Rational tuning of a membrane-perforating antimicrobial peptide to selectively target membranes of different lipid composition

The use of designed antimicrobial peptides as drugs has been impeded by the absence of simple sequence-structure-function relationships and design rules. The likely cause is that many of these peptides permeabilize membranes via highly disordered, heterogeneous mechanisms, forming aggregates without well-defined tertiary or secondary structure. We demonstrate that the combination of high-throughput library screening with atomistic computer simulations can successfully address this challenge by tuning a previously developed general pore forming peptide into a selective pore former for different lipid types. A library of 2,916 peptides was designed based on the LDKA template. The library peptides were synthesized and screened using a high-throughput orthogonal vesicle leakage assay. Dyes of different sizes were entrapped inside vesicles with varying lipid composition to simultaneously screen for both pore size and affinity for negatively charged and neutral lipid membranes. From this screen, nine different LDKA variants that have unique activity were selected, sequenced, synthesized, and characterized. Despite the minor sequence changes, each of these peptides has unique functional properties, forming either small or large pores and being selective for either neutral or anionic lipid bilayers. Long-scale, unbiased atomistic molecular dynamics (MD) simulations directly reveal that rather than rigid, well-defined pores, these peptides can form a large repertoire of functional dynamic and heterogeneous aggregates, strongly affected by single mutations. Predicting the propensity to aggregate and assemble in a given environment from sequence alone holds the key to functional prediction of membrane permeabilization.


Introduction
Our results suggest that in-vitro activity, lipid selectivity, and aggregation propensities of 87 AMPs depend highly on even the most conservative sequence changes. While the broad 88 underlying properties correlate with simple descriptors that can be directly derived from the 89 peptide sequence (e.g. hydrophobic moment, overall charge, and amphiphilicity), these 90 quantities do not allow us to directly determine which sequence will be selective, or porate 91 membranes at all. The peptides form a large repertoire of functional dynamic and heterogeneous 92 structures in the membrane, and each sequence change can dramatically affect the 93 oligomerization propensity, structure of the aggregates, ability to porate, and selectivity for 94 different membrane compositions so desired for pharmaceutical application. This suggests that 95 ultimately only structure (rather than sequence) based approaches, such as direct pore 96 aggregation and equilibrium simulations, will enable predictive, rather than descriptive de novo 97 AMP design. 98 99

100
Rational peptide design. LDKA is a small pore-former in neutral POPC and anionic POPG vesicles 101 and has low micromolar antimicrobial activity against bacteria. The goal of this library is to 102 explore whether simple rearrangements of the LDKA sequence using four amino acids (Leu, Asp, 103 Lys, and Ala), will allow modulation of pore-forming potential, pore-size, and targeting of specific 104 membrane compositions. To achieve this, we have designed a combinatorial peptide library 105 containing 2,916 LDKA analogues (Figure 1a-b). The LDKA template sequence was mutated in 106 order to: (i.) adjust peptide hydrophobicity, (ii.) promote more salt bridge formation between the 107 peptides, (iii.) introduce a central proline kink to the structure, and (iv.) substitute more 108 positively charged residues on the C-terminus, which is one of the common motifs in the 109 Antimicrobial Peptide Database (APD; http://aps.unmc.edu/AP) 49 . 110 Peptide hydrophobicity is modulated by interchanging leucine and alanine residues as 111 well as substituting more positive (lysine) and negative charges (aspartic acid) in the sequence. 112 The goal of these mutations is to fine-tune the peptide solubility and membrane-partitioning. To 113 further allow for more structural plasticity of the peptide, we introduced a proline near the 114 center of the peptide sequence, which is common in naturally occurring AMPs 50,51 . More charged 115 residues (aspartic acid and lysine) were introduced to both facilitate inter-peptide salt-bridge 116 formation and strengthen the peptide-peptide interface 52 , as well as to allow for a more polar 117 central pore enabling larger multimeric channel structures 24,53 . Additional positive charges were 118 introduced at the C-terminus to enhance peptide binding to anionic lipids, which is a common 119 motif in many antimicrobial peptides from natural sources, such as Hylaseptin-P1 22,54 , Hylain 2 55 , 120 melittin 56,57 , and maculatin 50,51 . 121 122 Membrane specific poration and pore size. The potency and membrane selectivity of the 2,916 123 LDKA library peptides for zwitterionic (POPC) and anionic (POPG) large unilamellar vesicles (LUVs) 124 was evaluated using a high-throughput liposome leakage screen. This approach allows us to 125 detect and quantify the release of small fluorescent dye ANTS (8-aminonaphthalene-1,3,6-126 trisulfonic acid disodium salt; MW = 427 Da) and its fluorescent quencher DPX (p-xylene-bis-127 pyridinium bromide; MW = 422 Da) encapsulated in LUVs after addition of the library peptides. 128 Neutral POPC LUVs serve as a simple model for mammalian membranes, while charged POPG 129 LUVs serve as a very simplistic model for bacterial membranes enriched in anionic lipids. 130 charged LUVs after addition of the library peptides. In this study, 11.2% of the LDKA analogues 132 have POPG-favourable selectivity and induce >50% encapsulated dye leakage from charged POPG 133 LUVs at low peptide concentration (P:L = 1:1000), while 0.4% cause leakage from neutral POPC 134 LUVs only, and 6.6% disrupt both POPC and POPG LUVs. LDKA analogues that induce >90% dye 135 leakage from POPC and POPG LUVs were screened for their ability to induce leakage of a larger 3-136 kDa TAMRA-biotin-dextran (TBD) dye 36 . Several LDKA-like peptides form larger pores in POPG 137 vesicles, while the pores induced in POPC vesicles are generally smaller (Figure 1d-e).

138
Eight LDKA peptides with different lipid selectivity and pore sizes were selected from the 139 high-throughput screen and sequenced using Edman degradation 58 . Table 1 shows that these 140 peptides have 1 to 4 mutations compared to the LDKA template sequence. The most common 141 mutation is leucine to alanine, occurring 13 times and in a total of 7 of the 8 peptides. Alanine to 142 leucine occurred 6 times in 5 peptides, leucine to aspartic acid occurred 3 times in 3 peptides, 143 and leucine to proline occurred once. 144 The analysis of selected peptide sequences showed positive-charged lysine is not a 145 favourable substitution in the non-polar face of the LDKA template helix and the C-terminal motif 146 (positions 6, 8, 10, 12, and 13). Instead, hydrophobic leucine and alanine are more preferable.

147
This is in agreement with the evolutionary derivatives of 26-residue melittin, in that the 148 positively-charged amino acids (lysine and arginine) are less likely to be favored in the non-polar 149 face 26,37 . 150 Other than fixed lysine at positions 7 and 11, no additional lysine residues were observed 151 in the analogues. Additional aspartic acids were observed at position 3 and 5, which is right next 152 to the fixed aspartic acid at position 4 that can further promote salt bridge formation in peptide-153 peptide interactions. The net charges of these analogues are between +1 and +2, and they are 154 consistent with the majority (net charge +1) of AMPs in the APD 23,49 . This shows that cationic 155 residues can promote peptide binding to anionic bacterial membranes; however, more cationic 156 charges may result in lower hydrophobicity and higher energy barriers to cross the hydrophobic 157 core of membranes. Therefore, a longer peptide length is needed to strengthen the 158 hydrophobicity when the sequence contains more charges. A natural membrane-active peptide, 159 melittin (sequence: GIGAVLKVLTTGLPALISWIKRKRQQ-Amide), is a good example. Although it has 160 four positive charges (-KRKR-) in its C-terminus, longer peptide length (26 amino acids) and the 161 hydrophobic N-terminus (GIGAVLKVL-) make it hydrophobic enough to span cell membranes. 162 Table 1 reveals that leucine to alanine mutations are generally sufficient to prevent 163 poration in neutral POPC membranes, while the peptides still porate charged POPG membranes, 164 which is similar to the L16G mutation of melittin 37 . More specifically, the LDKA analogues that 165 only induce ANTS/DPX leakage from anionic POPG LUVs have 4-5 leucines, while the analogues 166 that can porate both POPC and POPG LUVs have 6-7 leucines in their sequences. The net charge 167 of all LDKA wildtype and analogues are between +1 and +2, and we did not observe any anionic 168 peptide, neutral peptide, or peptide that has net charge greater than +2. This suggests that the 169 membrane-selectivity is driven by hydrophobic moment to POPC but including electrostatics on 170 POPG.

172
Binding to mixed membranes. To investigate the root cause of the different leakage preferences 173 of LDKA analogues for POPC and POPG membranes, we studied the binding and secondary 174 structural properties of LDKA analogues using tryptophan fluorescence and circular dichroism 175 (CD) spectroscopy, respectively. Peptide solutions (50 µM peptide concentration) were titrated 176 with POPC and POPG LUVs (between 0-5 mM) and the corresponding changes of tryptophan 177 fluorescent spectra were collected, yielding binding free energies and helicities of the peptides, 178 albeit without any structural information on the underlying poration process ( Figure S1, 179 supplement). 180 Further studies were performed to answer why some peptides (i.e. 7D12, 7G6, 28H6, 181 11D12, and 24F1) show selectivity for either membrane type. First, we characterised peptide 182 secondary structural changes and binding to LUVs containing binary mixtures of POPC and POPG 183 lipids. Figure 2 (7D12, 7G6, and 28H6) and Figure S5 (11D12 and 24F1) show changes in the 184 tryptophan fluorescence and CD spectra for these peptides upon addition of LUVs for whom the 185 ratio of POPG was elevated from 0 to 100% with 20% increments (0, 20, 40, 60, 80, and 100% 186 POPG). These analogues are sensitive to the anionic POPG lipid and have significant structural 187 change with small PG fraction (20% POPG), except 7D12 which is less sensitive to anionic lipid. 188 These membrane selective peptides only bind to POPG and show little or no binding to POPC, 189 which is consistent to our liposome leakage assay.

191
Simulations of two similar sequences.

192
The information gained from the experimental screen is limited in that there is an absence of a 193 nuanced correlation between simple peptide descriptors and selectivity and leakage propensity.

194
Since a multitude of AMP structures can cause membrane permeabilization, 14 it is critical to 195 identify which overall mechanism applies for the chosen library template. Both peptides have a net charge +1 and have the same C-terminal motif (-KLAGW-Amide).

210
The only differences are (i.) aspartic acid shifts from position 3 in 25B2 to position 5 in 7D12, and 211 (ii.) hydrophobic position 10 where 25B2 is leucine and 7D12 is alanine. A quick analysis shows 212 these simple modifications result in a hydrophobic dipole moment of 4.8 in 25B2 and 1.9 in 7D12.

213
Mirroring the biophysical experiments, we performed peptide-assembly simulations of 7D12 and 214 25B2 in both POPC and POPG bilayers. (Figure 3 and Table S3). Simulations and experiments 215 show that 25B2 results in higher helical content than 7D12 (Table S2 and Table S3). Similar to our 216 prior simulations of LDKA, the peptides spontaneously insert and form a large number of 217 heterogeneous oligomeric pore-structures. These can range from 3-9 peptides, with a core of 218 mainly 4-5 tilted TM inserted peptides, supported by several surface-bound peptides that are 219 more loosely attached. Since the sequences are short, the main arrangements are strongly tilted 220 and double-stacked, rather than a membrane-spanning barrel-stave layout. Peptides align both 221 parallel and anti-parallel at various levels of insertion. The large number of charged sidechains, 222 both cationic and anionic, enable small water-filled bilayer channels with many cross-peptide 223 salt-bridges, pulling in both lipid headgroups and ions. Peptides usually leave and join these small 224 aggregates, resulting in no overall stable structures but rather in a wide variety of different pore 225 assemblies. There is substantial water and ion flux across these, with higher oligomers yielding 226 larger flux. Both cations and anions can translocate across the pores, with a preference for 227 cations in the POPG simulations, presumably due to the more anionic environment of the pore 228 aggregates, where PG headgroups are pulled into the membrane. The heterogenous nature of 229 the pore aggregates indicates a highly dynamic equilibrium which is strongly influenced by 230 individual sequence changes. 7D12 is shown to be selective: It does not insert and form 231 aggregates in POPC, but remains on the surface, indicating that pore aggregates are not stable in 232 this membrane, and the surface state is preferred.

243
The results of MD simulations and ITC are consistent. 7D12 in POPG, and 25B2 in both 244 membrane types assembled channel-like architectures in MD simulations and showed significant 245 heat release in ITC. In contrast, 7D12 in POPC bilayers neither formed any structure, nor induced 246 any heat release/absorption. Thus, there is a remarkable agreement between experiments and 247 simulations. The lower hydrophobic moment of 7D12 appears to explain the less thermostable 248 helical structures than other peptides ( Figure S2), and the unfolded structures are more 249 disordered than the helical structure of 25B2 as compared to what we observed in ITC (Figure 3c-250 d). Therefore, it suggests hydrophobic moment is a determinant to promote membrane 251 selectivity.

253
Hemolysis and antibacterial activity. To test toxicity of the LDKA analogues against human cells, 254 we performed a hemolysis assay (Figure 4a). LDKA wildtype is hemolytic at moderate micromolar 255 concentrations with a hemolytic concentration lysing 50% of red blood cells (HC 50 ) of 55.1 µM 256 ( Table 2). The peptide-induced POPC LUV leakage is correlated with the hemolytic activity ( Figure  257 4b-c). The peptides that induce leakage from POPC LUV at low peptide concentration (P:L = 258 1:1000) are hemolytic (HC 50 = 1-57 µM). More specifically, 7F3 (HC 50 = 1.1 µM) and 28H6 (HC 50 = 259 1.2 µM) are as powerful as natural toxin melittin and its gain-of-function derivative MelP5 (HC 50 = 260 1-3 µM) 37 . All POPG-favourable peptides have no effect to human red blood cell, even at 75 µM 261 peptide concentration. 262 The real test is how selectively the selected peptides target and kill various bacteria. to 16-fold resistance to these antibiotics compared to their 1 st generation strain (Figure 5a).

284
LDKA analogues were tested against these four drug-resistant E. coli cultures. Membrane Activity against biofilms. In clinical settings, bacteria are mostly found in biofilms that are the 293 key drivers of infections 67,68 . We therefore challenged our LDKA analogues against bacterial 294 biofilms, which are generally much more resistant than planktonic equivalents 69 . The results 295 showed that the selected LDKA analogues (4H9, 7F3, 25B2, 7G6, 11D12, and 24F1) can eliminate 296 ~50% of the E. coli biofilm in the presence of 67-150 µM peptide. Only 7F3 is capable of reducing 297 S. aureus biofilms by ~50% with 100 µM peptide concentration, and none of the analogues work 298 against P. aeruginosa biofilms (Figure 5c-e). 299 300

301
In this study, we used the leucine-rich membrane-active peptide LDKA as a sequence template in 302 order to test whether the combination of database-guided combinatorial peptide library 303 screening, and direct MD simulation of membrane aggregation, can tune the template to 304 significantly change its secondary structure, potency, and membrane specificity. Similar rational 305 combinatorial design has been used before to develop and tune the activity of other 306 AMPs 26,37,50,70 . The LDKA library peptides were designed using only four different amino acids 307 (Asp, Lys, Leu, and Ala) in a template sequence of GxxDxxKxxxKxxGW-Amide, where 'x' 308 represents one of the four amino acids. Our LDKA analogues reveal that a small number of 309 conservative substitutions (Leu to Ala) in the LDKA sequence can dramatically change the 310 selectivity toward different membrane types (anionic and neutral LUVs), resulting in specificity to 311 different bacteria species and human red blood cells. This is consistent with Krauson et al., who 312 showed a single-residue change of a leucine at position 16 to glycine (L16G) can redirect the 313 general toxicity of melittin towards bacteria only, leaving red blood cells unharmed 37 . A similar 314 study introduced charged amino acids in the C-terminal motif of HYL-20 peptide, fine-tuning the 315 selectivity against several bacteria strains with negligible hemolytic activity 70 . The fact that we 316 did not observe this feature in our LDKA peptide library suggests, again, that simple generic 317 structure-function rules are not applicable to membrane active peptides. 318 The dependence of the drastic changes in selectivity and leakage propensity upon small 319 sequence changes demonstrates the limitation of overall macroscopic peptide descriptors such 320 as hydrophobic moment and polar angle as design criteria. For example, the hydrophobic 321 moment is somewhat correlated to hemolytic activity (Figure S4d), and the LDKA peptide library 322 suggests a hydrophobic moment of 3.37 as a cut-off for toxicity toward human red blood cells; 323 however, this does not apply to 26-residue peptide melittin (hydrophobic moment = 3.94) and its 324 membrane-selective analogue (hydrophobic moment = 3.44-3.46). Hydrophobic moment 325 estimates could be limited as they are based on a single, perfectly helical peptides, and do not 326 consider peptide-peptide aggregates and assemblies, as observed in our MD simulations. 71-76 327 Experimentally, fluorescent dye leakage from POPC vesicle is also a reliable model to predict the 328 hemolytic activity with a linear correlation (R-squared value = 0.87; Figure 4b). It is similar to 329 structure-function relationship that shows R-squared value 0.78 between helicity in POPC LUV 330 and hemolysis (Figure 4c).

331
The absence of strong correlations between macroscopic peptide descriptors and 332 selectivity and leakage propensity means detailed structural models are needed to show what is 333 going on. Advances in computer performance have enabled long-scale (multi-µs), fully atomistic 334 MD simulations to provide that picture. 53,[77][78][79][80][81][82][83][84][85][86][87] We have demonstrated before that MD 335 simulations are now able to directly predict aggregate structures in membranes. 45,53,87 The 336 simulations here show the atomic details of how these short membrane-spanning peptides 337 selectively fold, aggregate, and form water pores in specific lipid bilayers (Figure 3). Key to these 338 simulations is that they are not stuck in initial conditions. The pore aggregates are predicted 339 without any prior information and fluctuate sufficiently to reveal the major structural assemblies 340 that these peptides are expected to populate at either equilibrium, or during a membrane 341 permeabilization event. Structures are highly heterogenous. The selectivity found in the 342 experiments is reproduced in the simulations: 7D12 only folds and assembles in anionic POPG 343 bilayers. There are no TM pores for 7D12 in POPC, with the peptides staying on the membrane 344 surface and no noticeable water leakage. This is consistent to the experimental findings, and it 345 demonstrates the extreme effect even tiny sequence changes can have on the pore forming 346 equilibrium. Anionic sidechains are known for their steep insertion penalties, 88 so the reason for 347 the lack of TM insertion likely is the shift of the 2 Asp residues one position towards the center of 348 the peptide. The propensity of a peptide sequence to aggregate and assemble in a given 349 environment depends in a highly complex and non-linear way on the its sequence.  (Figure 5a-b) and biofilm (Figure 5c-e) with micromolar peptide concentrations. 369 Our study demonstrates a simple methodology of the rational design of membrane-370 selective peptides, revealing the potential of using MD simulations to fine-tune the membrane 371 selectivity for peptide design and protein engineering for different cell types. This demonstrates 372 the feasibility of computer-guided antibiotics design 24,90-92 , developing potent antimicrobial 373 peptides that have effective membrane selectivity to distinguish between human red blood cells 374 and bacterial membranes, and even between different bacterial species. The key advantage of in-375 silico techniques is the vastly larger combinatorial space that can be explored in comparison to 376 experimental library screening. In this study, the large-scale all-atomistic simulation effort was 377 limited to only a few sequences and target membranes due to the heavy resources required. 378 However, the strong correlation to the experimental results demonstrates the maturity of these 379 techniques. With rising computing power in the near future, the library screening effort will be 380 shifted towards the computational side. This combined experimental/computational approach 381 opens the path to apply these LDKA analogues, and numerous other designed peptides to various 382 different biomedical applications, e.g. antibiotics, biosensors, and drug delivery. 383 384

385
Combinatorial peptide library synthesis. The synthesis of combinatorial peptide library was 386 modified from the method described by Krauson, et al 26 . Peptides library synthesis was 387 performed using Tentagel® NH 2 macrobeads (280-320 µm bead diameter) particle size (~65,550 388 beads/g) using Fmoc solid-phase peptide synthesis. Each bead only has one peptide sequence. A 389 photolinker is attached between peptide and bead to allow the UV light-induced cleavage of 390 homogenous peptide from bead in each well. The quality of the peptide library was verified by 391 mass spectrometry (e.g. MALDI) and Edman sequencing. After placing one bead in each well of 392 96-well microplate, the photolinker between peptide and bead was cleaved with 5 hr of low-393 power UV light on dry bead, which were spreading to a dispersed single layer in a glass dish. The 394 peptides were each dissolved in DMSO, quantified by tryptophan absorbance using nanodrop, 395 and stored in -20 ˚C freezer.

397
Bulk peptides and chemicals. The selected LDKA-like peptides were synthesized using standard 398 Fmoc chemistry and purified to 98% purity using reverse phase HPLC by GenScript, Inc  399 (Piscataway, NJ, USA). The N-terminus was positively charged amine group and C-terminus is 400 neutral amide group. Peptide purity and identity were confirmed by HPLC and ESI mass 401 spectrometry. The solubility test was performed by GenScript, Inc (Table S1).

449
Tryptophan fluorescent binding assay. The protocol was modified from the original method 450 described by Christiaens,et al. 104 . LDKA peptides (50 µM) and POPC/POPG (600 µM) were 451 prepared in 10 mM phosphate buffer (pH 7.0). The solutions were incubated and measured after 452 60 minutes. Excitation was fixed at 280 nm (slit 9 nm) and emission was collected from 300 to 453 450 nm (slit 9 nm). The spectra were recorded using Synergy H1 Hybrid Multi-Mode Reader and 454 Cytation™ 5 Cell Imaging Multi-Mode Reader from BioTek and were averaged by 3 scans.

471
Biofilm. The formation of biofilm and quantification was modified from the method described by 472 O'Toole, et al. 105 . Escherichia coli strain ATCC 25922, Staphylococcus aureus strain ATCC 25923 473 and Pseudomonas aeruginosa ATCC PAO1 were overnight cultured to log phase OD 600 = 0.3-0.6. 474 Dilute the overnight culture 1:100 into fresh medium for biofilm assays. Add 100 µL dilutions to 475 each well and culture it at room temperature without shaking. After 48 hr incubation, remove 476 the media and rinse each well with 150 µL water for three times. Prepare elevated concentration 477 of AMPs and treat the biofilm using total 150 µL volume in each well. Incubate it for 3 hr at room 478 temperature and remove the supernatant. Rinse each well for three times using water. Add 150 479 µL of a 1% crystal violet in water to each well and incubate the plate at room temperature for 15 480 min. Rinse the plate three times with water to remove the free crystal violet. Turn the plate 481 upside down and dry for overnight. Add 150 µL of 30% acetic acid in water to each well of the 482 plate to solubilize the crystal violet on the cells. Incubate the plate at room temperature for 15 483 min. Transfer 100 µL of the solubilized crystal violet to another plate and quantify the 484 absorbance at 550 nm using Cytation™ 5 Cell Imaging Multi-Mode Reader from BioTek.

486
Drug resistant Escherichia coli. Escherichia coli strain ATCC 25922 was overnight cultured to log 487 phase OD 600 = 0.3-0.6. Initial bacterial cell density was prepared with ~3 x 10 5 CFU/mL in LB broth 488 in 96-well plate. Bacteria were added to serially diluted antibiotics (e.g. ceftazidime, ciprofloxacin, 489 streptomycin, and gentamicin) and co-incubated at 37 ˚C. After 12 hr incubation, the optical 490 density of each well was recorded on a plate reader to determine whether they were sterilized or 491 were at stationary phase growth. The E. coli which survived at the highest antibiotic 492 concentration (below or near the MIC) was collected and cultured for another generation. This 493 cycle is repeated for 10 times until the e. coli have resistant (2-fold higher MIC than its wildtype) 494 against the antibiotics. 495 496 Hemolysis assay. Fresh human red blood cells were obtained from Interstate Blood Bank, Inc., 497 and thoroughly washed in PBS until the supernatant was clear. hRBC concentration was 498 determined using a standard hemocytometer. In hemolysis assays serial dilutions of peptide 499 were prepared, followed by the addition of 2 × 10 8 hRBC/mL. After incubation for 1 hr at 37 ˚C 500 the cells were centrifuged, and the released hemoglobin was measured by optical absorbance of 501 the heme group (410 nm). Negative control was buffer only (0% lysis), and the positive controls 502 were 20 µM melittin and distilled water (100% lysis). The measurements were made in triplicate.

504
Molecular dynamics simulations and analysis. Unbiased all-atom MD simulations were 505 performed and analyzed using GROMACS 5.0.4 106 and Hippo BETA simulation packages 506 <http://www.biowerkzeug.com>, and VMD molecular visualization program 107 507 <http://www.ks.uiuc.edu/Research/vmd/>. The pdb structure of extended peptides (GL 5 KL 6 G, 508 LDKL, and LDKA) were generated using Hippo BETA (see Table S1, Table S2, and Table S3). These 509 initial structures were relaxed via 200 Monte Carlo steps, with water treated implicitly using a 510 Generalized Born solvent. 511 After relaxation, the peptides were placed in all atom peptide/lipid/water systems 512 containing model membranes with 100 mM K and Cl ions using CHARMM-GUI 108 513 <http://www.charmm-gui.org/>. Four helical peptides were initially placed on both interfaces of 514 the bilayer and equilibrated and relaxed for ~600 ns. After equilibration, the system was 515 multiplied by 2x2 matrix in both the x and y directions and results in a bigger system with total 16 516 surface-bound peptides on the bilayer. The simulations were performed at 120 ˚C to speed up 517 the kinetics, and we confirmed their simulated helicity with the liquid-state circular dichroism 518 spectroscopy (Table S2 and Figure S2). MD simulations were performed with GROMACS 5.0.4 519 using the CHARMM36 force field 109 , in conjunction with the TIP3P water model 110 . Electrostatic 520 interactions were computed using PME, and a cut-off of 10 Å was used for van der Waals 521 interactions. Bonds involving hydrogen atoms were constrained using LINCS. The integration 522 time-step was 2 fs and neighbor lists were updated every 5 steps. All simulations were 523 performed in the NPT ensemble, without any restraints or biasing potentials. Water and the 524 protein were each coupled separately to a heat bath with a time constant τ T = 0.5 ps using 525 velocity rescale temperature coupling. The atmospheric pressure of 1 bar was maintained using 526 weak semi-isotropic pressure coupling with compressibility κ z = κ xy = 4.6 · 10 −5 bar −1 and time 527 constant τ P = 1 ps.

529
Oligomer population analysis. In order to reveal the most populated pore assemblies during the 530 simulations, a complete list of all oligomers was constructed for each trajectory frame. An 531 oligomer of order n was considered any set of n peptides that are in mutual contact, defined as a 532 heavy-atom (N, C, O) minimum distance of <3.5 Å. Frequently, this definition overcounts the 533 oligomeric state due to numerous transient surface bound (S-state) peptides that are only loosely 534 attached to the transmembrane inserted peptides that make up the core of the oligomer. These 535 S-state peptides frequently change position or drift on and off the stable part of the pore. To 536 focus the analysis on true longer-lived TM pores, a cut-off criterion of 65° was introduced for the 537 tilt angle τ of the peptides. Any peptide with τ ≥65° was considered in the S-state and removed 538 from the oligomeric analysis. This strategy greatly reduced the noise in the oligomeric clustering 539 algorithm by focusing on the true longer-lived pore structures. Population plots of the 540 occupation percentage of oligomer n multiplied by its number of peptides n, were then 541 constructed. These reveal how much peptide mass was concentrated in which oligomeric state 542 during the simulation time.

544
Membrane partitioning and secondary structure. 545 Peptide solutions (50 µM peptide concentration) were titrated with POPC and POPG LUVs 546 (between 0-5 mM) and the corresponding changes of tryptophan fluorescent spectra were 547 collected ( Figure S1). 7F3 and 28H6 show maximum fluorescent emission of ~331 nm in 548 phosphate buffer, suggesting aggregate formation. Other peptides have tryptophan fluorescence 549 peaks at ~348 nm, indicative of monomeric peptides or low multimeric soluble aggregates. 550 Change of the maximum wavelength indicates the partitioning between water and lipid phases. It 551 gives a direct measure of the binding free energy (∆G binding ) for each peptide with different lipids. 552 Binding free energy of toxic peptides (Figure S1a-j; 4H9, 7F3, 28H6, and 25B2), which porate both 553 POPC and POPG LUVs, are between -5.5 and -9.5 kcal/mol. The membrane-selective peptides 554 (Figure S1k-r; 7G6, 7D12, 11D12, and 24F1) have lower binding free energy toward POPC LUV 555 (∆G binding = -4.0 to -5.7 kcal/mol) than POPG LUV (∆G binding = -5.7 to -7.1 kcal/mol). It shows that 556 the strength of peptide binding is essential for membrane-selectivity and ∆G binding = -5.7 kcal/mol 557 is the cut-off. 558 We further performed CD spectroscopy to study the secondary structure of these 559 peptides with each POPC and POPG LUVs at elevated temperature ( Figure S2 and Table S2). CD 560 spectroscopy shows that all the toxic peptides are helical structure (54-75% helicity) in the 561 solution, and membrane-selective peptides are mostly coiled structure (22-38% helicity). The 562 secondary structure of the peptides in solution explain why these LDKA analogues have 563 selectivity toward different membrane types and result in different binding free energy. Coiled 564 structure exposes its intramolecular hydrogen bonds to water that make the compound more 565 polar; in opposite, helical structure makes it more hydrophobic. Therefore, toxic peptides have 566 higher helical content and strong interaction with both membrane types. Interestingly, 28H6 only 567 folds beta-strand structure in POPG LUV, and the temperature at 95 ˚C can break the 568 intermolecular hydrogen bonds and reverse it to helix. As expected, the membrane-selective 569 peptides only fold helix in POPG LUV and have no response to POPC LUV. Most of the helical 570 structures are highly resistant to thermal denaturation (at 95 ˚C) when they once fold in the 571 membrane ( Figure S2).

572
The linear regression analysis shows strong correlation between hydrophobic moments, 573 helicity in POPC LUV, and ANTS/DPX leakage fraction from POPC LUV (Figure S3a). It confirms the 574 interaction between peptide and POPC LUV is strongly dependent on the peptide's hydrophobic 575 moment; however, it does not correlate to the membrane pore size. Figure S3b shows that the 576 helicity of a peptide is linearly correlated to the hydrophobic moment, which is promoted by the 577 hydrophobicity. We further analyzed the AMPs from APD that have peptide length between 5 578 and 30 amino acids, which dominate >50% peptide population (1,500 AMPs) in APD. We grouped 579 the AMPs by their peptide length and averaged each of their net charge and hydrophobic 580 moment. It shows that increased hydrophobic moment corresponds to higher net charge ( Figure  581 S4a). 582 We analyzed the sequence of LDKA analogues and compared them to the AMPs from APD 583 that have same peptide length to LDKA (Figure S4b). It showed that the toxic LDKA peptides have 584 higher hydrophobic moment 3.41-4.78 than membrane-selective LDKA peptides with 585 hydrophobic moment 1.92-3.32 (Table 1), which correspond to their specificity toward different 586 membrane types ( Figure S4c) and toxicity to human red blood cell ( Figure S4d). We found 587 hydrophobic moment 3.37 is a cut-off between membrane-selective and toxic peptides in the 588 LDKA library peptides. However, the cut-off may shift in different peptide length and charge 589 distribution (Figure S4e-h and Table S5); therefore, bigger sample size is necessary to improve 590 this sequence-based prediction of membrane selectivity. 591 592

593
The data that support the findings of this study are available from the corresponding author on 594 reasonable request.

596
Acknowledgements 597 We thank the Karlsruhe Institute of Technology (KIT) ANKA synchrotron CD beamline staff for 598 support and beamtime. We thank Jochen Bürck at KIT for valuable discussion and technical 599 support for ANKA synchrotron CD beamline. We thank Guangshun Wang at the University of 600 Nebraska Medical Center for providing the raw data of the antimicrobial peptide database. We 601 thank from each POPC and POPG LUVs with P:L = 1:1000 at pH 7 phosphate buffer. The fluorescent dye 790 leakage fraction has been normalized to fit between 0 and 100 % by the positive control (LUV 791 with potent peptide) and negative control (LUV with non-active peptide). †N-terminus is free, C-792 terminus: -NH 2 . ‡Large pore for PG only. aggregates (peptides colored blue (N-) to red (C-) terminal, lipid phosphates as orange beads), 835 oligomeric occupation plots (blue = S-state, yellow = single TM, red-dark = higher TM oligomers, 836 overall distribution on right), and cross-membrane water and ion flux caused by the pore 837 assemblies. c and d, Isothermal titration calorimetry of the heat release/absorption of the 838 peptide-lipid interactions. The integrated ITC data curve of 7D12 and 25B2 with each POPC and 839 POPG LUVs is also shown. The concentration is fixed at 100 µM with titrated lipid LUVs in 10 mM 840 phosphate buffer at pH 7.0. The ITC data is consistent with the simulation results for the binding 841 selectivity of 7D12 for POPG. ciprofloxacin, streptomycin, and gentamicin) were treated with serial E. coli generations. The E. 859 coli that survives below/near the MICs was selected for the next generation. b, MICs of LDKA 860 analogues (membrane-selective peptides: 24F1, 11D12 and 7G6; toxin peptides: 25B2, 7F3 and 861 4H9) against four different strains of drug resistant E. coli. Antibacterial activity of LDKA 862 analogues against quantitative biofilm formation on polystyrene 96-well plate for 3 hr treatment.