoding DNA sequences were ampli ed by primer 1 and primer 2 using 100 ng of plasmid and the PCR-ampli cation protocol mentioned before, but only for 10 cycles to avoid ampli cation bias. The ampli cation product was puri ed using an agarose gel. in Me x were listed with a unique identi er (ID) in File S2.

Generation of M e x growth curves
We used the standard work ow of DESeq2 (NGS read count normalization, dispersion estimates, and Wald's tests) to analyze NGS read counts. [32] Only sequences that passed independent ltering were included in further analyses (= 10,633). To draw growth curves for each peptide-expressing strain, we calculated the log2-fold changes of NGS read counts (listed for each ID in File S2) between the time of induction and all other time points (1.5 h, 3.0 h, and 4.5 h post-induction). A Bayesian shrinkage estimator was employed to shrink the log2 fold-change for each ID (lfcShrinks ID ) between all time points using the R/Bioconductor package 'apeglm'. [33] To draw the Me x growth curves, we calculated a strain-speci c OD ID at each time point according to equation (1). OD values at the speci c time points were averaged values from all three biological replicates ( Figure S5). The OD ID (0 h) for each peptide-expressing strain was set to 0.2 at the time of induction as lfcShrink ID (0 h) = 0 and OD = 0.2. This enabled us to compare peptide-expressing strains of different abundancies (see Figure S6). OD ID values can be interpreted as the OD values that would have been measured when incubating the respective strain individually in the same experiment, i.e. in this case in LB medium in a 100 ml shake asks.
To nd Me x -active peptides, we also performed a one-sided Wald's test, with the alternative hypothesis that the expression of a given peptide leads to a reduced OD ID 1.5 h and 4.5 h post-induction. We rejected the null hypothesis at signi cance level alpha = 0.05. Peptides with a p<0.05 (after adjustment for multiple testing using the Benjamini-Hochberg method) after 4.5 h are considered Me x -active peptides. Peptides with p<0.05 after 1.5 h do signi cantly inhibit growth already after 1.5 h. All values and results are reported in File S1.

Monoseptic growth experiments
Taking the OD ID (4.5 h) of each peptide-expressing strain, we could rank all peptides by their growth inhibitory effect. We selected 110 peptides (Ranks 1-50, 100-119, 1000-1019, and 10,000-10,019) and then generated an identical copy of the strain previously used in Me x for its expression. First, the corresponding peptide-encoding DNA-sequences were synthesized as gene fragments. An aliquot of 400 ng of each gene fragment was directly used for a restriction digest using enzymes HindIII-HF and Pst-HF in Cutsmart buffer. The product was puri ed using a DNA puri cation kit. Next, T4 ligase (800 units) was used to ligate 50 ng of identically digested pBAD vector and 10 ng of puri ed gene fragment in T4 ligase buffer for 14 h at 16°C. The ligation product was puri ed using a DNA puri cation kit. An aliquot of 5 µl of the puri ed ligation product was then used to transform chemically competent E. coli TOP10 cells. From the resulting colonies, we isolated one strain, sequence-veri ed the correct assembly of the expression plasmid, and stored it after overnight growth in glycerol at -80°C. For the growth experiment, we rst re-isolated single colonies on solid media and then picked three clones, incubated them separately overnight, and inoculated them into 200 µl fresh LB medium containing 0.3% (w/v) l-arabinose to a nal OD of 0.01 into 96-well microtiter plates (Greiner). Growth was recorded by measuring OD in a Tecan In nite 200 PRO (Tecan, Männedorf, CH) for 4.5 h (37°C, 1.5 mm orbital shaking).

Enrichment analyses
We used Fisher's exact test to assess the over-or underrepresentation of Me x -actives in various groups. This amounts to a hypergeometric test to assess the signi cance of drawing n active peptides in a group of k, from a population of size N containing K active peptides. We rejected the null hypothesis at signi cance level alpha = 0.05. Groups with a p<0.05 had a signi cantly different representation of active peptides compared with the overall population. When adjusting for multiple testing, we used the Benjamini-Hochberg method.

Peptide classi cations
The physicochemical parameters of the peptides were calculated at pH 7 using the R package 'Peptides' (https://cran.r-project.org/package=Peptides). For charge, we used the method by Lehninger. [34] For hydrophobicity, we used the calculations by KyteDoolittle. [35] The information for each parent such as the name, chemical modi cation, activity, 3D-structure, was extracted from the APD website (https://aps.unmc.edu/) using an in-house R script. The information on the species from which a speci c peptide sequence originated, was extracted from the tblastn search and the APD website. The entire taxonomic classi cations (kingdom, phylum, class) for each species were extracted, if available, from the Global Biodiversity Information Facility Data Portal (https://gbif.org) using the R package 'taxize' (https://cran.r-project.org/package=taxize). The results are summarized in File S1.
Membrane damage assay assay using intracellularly synthesized peptides We selected the peptide-expressing strains of rank 1-50 in Me x that we had previously constructed for the monoseptic growth assay. Additionally, we selected the strain expressing the inactive control peptide HNP-1 3425 APD , a peptide known to be inactive if expressed in E.coli. [8] Each strain was re-isolated on solid media from frozen stock and incubated overnight. Then, two colonies were picked and incubated overnight in 96-deep-well polypropylene plates. These cultures were used to inoculate fresh media containing 0.3% (w/v) l-arabinose to a nal OD of 0.01 into 96-well microtiter plates. The plates were then incubated on for 4.5 h (37°C, 1.5 mm orbital shaking). After 4.5 h, an aliquot of 50 µl of cell suspension a Tecan In nite 200 PRO plate reader was added to 150 µl of phosphate-buffered saline into a fresh 96-well microtiter plate. Propidium iodide (PI) was added to a nal concentration of 1 µg ml −1 . PI is a DNAintercalating dye that cannot pass an intact cytoplasmic membrane. [36] For each sample, PI uorescence (λ Ex = 579 nm / λ Em = 616 nm) of ~10,000 cells were analyzed using a ow cytometer LSR Fortessa (BD Biosciences, Allschwil, CH). To determine the membrane damaging properties of each of the expressed peptides, we calculated the fraction of cells in percent for which a PI uptake was measured using the software FlowJo V10 (BD Biosciences).
Stress response assay using intracellularly synthesized peptides We selected peptide-expressing strains of rank 1-50, previously generated for the monoseptic growth assay. Additionally, we selected the strain expressing the inactive control peptide HNP1 3425 APD . Moreover, two plasmids (cloning vector: puA66) containing either the promoter of the gene for recombinase A (P recA ) or for the gene for cold shock protein A (P cspA ) were puri ed from the E. coli Alon collection. [15] Both plasmids contained a transcriptional fusion of their promoter with a downstream gene for green uorescent protein (gfp), an additional kanamycin resistance cassette, and a pSC101 origin of replication. We transformed each of the 51 peptide-expressing E. coli strains with each of the two plasmids to generate 102 different strains and incubated them overnight on solid media. Then, three colonies were picked and incubated overnight. These cultures were used to inoculate fresh media containing 0. % (w/v) l-arabinose to a nal OD of 0.05 into 96-well microtiter plates. We recorded OD and GFP expression (λ Ex 488 nm/λ Em 530nm) after 1.5 h and 4.5 h using a Tecan In nite 200 PRO (37°C, 1.5 mm orbital shaking). For each strain, we calculated the speci c uorescence change between the two time points (GFP/OD (4.5 h) -GFP/OD (1.5 h)). Statistical signi cance was calculated by one-sided t-tests, adjusted for multiple testing by Benjamini-Hochberg, using the signal of HNP-1 3425 APD as null distribution. We rejected the null hypothesis at signi cance level alpha = 0.05.
Puri cation of chemically synthesized peptides Peptides were obtained from Pepscan (Lelystad, NL) in >90% purity or in crude format and subsequently puri ed to >90% purity in-house. For the latter, crude peptides were dissolved in 5 ml DMSO and 15 ml 0.1% aqueous tri uoroacetic acid, TFA. HPLC-puri cation of the dissolved crude peptides was performed on an ӒKTAexplorer chromatography system (GE Heathcare, SE). The entire peptide sample was loaded onto a RP C18 column (PRONTOSIL 120 C18 10 µm, 250 x 20 mm, 50 x 20 mm precolumn, Bischoff, Leonberg, DE), heated to 30°C and operated at a ow rate of 10 ml min 1 using 0.1% aqueous TFA as solvent A and acetonitrile supplemented with 0.1% TFA as solvent B. The ratios of A to B were adapted for each peptide and typical values are given below. The column was equilibrated with the peptide-speci c mixture of solvent A and solvent B 0-20%) prior to injection. After injection and an initial wash step of 6 min a gradient was imposed with the same mixture, and then a gradient was applied, in the course of which the amount of solvent B was increased to 50-0 % in 40 min. The column was washed with 5 % solvent B for 8 min and equilibrated with the speci c solvent A/solvent B mixture for the next run for 13 min. Peptide elution was monitored spectrophotometrically at 205 nm, and generally the main peptide peak was collected. . Separation was performed using the same concentration pro le previously used for puri cation. The concentration was measured using the integrated peak area at 205 nm and then calculated using peptide-speci c absorption properties. [37,38] Measurement of the MIC using chemically synthesized peptides On the same day at which MIC assays were executed, puri ed peptides were thawed and the concentration was determined by HPLC as described before. E.coli TOP10 cells were grown in Mueller Hinton Broth (MHB) or diluted MHB (25% of the original strength) overnight to stationary phase. Diluted MHB has been frequently used to assay antimicrobial peptides. [39] The cultures were then supplemented with 20% glycerol, aliquoted, and frozen at 80°C. For MIC measurements, a frozen stock of the cells was thawed, resuspended in MHB or 25% MHB to adjust to a density of 5*10 5 cells ml − 1 in the experiment, and distributed to microtiter plate wells by an automated liquid handling system (Hamilton, Bonaduz, CH). Then the peptides were added by the liquid handling system in 2-fold dilutions using minimum of 100 µg ml − 1 as the highest concentration. MICs were determined as broth microdilution assay in 384-well at bottom polypropylene plates (Falcon® 96-Well Flat-Bottom Microplate) adapted from the protocol of Wiegand et al. [40] The plates were sealed airtight and incubated for 18 h without shaking at 37°C before reading the OD using a Tecan In nite 200 PRO plate reader. The MIC value corresponded to the concentration at which no growth of the bacterial strain was observed (< 5% of the OD value of the growth control). MIC experiments were performed at least in triplicate.
Membrane damage assay using chemically synthesized peptides To measure extracellular membrane damage, E. coli ATCC 25922 [pSEVA271-GFP] and the peptide dilutions were prepared as described for the MIC measurements but covering a concentration range of 16 x MIC to MIC/16 in 2-fold dilutions steps with a nal assay volume of 200 µl. The bacterial strain suspension was furthermore supplemented with a nal concentration of 1 µg ml −1 propidium iodidejust before pipetting the assay. After 1 h incubation at room temperature membrane damage (=release of intracellularly expressed GFP and/or uptake of extracellularly added PI) was assessed by ow cytometry using a Fortessa Analyzer (BD Biosciences; 488 nm laser with 530/30 nm bandpass lter and 579 nm laser with 610/20 nm bandpass lter). The fractions of PI-positive and GFP-positive were determined with the same gate for all populations using the FlowJo V10 software (BD Biosciences). The extracellular membrane integrity assay was performed in biological duplicates analyzing at least 10,000 cells in each experiment.
Hemolysis assay using chemically synthesized peptides Two samples of human blood were obtained from a blood bank (Blutspendezentrum SRK at the University Hospital Basel). The samples were pooled and erythrocytes were isolated by repeated centrifugation at 500 x g for 10 min, removal of the blood plasma and resuspending the remaining cells in an equal volume of DPBS. Following last resuspension, erythrocytes were diluted 1:50 in DPBS. For the hemolysis assay, a log2 serial dilution of each peptide was prepared as described for the MIC but using DPBS and a 96well plate (Ubottom, PP, 650201, Greiner) with a nal volume of 200 µl. As lysis control, 2.5% TritonX100 in DPBS was used in well 10, well 11 served as non-treated control (no peptide added), and well 12 as blank. To each well of the dilution plate, 100 µl of the red blood cells suspension was added. The plate was incubated for 1 h at 37°C. After the incubation, the plate was centrifuged at 500 x g for 10 min and 100 µl of the supernatant was transferred to a clean 96-well plate (F-bottom, PS, 655101, Greiner). The absorbance was measured at 540 nm using an In nite M1000 PRO plate reader (Tecan) and corrected by the measurements from the untreated wells. The lysis of each peptide concentration was expressed relative to the lysis control (set as 100% lysis). The hemolysis assay was performed in triplicate.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.