The War Against Drug-Resistant Bacteria
Towards the Rational Design of Potent Peptide Antibiotics to Combat Drug-Resistant Bacteria
Prathik Naidu
Johns Hopkins University
Thomas Jefferson High School for Science and Technology
Abstract
The rapid emergence of drug-resistant bacteria has prompted the pharmaceutical industry to research new methods to combat these deadly microorganisms. Antimicrobial peptides (AMPs) are seen as a possible alternative to current drugs; however, they have not entered the pharmaceutical industry because there is minimal information on their interactions with bacterial membranes, they are toxic to human cells, and synthesizing AMPs is expensive and time-consuming. Atomic-scale molecular dynamics (MD) simulations were used in order to model peptide-membrane interactions. The MD simulations revealed the existence of six different peptide structures and helped visualize the function of AMPs. In addition, the simulations also characterized a biochemical mechanism of AMP membrane disruption. An AMP prediction algorithm, which was more accurate than existing methods, was used to expedite the AMP discovery process, making peptide synthesis faster and less expensive. The peptides were screened on eukaryotic-based vesicles and had minimal activity, indicating that the peptides were not toxic to humans. Now, with a detailed understanding of their mechanism of action and a computational-method to speed up drug design, AMPs can potentially serve as the new replacement for current antibiotics.
Introduction
Each year in the United States, there are at least 2 million cases and 23,000 deaths be- cause of antimicrobial resistance, costing our nation nearly 26 billion dollars in healthcare [1]. Current antibiotics, such as penicillin and vancomycin, are becoming ineffective because bacteria have become resistant to these treatments [8]. Antibiotics are designed to bind to specific proteins either on the surface or inside of a cell. When the antibiotics attach to proteins, an important metabolic pathway is disrupted, causing bacterial cell death. However, among a large colony of bacteria, there may be a few that have a mutation in the specific protein that the antibiotic targets. This mutation causes variations in the tertiary structure of the protein and changes the overall structure of the binding site. Although the antibiotics will kill the bacteria that do not have the mutation, the cells that do have this mutation will be unharmed [8]. The bacteria that have not been killed rapidly reproduce, creating a new colony of drug-resistant bacteria. This simple evolutionary process represents the fundamental problem with the conventional antibiotic treatments for bacterial disease [5].
Many pharmaceutical companies tried to overcome antimicrobial resistance by working towards developing more selective and effective drugs. Despite efforts to develop better treatments against bacteria, current antibiotic drug design has dwindled because of its inefficient and expensive process. Designing a drug through common de novo methods can take years because of extensive time needed to screen through chemical libraries consisting of over 600,000 compounds [1]. Each compound is checked individually and even if a potential drug candidate seems to work, it often times exhibits a high cytotoxicity, making it impermissible for use in humans [2]. In addition the cost of screening such large libraries is high because a variety of chemicals, cell cultures, and assays are needed to test a single compound. As a result of these problems, investors have been decreasing the amount of money spent on antibiotic research [5]. Thus, given the ineffectiveness of current antibiotic treatments and drug design, there is an urgent need to develop new treatments that can destroy drug-resistance bacteria.
Antimicrobial peptides (AMPs), which consist of a specific amino acid sequence, are relatively untouched biological molecules that have the potential to combat drug-resistant bacteria [1]. However, there are fundamental problems with AMPs that prevent their use in clinical settings. First, although they occur naturally, most AMPs that are discovered in other organisms such as tree frogs are highly toxic to mammalian cells, rendering them impossible to use with humans. This toxicity is predicted to occur because each organism has unique and specific cellular structures [2]. Second, a majority of discovered AMPs consist of approximately 50 amino acids, making the process of synthesizing individual peptides inefficient and expensive [4]. Scientists have tried to expedite the peptide synthesis process by using FMOC solid-phase methods that consists of over 10,000 different amino acids combinations. Although solid-phase synthesis was shown to be an efficient technique to create peptides, the process of creating thousands of versions in search for a potent peptide is still expensive [4]. Machine learning prediction algorithms, have been implemented to make AMP discovery efficient, but many current methods do not take into account a variety physiochemical properties of the peptides and rely on extensive training before functioning optimally. Moreover, the current algorithms show accuracy percentages that are maximized at approximately 90 percent [9]. Third, the exact mechanism of how antimicrobial peptides function is largely unknown. AMPs are known to interact with cell membranes, but it is unclear how these interactions work [5]. Researchers have proposed possible models of AMP function such as barrel-stave pores, which is when the peptides insert themselves into the membrane, and the toroidal pore, which is when the peptides cause the membrane to fold [7]. Despite these proposals, there is no direct evidence that both explains the mechanism of action and identifies the key components of AMP structure.
The goal of this research was to first use computational simulations in order to build a complete picture of how AMPs function by isolating high-atomic detail structures of these peptides. In this first phase of the project, it was hypothesized that rigorous computational analysis using molecular dynamics can provide information about how AMPs interact with membranes. In addition, it was also hypothesized that the simulations, combined with novel prediction algorithms, can be used to rationally design a small, and inexpensive peptide library that is optimized for antimicrobial activity without toxicity to human cells.
Materials, Methods, and Procedures
Molecular Dynamics Simulations
For the molecular dynamics (MD) simulations, a CHARMM36 package was downloaded and executed on a Linux computer. A protein data bank (.pdb) file for a bilayer containing a total of 104 lipids was acquired through the CHARMM-GUI application. The composition of all the lipids in the .pdb file was 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphatidylglycerol (POPG), which is the primary lipid composition of bacterial cells. A wild-type peptide .pdb file from Litoria genimaculata, which is publically available on the SWISS-PROT protein database, was acquired. The POPG lipid bilayer and the peptide .pdb files were loaded into the VMD visualization software and eight peptides were positioned horizontally above the surface of the lipid bilayer (Figure 1). All peptides were exported into individual .pdb files containing coordinates for each atom, which were copied into the POPG .pdb file. GROMACS, which is a software used to run MD simulations, was downloaded and the grompp command was used to solvate and equilibrate the system with water molecules.
Solvation with water molecules helps to increase the similarity between the simulation and an actual biological environment. However, in order to have a more realistic simulation, mathematical models were implemented in order to account for important properties such as bond length, bond angle, and bond rotation as shown below.
In addition, further parameters, including electrostatic interactions and van der Waals forces were used as shown below. Specifically, these two parameters ensure a realistic representation of how atoms interact with one another and thus play an important role in simulating peptide insertion into the membrane.
These parameters also help in identifying the most stable conformation of the peptides. In each of the formulas, the V represents a potential energy calculation for all of the peptides in the simulation. These formulas aim to minimize the potential energy of the peptides, which would help identify the most stable peptide structure while maintaining high-atomic detail.
After equilibration of the system through the GROMACS package via the trjconv and grompp commands, the simulation .pdb file was transferred to the Maryland Advanced Research Computing Center (MARCC) and was executed on 12 cores using the mdrun command. Once the simulation finished running, VMD was used to visualize the peptide insertion process and PyMol was used to analyze the peptide structures that formed. Finally, scripts were written in Python to quantify ion conductance across the lipid bilayer to determine the activity and mechanism of disruption for the peptide structures.
AMP Prediction Algorithm
The AMP prediction algorithm was developed through the use of programmed machine learning algorithms that are trained and validated. From the SwissProt AMP database, 3200 known AMP sequences were consolidated into a text file. This data set served as a positive test set for the machine learning algorithms. A random set of 3200 non-antimicrobial proteins was downloaded from the UniProt database. This dataset served as a negative control of proteins that did not have any antimicrobial activity and were thus classified as not AMPs (NAMP). Since the protein sizes downloaded from the UniProt database were much larger than the peptides, algorithms would classify based on this length discrepancy, resulting in an imprecise model. Thus, a program was written using MATLAB both to cut down the sequences of proteins into smaller lengths similar to those of AMPs and to reformat the data for machine learning testing.
Once these datasets were created, they were processed through a python script that calculated 10 unique properties of the peptides and proteins: length, GRAVY hydrophobicity, instability index, aromaticity, isoelectric point, secondary structure, turning points, and helicity. These properties served as classification features that helped distinguish between an AMP and NAMP. The two datasets were combined into one large file containing the sequences and their attributes. This combined dataset was inputted into R for training and 10-fold cross validation using the randomForest package for Random Forest, nnet package for Multilayer Neural Network, rpart package for Alternate Decision Tree, and e1071 package for Naive Bayes algorithms. The neural network structures were tested with varying numbers of hidden neurons and the tree algorithms were tuned based on leaf size. The outputs from the training and cross validation included the accuracy, precision, false positive rate, and AUROC, which provide indications of the effectiveness of AMP prediction.
Peptide Library Design and Synthesis
A helical wheel was created from the simulation structures by analyzing the positions of amino acids on the peptides. Various amino acid combinations were designed in order to increase favorable properties such as more polar amino acids or more hydrophobic amino acids. These combinations were tested on the top performing machine-learning algorithm, which helped determine optimal peptide sequences for the library. Once the library was designed, the peptides were synthesized using solid-phase FMOC reactions on 0.5g of Tentagel resin beads. The beads were initially soaked with dimethylformamide (DMF) for 30 minutes in a reaction vessel. After this incubation period, two times the moles of beads of the amino acids, HOBT, and HBTU, were measured added into the reaction vessel containing the beads. The beads, along with the amino acids and reagents, were mixed for 1.5 hours. Once the reacting was completed, the standard Kaiser test procedure was conducted to determine if the reaction was complete. If the reaction tionwas not finished, the beads were mixed for an additional 45 minutes. A deprotection solvent, consisting of 1 milliliter of pyridine and 3 milliliters of DMF was added to the reaction vessel containing the beads and was placed on a rotator for 15 minutes. The beads were rinsed with DMF using a vacuum that connected to the reaction vessel. The above process was repeated for all of the amino acids in the peptide sequence.
Purity of Synthesis Experiments
Matrix Assisted Laser Desorption/Ionization (MALDI) mass spectrometry was used to confirm if the peptides were the correct mass after synthesis. Eight beads were randomly selected using tweezers and were placed into individual wells of a 96 well plate. The beads were incubated with HFIP and MilliQ water and placed under UV light for 3 hours. A micropipette was used to transfer two microliters of the liquid sample from each well onto the mass spectrometry matrix cassette. The samples were spaced evenly and separated to prevent cross-contamination. The cassette was loaded into the MALDI mass spectrometry machine and the spectra were collected for all of the samples. High-performance Liquid Chromatography (HPLC) was used to ensure that the peptide samples did not have any contaminants. A C18 reverse phase LC column was attached to the HPLC and the machine was then calibrated and washed with water and acetonitrile (ACN). All of the remaining solution for each peptide in the 96 well plate was loaded into the machine. The HPLC was run for 30 minutes and the data was exported for further analysis.
Screening of Peptide Library
The remaining beads in the reaction vessel were washed with dichloromethane and solvated with a 10 mL solution containing 88% TFA, 5% phenol, 5% MilliQ water, 2% tri-isopropylsiline. After two hours of solvation, the beads were washed with dichloromethane and were spread across a glass tray. The tray was placed under UV light for a total of four hours. The beads were each placed into wells of 96-well plates using tweezers and 25 uL of MilliQ water was added to ensure the peptide was in solution. Vesicles consisting of 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphatidylcholine (POPC) lipids, which are the primary lipid in eukaryotic mammalian cells, were prepared by combining 10 mL of ANTS/DPX dye with 30 mL of POPC lipid solution purchased from Avanti Polar Lipids in a falcon tube. The mixture was passed through a G-100 extruder to create vesicles that are 100 nm in diameter and contain dye inside of them. A multichannel pipet was then used to transfer 50 uL of POPC vesicles into each well containing the synthesized peptides. The 96-well plate was immediately inserted in a H4 microplate reader in order to detect fluorescence in each of the wells, which would indicate peptide activity on the membrane.
Results
The molecular dynamics simulations, which utilized the implemented mathematical models that ensured a close-to-realistic environment, were successful at identifying peptide structures in the membrane. In just 60 µs after starting the simulation, there were early signs of pore-forming structures in the membrane (Figure 1B). By the time the simulation length reached 1 millisecond, six different peptide structures formed within the membrane, ranging from a dimer pair to an eight-peptide octamer complex (Figure 2).
After observing the various structures, it was evident that the peptides are in an antiparallel formation, with neighboring peptides facing in opposite directions. The structure also indicates a clear central pore that was formed by the interaction of multiple peptides. This pore was predominant in higher-order structures, such as the tetramer, pentamer, hexamer, and octamer. The dimer and trimer have stable structures, but do not seem to exhibit any pore formation in the membrane (Figure 2).
The structures discovered and represented in Figure 2 were subjected to a conductance analysis in order to determine to what degree the different structures disrupt bacterial membranes. In Figure 3, the setup of the conductance analysis from the molecular dynamics simulation is shown. The peptide structure, represented in red, is centered inside the lipid bilayer, which is represented by the orange circles as the phosphate groups and the gray lines as the lipid tails. Each side of the membrane was solvated with water molecules as well as chloride, sodium, and potassium ions. The data in Table 1 summarizes the results of conductance for the higher-order structures represented as the number of events per microsecond.
One event occurs when a single ion passes completely through the membrane to the other side. Figure 4, which is shown below, illustrates the high-atomic detail in the simulations through the identification of specific locations of amino acids on the peptides. After analyzing the amino acids and their respective physiochemical properties, a helical wheel was developed that represents the locations of amino acids on a single peptide (Figure 5). The red region of the helical wheel corresponds to polar amino acids that face the inside of the peptide structure because the inner pore conducts polar molecules like water. The two blue regions consist of amino acids such as valine and alanine, which are able to form hydrogen bonds that stabilize the overall structure. The orange region contains nonpolar amino acids that reduce the free energy barrier during peptide-membrane insertion.
As mentioned in the methods, the helical wheel was used to design various amino acid combinations that optimize specific properties of the peptide, such as polarity. The results of the machine learning al gorithms, in Table 2, indicate that the Random Forest method has the highest performance in AMP prediction. Thus, the various changes in peptide sequences were tested with the Random Forest algorithm and a library consisting of 2916 combinations was designed (Figure 6). The wild-type sequence is shown in Figure 6, along with the amino acid substitutions at various points in the sequence that result in a total of 2916 peptide combinations.
The results of the HPLC test indicate that the peptides did not have any foreign contaminants because there was one single peak at 12.712 minutes, which is when the peptide flowed through the HPLC machine. The mass spectrometry result indicates that the peptides were synthesized to the correct length because the mass of 2223.114 falls within the expected range. There is a second peak on the mass spectrometry graph, however this peak, which is at a mass of 2245.121, differs from the first peak by a mass equivalent to a sodium ion. This difference indicates that the second peak is not a contaminant but rather an additional sodium ion that was attached to the peptide from the mass spectrometry matrix solution.
Discussion
The formation of a peptide structure in the membrane in just 60 µs, as shown in Figure 1, illustrates the rapid speed at which the AMPs can interact with the membrane. Moreover, the molecular dynamics simulations yielded results which show that peptides interact with other peptides to form complexes that exist in the bacterial membrane. All of these structures in Figure 2, from dimer to octamer, were found to coexist simultaneously, suggesting that the peptides have multiple paths towards disrupting the bacterial membrane. The structures themselves provide the first clear representation of the mechanism of how the peptides create pores in the membrane. These results provide useful information for future development of AMP-based drugs because the simulations provide a visual representation of the process that the peptides take to insert and create complexes in the membrane. Pharmaceutical companies can use the novel simulation data to further develop optimized antimicrobial drugs based on the discovered mechanism of action for the AMPs. The molecular dynamics approach in this project utilized mathematical formulas to simulate realistic atomic interactions and thus provides further confirmation of these results.
Each of the peptides structures were in an environment that contained water and ions. The setup in Figure 3 illustrates a cross section view of the peptide structure inside the bacterial membrane. This result provides substantial evidence that the specific model for peptide insertion in the membrane follows the ‘barrel-stave’ pore. There has been previous speculation about what type of pores AMPs form inside of a membrane because the peptides are hard to visualize even with high-resolution electron microscopes. However, the molecular dynamics simulation in Figure 3 shows that the peptide inserts and creates a pore in the membrane, which means it follows the ‘barrel-stave’ model.
The molecular dynamics simulations further illustrate the ability to understand AMP structures on a high-atomic level detail as shown in Figure 4 by being able to identify and analyze specific amino acids of the peptides, which supports the initial hypothesis in this experiment that the simulations can provide detailed representations of the AMPs to understand the mechanisms of their function. The helical wheel in Figure 5 was used along with the random forest machine-learning algorithm to develop a 2916-member peptide library as shown in Figure 6. With respect to the AMP prediction, the random forest method likely outperformed the other algorithms (Table 2) because it uses an in depth multilayer decision tree that takes into account each attribute, such as GRAVY hydrophobicity, as its own individual entity unlike the other algorithms, which build relationships among attributes. Although it takes slightly longer to process data and predict, the accuracy of the algorithm using the specific attributes in this project exceed the max accuracy of current algorithms by nearly four percent. In addition, the high precision and low false positive values suggest that this prediction method did not receive the accuracy score by random chance, but instead that the algorithm is strategically trained to classify sequences. The peptide library containing 2,916 peptides may seem large; however, compared to current approaches, which create libraries consisting of more than 17,000 peptides, this project’s library size is much smaller. The small library size was generated through the use of the prediction algorithm, which automatically filters sequences that would not have antimicrobial activity without having to synthesize the peptides, thus saving over 80% of the total cost in synthesizing conventionally sized peptide libraries.
Results from the HPLC and mass spectrometry experiments indicate that the peptide synthesis did not have any contaminants and that the synthesis was done correctly. The peptides in the library were testing against POPC lipid vesicles, which are the mammalian cell models, in order to determine toxicity. The results in Figure 9 illustrate that a vast majority of the peptides in the library had less than 5% leakage efficiency on the POPC vesicles. Thus, the peptides were not efficient at creating pore structures and causing membrane disruption on these POPC vesicles, which means the designed peptides are not toxic to human cells and can be used clinically. The results from the AMP prediction algorithm and the leakage efficiency experiment support the second hypothesis in this research.
The discovered AMPs are more effective at targeting bacteria and preventing resistance when compared to conventional antibiotics because these peptides target the bacterial membrane. Thus, if bacteria wanted to gain resistance, they would have to undergo a system-wide alteration to change the lipid composition of their membrane. This process would not be favorable since the bacteria would also need to change the structure of proteins in the membrane. In addition, the ensemble of various structures used by the AMP in this project acts as a combination of different “drugs” so the bacteria would need to gain resistance to all of these different structures. Thus, the characterized AMPs not only provide a potent approach to targeting and destroying deadly bacteria, but also reduce the likelihood of the development resistant bacteria.
One of the primary limitations in this research is the use of computational simulations to understand the mechanism of AMP function. Current experimental technologies to analyze such small scale processes have not been completely developed for use in labs, and the simulations use models and a realistic system to represent natural conditions as closely as possible. Molecular dynamics simulations are well known and accepted methods in computational biology, but further research should focus on confirming the structures in the lab once the technology has sufficiently developed. In addition, while the results confirm minimal toxicity against mammalian cells, the peptides need to be tested on bacterial membranes, even though the computational simulations suggest potent activity against bacteria. On going research is currently being done at the lab to continue to screen the peptides for activity on bacterial membranes.
Conclusion
In the race for new methods to combat drug-resistant bacteria, antimicrobial peptides (AMPs) have gained attention, but the lack of information on how AMPs function, the expensive and time consuming process of synthesizing peptides, and the high toxicity against human cells currently makes AMPs an ineffective option. Through the use of molecular dynamics simulations that model real-life atomic interactions, six unique structures were identified that explain how AMPs insert into the membrane and create a pore structure. A conductance analysis on these structures, which details how molecules pass through the peptide pores, provides further evidence on both the effectiveness of AMPs and the mechanism of bacterial membrane disruption. These structural and biochemical mechanisms of AMPs further the understanding of these complex molecules in the scientific community. Moreover, the use of an accurate and precise AMP prediction algorithm provides the basis for a rational design of peptide libraries that are smaller and thus, easier and less expensive to synthesize. The results from this experiment indicate that these rationally designed peptides had minimal leakage on mammalian cell model vesicles, indicating the discovery of a set of AMPs that can be used to treat antibiotic-resistant infections in humans. Overall, this research provides insight into a new biological molecule that may serve as the next-generation of antibiotics.
References
[1] Christopher Fjell, Jan Hiss, Robert Hancock, and Gisbert Schneider. Designing antimi- crobial peptides: form follows function. Nature Reviews, 11:37–51, 2012.
[2] Julia Koehler Leman, Martin Ulmschneider, and Jeffrey Gray. Computational modeling of membrane proteins. Proteins, 83:1–24, 2015.
[3] Aram Krauson, Jing He, and William Wimley. Gain-of-function analogues of the pore- forming peptide melittin selected by orthogonal high-throughput screening. Journal of the American Chemical Society, 143:12732–12741, 2012.
[4] G. Maccari, R. Nifosi, and M. Di Luca. Rational development of antimicrobial pep- tides for therapeutic use: design and production of highly active compounds. Microbial Pathogens and Strategies for Combatting them, pages 1265–1277, 2013.
[5] X. Ng, B. Rosdi, and S. Shahrudin. Prediction of antimicrobial peptides based on se- quence alignment and support vector machine-pairwise algorithm utilizing lz- complexity. BioMed Research International, 2015.
[6] R Palffy, R. Gardlik, M. Behuliak, et al. On the physiology and pathophysiology of antimicrobial peptides. Molecular Medicine, 2009.
[7] Sander Pronk, Szilard Pa, Roland Shculz, Per Larsson, Par Bjelkmar, et al. Gromacs 4.5: a high-throughput and highly parallel open source molecular simulation toolkit. Bioinformatics, 29(7):845–854, 2013.
[8] Drug Resistance. National institute of health. NIAID, 2015.
[9] Ng Xin, Rosdi Bakhtiar, and Shahrudin Shahriza. Prediction of antimicrobial peptides based on sequence alignment and support vector machine- pairwise algorithm utilizing lz-complexity. BioMed Research International, 2015.
