ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. RELATIONSHIP BETWEEN BIOPHYSICAL PROPERTIES OF ANTIMICROBIAL PEP- TIDES (AMPS) AND THEIR ASSOCIATED DRUG EFFI- CACIES ALISHA ZHU ✵ ABSTRACT Antibiotic resistance is a growing global health threat. One consequence is that patients with cystic fibrosis (CF) are prone to developing antibi- otic resistant lung infections caused by multiple strains of bacteria, including Pseudomonas aeru- ginosa. Due to the limited number of treatment op- tions for patients with chronic antibiotic resistant in- fections, there is a need for finding new antibiotics that allow for effective eradication of bacterial infec- tions, such as those in the CF lung. Many antimicro- bial peptides (AMPs) have been annotated in data- bases and are considered as potential alternatives for current antibiotics. However, in many instances, the suitability of AMPs as drug molecules has not been extensively explored. Here, we propose that certain molecular properties of AMPs favor high an- tibiotic efficacy. Using information from AMP data- bases, we combined statistical analyses and ma- chine learning techniques to identify relationships between various biophysical properties of AMPs and their drug efficacies. Analyses from classifica- tion and regression trees (CART) and random for- ests suggest that net charge and maximum average hydrophobic moment are the most important prop- erties in determining if a peptide is useful against P. aeruginosa infections in CF patients. Maximum av- erage hydrophobic residue, average alpha helix propensity score, hydrophobic proportion, and peptide length still contribute to this determination but to lesser degrees. Cation-pi interactions, on the other hand, do not appear to factor into this deci- sion at all. Based on these properties, our current work is focused on designing and experimentally testing new peptides that may have activity against P. aeruginosa infections. 2 INTRODUCTION Antibiotic resistance is a persistent issue in patient care that affects more than 2.8 million indi- viduals, with an estimated occurrence of ~35,000 deaths in the U.S. each year (CDC, 2019). Deemed as a serious threat by the CDC, Pseudomonas aeru- ginosa, a type of rod-shaped, Gram-negative bacte- ria, is a major source of infection that requires greater attention. Gram-negative bacterial infec- tions tend to be difficult to treat because they have an extra outer protective membrane layer. Due to buildup of mucus and formation of bacterial bio- films that prevent the delivery of antibiotics at the levels needed for bacteria eradication, patients with cystic fibrosis (CF) are susceptible to developing an- tibiotic resistant lung infections caused by bacteria, of which the most common in adults is P. aeru- ginosa. CF is a recessive genetic disorder that arises from mutations in the cystic fibrosis transmembrane conductance regulator (CFTR) gene. Loss of func- tion of the CFTR decreases hydration in the airways, leading to decreased mucociliary clearance and bacterial retention (Allen et al., 2020). Conse- quently, 80% of CF patients develop chronic infec- tions (Moreau-Marquis et al., 2008). Thus, there is a need for the discovery of new antibiotic agents. An- timicrobial peptides (AMPs) are a class of small pep- tides that exist in nature. Because they exhibit a broad range of activity against various bacteria, AMPs hold potential as small molecule antibiotics (Ageitos et al., 2017). AMPs have various modes of action. First, AMPs can directly kill bacteria by disrupting the cell membrane. AMPs can destabilize and permeabilize the bacterial cell membrane through both hydro- phobic and electrostatic interactions. In many cases, the positive net charge of AMPs allows them to in- teract with the negatively charged regions of the ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V bacterial surface. These interactions often lead to the formation of membrane pores resulting in lysis, or the rupture of the cell membrane (Li et al., 2021). Second, AMPs can prevent the formation of bacte- rial cell walls by blocking peptidoglycan elongation. Peptidoglycan is a polymer consisting of sugar chains interlinked with peptides. Since the bacterial cell wall is an essential protective barrier and rein- forces cell shape, inhibiting cell wall biosynthesis ef- fectively kills the bacteria (Chen et al., 2020). Third, AMPs can penetrate the cell membrane and act on intracellular targets, including RNA, DNA, and ribo- somes. Although this mode of action does not result in cell lysis, it interferes with bacterial bioprocesses. For instance, some AMPs block translation initiation by binding to the P-site on ribosomes, while others inhibit protein synthesis by preventing RNA splicing post-transcription (Ageitos et al., 2017). Promising AMPs proposed in literature, par- ticularly cationic antimicrobial peptides (CAMPs)— i.e., AMPs that have an overall net positive charge— suggest that certain characteristics of peptides favor activity against Gram-negative bacterial infections, such as those caused by P. aeruginosa (Lei et al., 2019). Although AMPs have been documented in databases, in many cases, their microbiological ac- tivity is not established. The rules governing which peptides are likely to be effective are not under- stood, especially for a specific species of bacteria such as P. aeruginosa (Ramazi et al., 2022). Thus, we propose to use database analysis to explore the re- lationship between biophysical properties of AMPs and their corresponding effectiveness. We define effectiveness based on a peptide’s minimum inhibi- tory concentration (MIC), which is the lowest con- centration of peptide that prevents visible bacterial growth. We selected the Database of Antimicrobial Activity and Structure of Peptides (DBAASP) for our studies. A set of candidate peptides designed to be efficacious will be identified and tested against clin- ical strains of P. aeruginosa. 3 METHODOLOGY DATABASE SELECTION AND FILTERING Numerous databases have been con- structed to provide classified information for productive AMP design and research. The DBAASP is one of the larger AMP databases, since it contains information on both synthetic and natural peptides (Pirtskhalava et al., 2021). The inclusion of both syn- thetic and natural peptides widens the scope of an- timicrobial research and development by allowing for a more comprehensive exploration of the pep- tide design space. At the time of use, the DBAASP contained information on 18,000+ peptides. How- ever, the complete DBAASP data set file contained 150,000+ entries; several peptides had multiple data entries containing results from different exper- iments. We implemented initial conditions to ac- quire a suitable subset. For the DBAASP, selected peptides must be monomers, cannot contain any D- amino acids (as we do not wish to work with syn- thetic amino acids), and cannot contain any uncom- mon or variable amino acids. These conditions en- sure that the subset consists of peptides, not pro- teins, that are more likely to be compatible with bi- ological systems. In addition, the peptides must have lengths less than or equal to 50 amino acids but also greater than or equal to 11 amino acids. Shorter lengths enable AMPs to have greater effi- ciency and flexibility when interacting with bacterial cell membranes, which is why an upper limit for peptide length was specified. However, the AMPs also had to have lengths of at least 11 amino acids long, since this represents three turns of an alpha helix spanning the minimum thickness of a Gram- negative bacterial cell membrane. Most im- portantly, the peptide must have known activity against P. aeruginosa, but its MIC value cannot ex- ceed 500 µM. A high MIC value indicates that more peptide is required for inhibiting growth; we de- fined a MIC > 500 µM as ineffective for treating P. aeruginosa infections. If a peptide’s reported MIC was a lower threshold value (i.e., greater than some value), then we considered its MIC value to be dou- ble its lower threshold. This was done to account for one more factor of two in the two-fold serial dilution performed in a MIC assay. If a peptide’s reported MIC was an upper threshold value (i.e., lower than some value), then we considered its MIC value to be equal to its upper threshold. If a peptide had numer- ous reported MIC values, the lowest MIC was ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V selected. After these initial conditions were imple- mented, we arrived at a subset of 4,218 unique pep- tides. MOLECULAR PROPERTIES OF BIOPHYSICAL SIGNIFICANCE Based on what is known about AMPs, there are several molecular properties that are relatively simple to compute and are likely to influence AMP activity: PEPTIDE LENGTH. The peptide length is the to- tal number of amino acid residues in the peptide se- quence. AMPs are typically quite short; most of them have lengths less than 50 amino acids, but some can have lengths as large as 100 amino acids (Ageitos et al., 2017). NET CHARGE. A peptide’s net charge is calcu- lated by subtracting the total number of negatively charged amino acid residues from the total number of positively charged amino acid residues. 𝑁𝑒𝑡 𝐶ℎ𝑎𝑟𝑔𝑒 = 𝐿𝑦𝑠 + 𝐴𝑟𝑔 − 𝐴𝑠𝑝 − 𝐺𝑙𝑢 (1) Net charge indicates how an AMP will elec- trostatically interact with the bacterial membrane. Studies have shown that decreasing the net charge to less than +4 rendered peptides inactive, and in- creasing net charge improved antimicrobial activity. However, an increase to net charges greater than +9 led to a dramatic rise in hemolytic activity (Jiang et al., 2008). This is undesirable because greater he- molytic activity leads to red blood cell damage and can result in anemia, inflammation, and organ dys- function. HYDROPHOBIC PROPORTION. The hydrophobic proportion refers to the percentage of hydrophobic amino acid residues within the peptide sequence. Hydrophobic residues facilitate interactions with the phospholipid fatty acyl chains of the bacterial mem- brane (Edwards et al., 2016). AMPs typically contain 40%-60% hydrophobic residues (Ye et al., 2019). We considered an amino acid to be hydrophobic if it exceeds 0.700 on Fauchere and Pliska’s (1983) hy- drophobicity scale based on octanol-water partition coefficients, which is defined as the ratio of the pep- tide’s concentration in the octanol phase to its con- centration in the aqueous phase. If the coefficient value is greater than one, then the peptide is more lipophilic (dissolves in lipids); if the coefficient value is less than one, then the peptide is more hydro- philic (dissolves in water). 𝐻𝑝𝑟𝑜𝑝 = 𝐶𝑦𝑠+𝑃ℎ𝑒+𝐼𝑙𝑒+𝐿𝑒𝑢+𝑀𝑒𝑡+𝑃𝑟𝑜+𝑉𝑎𝑙+𝑇𝑟𝑝+𝑇𝑦𝑟 𝑃𝑒𝑝𝑡𝑖𝑑𝑒 𝐿𝑒𝑛𝑔𝑡ℎ (2) MAXIMUM AVERAGE HYDROPHOBIC MOMENT. The hydrophobic moment is a measure of the am- phiphilicity, or asymmetry of hydrophobicity, of a peptide’s structure. A large hydrophobic moment corresponds to a structure that is primarily hydro- phobic on one side and primarily hydrophilic on the other side. Eisenberg et al. defined the hydropho- bic dipole moment as 𝜇 = √[∑ 𝐻𝑖 cos(𝑖𝛿)𝑖 ]2 + [∑ 𝐻𝑖 sin(𝑖𝛿)𝑖 ]2 (3) where δ represents the angle between the amino acid side chains (δ = 100 for an alpha helix), i represents the residue number in the ith position, and Hi represents the ith amino acid’s hydrophobi- city based on an averaged consensus scale (Eisen- berg et al., 1982, 1984). We modified this equation to determine the maximum average hydrophobic moment across an 11 amino acid window, which represents three turns of an alpha helix, with δ rang- ing from 90° to 110° instead of fixing it at 100°. By adding these modifications, we hope to account for peptides that have alpha helical segments but may not have a complete alpha helical structure. Addi- tionally, instead of using the averaged consensus scale, values from Fauchere and Pliska’s (1983) scale were used. < 𝜇𝑚𝑎𝑥 > = √[∑ 𝐻𝑖 cos(𝑖𝛿)𝑖 ]2 + [∑ 𝐻𝑖 sin(𝑖𝛿)𝑖 ]2 11 (4) MAXIMUM AVERAGE HYDROPHOBIC RESIDUE. The hydrophobic residue calculation provides a meas- ure of the average hydrophobicity across an 11 amino acid window. As in the maximum average hy- drophobic moment calculation, Hi represents the ith amino acid’s hydrophobicity based on Fauchere and Pliska’s scale (1983). ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V < 𝐻𝑚𝑎𝑥 > = ∑ 𝐻𝑖𝑖 11 (5) AVERAGE ALPHA HELIX PROPENSITY SCORE. Alpha- helices are the most common secondary structure seen in AMPs. Studies have demonstrated that in- creased alpha-helical content correlates to greater antimicrobial activity (Li et al., 2021). The alpha helix propensity score describes how likely a peptide will form an alpha helix. Pace and Scholtz (1998) created a helix propensity scale based on experimental studies of proteins and peptides. The propensity score is given in terms of the amount of energy re- quired for a given amino acid to fold into an alpha helix, with a higher propensity score corresponding to a smaller probability that the peptide will form an alpha helical shape. We modified Porto et al.’s (2018) alpha helix propensity score calculation by dividing it by peptide length to standardize the pro- pensity score across the entire peptide. < 𝛼 > = ∑ 𝑒𝐻𝑥𝑖𝑖 𝑃𝑒𝑝𝑡𝑖𝑑𝑒 𝐿𝑒𝑛𝑔𝑡ℎ (6) Hxi represents the ith amino acid’s helix propensity, based on the Pace–Schols scale. CATION-PI INTERACTIONS. Cation-pi interactions between the positively charged side chain of an ar- ginine amino acid residue and the aromatic side chain of a tryptophan amino acid residue are of in- terest because studies suggest that tryptophan con- taining peptides have higher affinity for deeper in- sertion into bacterial cell membranes (Hicapie et al, 2018). Assuming that the peptide adopts an alpha helix secondary structure, a cation-pi interaction ex- ists only for the configuration where an arginine is four residues after a tryptophan (Shi et al., 2022). We computed a simple count for the number of cation- pi interactions within a peptide sequence. 𝑇𝑟𝑝 ⟶ 𝐴𝑟𝑔 (𝑖, 𝑖 + 4) (7) STATISTICAL ANALYSES AND MACHINE LEARNING TECH- NIQUES SCATTERPLOT MATRIX. Using the psych R pack- age, a scatterplot matrix was generated to visualize the pairwise relationships between the selected bi- ophysical features. A scatterplot matrix helps to graphically and quantitatively identify if multicollin- earity exists among the various molecular proper- ties of interest. PRINCIPAL COMPONENT ANALYSIS (PCA). As a di- mensionality reduction technique, PCA is appropri- ate for our purposes because the DBAASP subset is a large data set in which each observation contains a high number of features (each peptide is associ- ated with seven molecular properties). PCA was conducted using the factoextra R package and is ac- complished by linearly transforming the data onto a new coordinate system. Information is projected onto a two-dimensional space, increasing interpret- ability while preserving the maximum amount of in- formation (Jolliffe & Cadima, 2016). DECISION TREES. Classification and Regression Tree (CART) is a predictive modeling approach. It il- lustrates how peptide effectiveness (response varia- ble) can be predicted based on the seven biophysi- cal properties (explanatory variables). While classifi- cation trees are used for predicting a qualitative out- come, regression trees are used for predicting a quantitative outcome (Krzywinski & Altman, 2017). Using the rpart and rpart.plot R packages, we gen- erated CART for the DBAASP subset. For the classi- fication tree, the peptides were separated into two classes based on a MIC threshold that we specified: noneffective (MIC ≥ 4 µM) and effective (MIC < 4 µM). For the regression tree, the peptides were par- titioned based on their biophysical properties, and an average log2(MIC) was reported for each sub- group. log2(MIC) was used in place of the MIC to re- duce the skewness of the MIC data. log2(MIC) was the most sensible transformation to apply because MIC assays are based on two-fold serial dilutions. RANDOM FORESTS. The random forest algo- rithm is an ensemble learning method that aggre- gates multiple decision trees to create more accu- rate predictions (Brieman, 2001). Like decision trees, random forests can handle both regression and classification tasks. However, it is a more robust approach than CART because it controls overfitting. Functions from the randomForest R package were implemented to construct a random forest ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V regression model. The model included all seven bi- ophysical properties as the explanatory variables with log2(MIC) as the response variable. Based on the method used to generate the prediction algo- rithm by Thomas et al. (2010), 70% of the 4,218 pep- tides in the DBAASP data set were used as the train- ing set and 30% were used as the testing set. The training set was used to build a random forest re- gression model that would predict the log2(MIC) of peptide sequences. DESIGNING AND TESTING PEPTIDES DESIGNING PEPTIDES. Based on the features learned from the bioinformatic analyses of the DBAASP, we wanted to design new, synthetic pep- tide sequences. We generated a set of 50,000 ran- dom peptide sequences that had the same length and amino acid distributions as those in our DBAASP subset (FIGURE 1). To control the amino acid distribution, we assigned each amino acid with a probability of occurrence based on the frequency that each appeared in the DBAASP subset. We then computed the seven biophysical properties of inter- est for all 50,000 peptide sequences. From there, we were able to use the random forest regression model to generate predictions for their log2(MIC) values. SELECTING PEPTIDES FOR EXPERIMENTAL TESTING. Our goal was to determine if any of the designed peptides were effective in vitro against P. aeru- ginosa infections. Assuming that the peptides with the lowest predicted MICs have the greatest likeli- hoods for being effective, we decided to examine 20 generated peptide sequences with the lowest predicted log2(MIC) values more closely. For the 20 peptide sequences, we used the Hemolytic Activity Prediction for Peptides Employing Neural Networks (HAPPENN) tool to calculate the predicted hemo- lytic activity of the peptides (Timmons & Hewage, 2020). The HAPPENN tool provides a PROB score between 0 and 1. A peptide with a score of 0 is pre- dicted to be most likely non-hemolytic, while a pep- tide with a score of 1 is predicted to be most likely hemolytic. Because the HAPPENN tool can only make predictions on sequences with lengths be- tween 7 and 35 amino acids long, we were only able to deduce the potential hemolytic activity for 11 out of the 20 peptides. Of those 11 peptides, 8 peptides were predicted to be most likely non-hemolytic. Therefore, we decided to select those 8 peptides for experimental testing. FIGURE 1: (a) Distribution of lengths for n = 4,218 peptides. (b) Distribution of amino acid residues that make up the n = 4,218 peptides ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V Sequence Predicted 𝒍𝒐𝒈𝟐(MIC) AKRTQRFPRWKKCLRLGFVGCKGNILKAA -0.175 RVRKGAGTSRKLKIVKNLGRHIVWFKGIP 0.105 IRKYRPGLFAKFHKLLKNRKIGGKNL 0.141 MKSKPTVIMRYRFWRVGKL 0.307 RQACGTKAKARLRKPRCLTIGRRVVRKFSKWR 0.329 KMVAKKVKIKRCKRKVHKLPGFGSISIL 0.335 LIAKYGKHAKFKAKKKPQGISGVPKRFYKALWIG 0.432 PRRIKTGAAKRKPKLSKKWNQKLLKKLTPFGW 0.464 TABLE 1: List of 8 peptide sequences selected from 50,000 computer generated sequences. These peptides are predicted to be both antimicrobial and nonhemolytic. 4 RESULTS AND DISCUSSION RELATIONSHIP BETWEEN MOLECULAR PROPERTIES The scatterplot matrix illustrates the rela- tionship between each pair of molecular properties (FIGURE 2). Since none of the correlation coefficients exceeds 0.70, it is generally agreed upon that none of the molecular properties is strongly linearly cor- related with each other (Ratner, 2009). This demon- strates that the properties are non-redundant, and they may all influence a peptide’s effectiveness. However, it is important to note that there exists a moderate, positive correlation between hydropho- bic proportion and maximum average hydrophobic residue. This was expected because both hydro- phobicity-based calculations were dependent on Fauchere and Pliska’s scale. FIGURE 2: Scatterplot Matrix. The diagonal panels provide histograms for each molecular property. The panels below the diagonal depict bivariate scatter plots, which provide a visual representation of the relationship between two variables, allowing for pattern identification, outlier detec- tion, and linear correlation assessment. The panels above the diagonal report the Pearson correlation coefficients as- sociated with the regression line in each scatter plot—a value close to +1 or -1 signifies strong correlation while a value close to 0 signifies weak correlation. ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V The PCA projects information from the seven-dimensional space onto orthogonal axes that are linear combinations and capture the variation of the original variables. The first two dimensions, also called principal components, account for 30.1% and 19.3% (FIGURE 3), respectively, of the total variation in the DBAASP subset. This means that the two-dimen- sional scatter plot, often used in PCA due to its facil- ity for visual interpretation, is not a faithful represen- tation of the original data in the seven-dimensional space, since it only includes 50% of the variation in the DBAASP data set. To account for greater varia- tion, higher dimensions are needed. Nevertheless, the PCA provides insight into how variation is dis- tributed among the molecular properties. The PCA shows that the greatest contributors to the variation are length, hydrophobic proportion, and maximum average hydrophobic residue. FIGURE 3: Principal Component Analysis. (a) Plot of molec- ular properties and their respective contributions. (b) Bip- lot of individual peptide data points and their molecular properties. FIGURE 4: CART. For both classification and regression trees, each decision node is associated with a condition. If the condition is true (or yes) for a peptide, the peptide will be sorted down the left branch and either reach a terminal node or proceed to the next condition. (a) Classification tree built for an n = 4,218 sample, where partitions are based on MIC = 4 µM threshold. Each box displays the overall classification at that node, and the left value repre- sents the number of peptides that are effective while the right value represents the number of peptides that are not effective. (b) Regression tree built for an n = 4,218 sample, where the response variable is log2(MIC). Each node box displays the predicted log2(MIC) value and the n number of peptides that contributed to that predicted value. RELATIONSHIP BETWEEN MOLECULAR PROPERTIES AND PEPTIDE EFFECTIVENESS The conditions used in the classification tree to split the DBAASP subset indicate that all the mo- lecular properties, except for cation-pi interactions, are important in partitioning the data. For classifica- tion purposes, we defined a peptide as effective if ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V its MIC value for P. aeruginosa is less than 4 µM. Based on the classification tree that we generated, it appears that the lowest ratio of noneffective to ef- fective peptides (69 to 139) is associated with a group of peptides that satisfies the following condi- tions: net charge ≥ 4.5, maximum average hydro- phobic moment ≥ 0.6981, and length ≥ 29.5 amino acids (FIGURE 4A). Regarding the regression tree that we generated, the subgroup of peptides with the lowest average log2(MIC) is composed of 208 pep- tides and satisfies the following conditions: net charge ≥ 4.5, maximum average hydrophobic mo- ment ≥ 0.6977, and length ≥ 29.5 amino acids (FIG- URE 4B). These conditions are similar to those de- scribed for the classification tree, suggesting that the information provided by the two trees agree. Us- ing these conditions, we can make predictions for new peptide sequences that were not included in the DBAASP subset. While the decision trees provide simple vis- ualizations and predictive methods, the predictions are not always accurate. By using multiple trees, random forests produce more accurate predictions and provide more reliable information of variable importance. The feature importance attribute of the random forest regression model indicates that net charge and the maximum average hydrophobic moment are the most distinctive biophysical prop- erties influencing peptide effectiveness (Figure 5). The feature importance attribute also confirms that the presence of cation-pi interactions contributes marginally to the fit of the regression model. The other four properties were of similar importance. Overall, the random forest regression model had a mean of squared residuals of 4.409, which suggests that the model is incorrect by 2.1 log2(MIC) units on average, and the % variance explained by the model is 37.83%. While the fit of the model is lacking, the regression model is still worth exploring, as our goal is not to be able to predict exact MICs, but to de- velop a selection method for identifying effective peptides based on their biophysical properties. FIGURE 5: The importance of a biophysical property is determined by measuring the increase in mean square error and the increase in node purity when that property is randomly permuted. Important properties result in substantial changes when randomly permuted, leading to greater increases in mean square error and node purity. ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V 5 CONCLUSION AMPs serve as a rich source of alternatives to conventional antibiotics and provide a potential solution for the antibiotic resistance crisis. Existing AMP databases are continuously being updated to account for the expanding knowledge known about them. However, a comprehensive understanding of the relationship between molecular properties of AMPs and their effectiveness against bacterial infec- tions, especially from a clinical perspective, remains elusive. To develop this comprehension and to ac- celerate the process for screening peptides, this study aimed to create a biophysical understanding and predictive model for classifying peptides as ef- fective or noneffective. To do so, we used a relevant subset of 4,218 AMPs from the DBAASP. To ensure that the selected molecular prop- erties are not correlated with each other and thus all contribute separately to peptide effectiveness, we constructed a scatterplot matrix. It is apparent that none of the properties are strongly linearly corre- lated with each other. This was unexpected, consid- ering that three of the seven selected properties were related to the hydrophobicity of the peptide. At the same time, this was promising because it in- dicates that multicollinearity was not present in our model, and thus each property individually contrib- utes to peptide effectiveness. To further examine the contributions of each molecular property to the total variation within the DBAASP, we performed a PCA. Length, hydrophobic proportion, and maxi- mum average hydrophobic residue had the great- est amount of information captured in the first two dimensions. We used CART to gain insight into which bi- ophysical properties were most important for decid- ing peptide effectiveness. Both the classification tree and regression tree reached a consensus that peptides with the highest potential to become effi- cacious drugs have a maximum average hydropho- bic moment ≥ 0.75 and have a net charge ≥ 4. The finding about net charge is not surprising and is consistent with available literature (Jiang et al., 2008). A predictive regression model using the random forest algorithm was generated to assess the potential antimicrobial activity of a given pep- tide sequence. This predictive tool was used to eval- uate a set of 50,000 computer generated peptide sequences. Based on the predicted antimicrobial and hemolytic activities of those peptide se- quences, we identified a set of eight peptides that may have activity against P. aeruginosa lung infec- tions. In the future, we wish to assess the accuracy of the random forest predictive tool, and we hope to confirm the activity of those 8 peptides experimen- tally by conducting MIC assays. Furthermore, we wish to expand upon our understanding of AMPs by exploring other methods, including non-linear di- mensionality reduction, PCA regression, and imple- menting a more formal search for designing pep- tides through following a genetic algorithm frame- work∎ 6 ACKNOWLEDGEMENTS This work is supported by grant #R21AI151746 from the National Institutes of Health. 7 REFERENCES [1] Ageitos, J. M., Sánchez-Pérez, A., Calo-Mata, P., & Villa, T. G. (2017). Antimicrobial peptides (AMPs): An- cient compounds that represent novel weapons in the fight against bacteria. Biochemical pharmacol- ogy, 133, 117–138. HTTPS://DOI.ORG/10.1016/J.BCP.2016.09.018 [2] Allen, P., Borick, J., & Borick, J. (2020). Acute and Chronic Infection Management in CF. Cystic Fibrosis in Primary Care, 69–87. HTTPS://DOI.ORG/10.1007/978-3-030-25909-9_8 [3] Breiman, L. (2001). Random Forests. Machine Learn- ing 45, 5–32. HTTPS://DOI.ORG/10.1023/A:1010933404324 [4] Centers for Disease Control and Prevention. (2019). Antibiotic resistant threats in the United States. U.S. Department of Health and Human Services. HTTPS://WWW.CDC.GOV/DRUGRESISTANCE/PDF/THREATS-RE- PORT/2019-AR-THREATS-REPORT-508.PDF [5] Chen, C.H.; Lu, T.K. (2020). Development and Chal- lenges of Antimicrobial Peptides for Therapeutic Ap- plications. Antibiotics, 9, 24. [6] Edwards, I. A., Elliott, A. G., Kavanagh, A. M., Zuegg, J., Blaskovich, M. A., & Cooper, M. A. (2016). Contri- bution of Amphipathicity and Hydrophobicity to the https://doi.org/10.1016/j.bcp.2016.09.018 https://doi.org/10.1007/978-3-030-25909-9_8 https://doi.org/10.1023/A:1010933404324 https://www.cdc.gov/drugresistance/pdf/threats-report/2019-ar-threats-report-508.pdf https://www.cdc.gov/drugresistance/pdf/threats-report/2019-ar-threats-report-508.pdf ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V Antimicrobial Activity and Cytotoxicity of β- Hairpin Peptides. ACS infectious diseases, 2(6), 442–450. HTTPS://DOI.ORG/10.1021/ACSINFECDIS.6B00045 [7] Eisenberg, D., Weiss, R.M., Terwilliger, T.C., & Wilcox, W. (1982). Hydrophobic moments and protein struc- ture. Faraday Symp. Chem. Soc., 17, 109-120. [8] Eisenberg, D., Weiss, R. M., & Terwilliger, T. C. (1984). The hydrophobic moment detects periodicity in pro- tein hydrophobicity. Proceedings of the National Academy of Sciences of the United States of America, 81(1), 140–144. HTTPS://DOI.ORG/10.1073/PNAS.81.1.140 [9] Fauchere, J., Pliska, V (1983). Hydrophobic parame- ters p of amino-acid side chains from the partitioning of N-acetyl-amino-acid amides. Eur. J. Med. Chem, 8, 369–375. [10] Hincapié, O., Giraldo, P., & Orduz, S. (2018). In silico design of polycationic antimicrobial peptides active against Pseudomonas aeruginosa and Staphylococ- cus aureus. Antonie van Leeuwenhoek, 111(10), 1871–1882. HTTPS://DOI.ORG/10.1007/S10482-018-1080-2 [11] Jiang, Z., Vasil, A. I., Hale, J. D., Hancock, R. E., Vasil, M. L., & Hodges, R. S. (2008). Effects of net charge and the number of positively charged residues on the bi- ological activity of amphipathic alpha-helical cationic antimicrobial peptides. Biopolymers, 90(3), 369–383. HTTPS://DOI.ORG/10.1002/BIP.20911 [12] Jolliffe I.T. and Cadima J. (2016). Principal compo- nent analysis: a review and recent developments. Phil. Trans. R. Soc. A.3742015020220150202. HTTPS://DOI.ORG/10.1098/RSTA.2015.0202 [13] Krzywinski, M., Altman, N. (2017). Classification and regression trees. Nat Methods 14, 757–758. HTTPS://DOI.ORG/10.1038/NMETH.4370 [14] Lei, J., Sun, L., Huang, S., Zhu, C., Li, P., He, J., Mackey, V., Coy, D. H., & He, Q. (2019). The antimicrobial pep- tides and their potential clinical applications. Ameri- can journal of translational research, 11(7), 3919– 3931. [15] Li, S., Wang, Y., Xue, Z., Jia, Y., Li, R., He, C., & Chen, H. (2021). The structure-mechanism relationship and mode of actions of antimicrobial peptides: A review. Trends in Food Science & Technology, 109, 103-115. DOI:10.1016/J.TIFS.2021.01.005 [16] Moreau-Marquis, S., Stanton, B. A., & O’Toole, G. A. (2008). Pseudomonas aeruginosa biofilm formation in the cystic fibrosis airway. Pulmonary Pharmacology & Therapeutics, 21(4), 595-599. HTTPS://DOI.ORG/10.1016/J.PUPT.2007.12.001 [17] Pace, C.N., Scholtz, J.M. (1998). A helix propensity scale based on experimental studies of peptides and proteins. Biophysical Journal 75, 422-427. HTTPS://DOI.ORG/10.1016/S0006-3495(98)77529-0 [18] Pirtskhalava, M., Amstrong, A. A., Grigolava, M., Chubinidze, M., Alimbarashvili, E., Vishnepolsky, B., Gabrielian, A., Rosenthal, A., Hurt, D. E., & Tartakov- sky, M. (2021). DBAASP v3: database of antimicro- bial/cytotoxic activity and structure of peptides as a resource for development of new therapeutics. Nu- cleic acids research, 49(D1), D288–D297. HTTPS://DOI.ORG/10.1093/NAR/GKAA991 [19] Porto, W.F., Irazazabal, L., Alves, E.S.F. et al. (2018). In silico optimization of a guava antimicrobial peptide enables combinatorial exploration for peptide de- sign. Nat Commun 9, 1490. doi: HTTPS://DOI.ORG/10.1038/S41467-018-03746-3 [20] Ramazi, S., Mohammadi, N., Allahverdi, A., Khalili, E., & Abdolmaleki, P. (2022). A review on antimicrobial peptides databases and the computational tools. Da- tabase: the journal of biological databases and cura- tion, baac011. HTTPS://DOI.ORG/10.1093/DATABASE/BAAC011 [21] Ratner, B. (2009). The correlation coefficient: Its val- ues range between +1/−1, or do they?. J Target Meas Anal Mark 17, 139–142. HTTPS://DOI.ORG/10.1057/JT.2009.5 [22] Shi, Z., Olson, C. A., & Kallenbach, N. R. (2002). Cat- ion-pi interaction in model alpha-helical peptides. Journal of the American Chemical Society, 124(13), 3284–3291. HTTPS://DOI.ORG/10.1021/JA0174938 [23] Thomas, S., Karnik, S., Barai, R. S., Jayaraman, V. K., & Idicula-Thomas, S. (2010). CAMP: a useful resource for research on antimicrobial peptides. Nucleic acids research, 38(Database issue), D774–D780. HTTPS://DOI.ORG/10.1093/NAR/GKP1021 [24] Timmons, P. B., & Hewage, C. M. (2020). HAPPENN is a novel tool for hemolytic activity prediction for ther- apeutic peptides which employs neural networks. Scientific reports, 10(1), 10869. HTTPS://DOI.ORG/10.1038/S41598-020-67701-3 [25] Ye, Z., & Aparicio, C. (2019). Modulation of supramo- lecular self-assembly of an antimicrobial designer peptide by single amino acid substitution: Implica- tions on peptide activity. Nanoscale advances, 1(12), 4679–4682. HTTPS://DOI.ORG/10.1039/C9NA00498J https://doi.org/10.1021/acsinfecdis.6b00045 https://doi.org/10.1073/pnas.81.1.140 https://doi.org/10.1007/s10482-018-1080-2 https://doi.org/10.1002/bip.20911 https://doi.org/10.1098/rsta.2015.0202 https://doi.org/10.1038/nmeth.4370 https://doi.org/10.1016/j.pupt.2007.12.001 https://doi.org/10.1016/S0006-3495(98)77529-0 https://doi.org/10.1093/nar/gkaa991 https://doi.org/10.1038/s41467-018-03746-3 https://doi.org/10.1093/database/baac011 https://doi.org/10.1057/jt.2009.5 https://doi.org/10.1021/ja0174938 https://doi.org/10.1093/nar/gkp1021 https://doi.org/10.1038/s41598-020-67701-3 https://doi.org/10.1039/c9na00498j ARESTY RUTGERS UNDERGRADUATE RESEARCH JOURNAL, VOLUME I, ISSUE V Alisha Zhu is a student in the Honors College at Rutgers University – New Bruns- wick. She is majoring in Biomedical Engineering and minoring in Statistics. Over the past three years, she has been working as an undergraduate re- searcher in Dr. Charles Roth’s Lab, which she first joined through the Aresty Research Assistant Program. Her research focuses on using computational sta- tistics and machine learning techniques to study the relationship between the biophysical properties of antimicrobial peptides and their drug efficacies. Out- side of research, she is the President of the Society of Women Engineers (SWE) and is the External Vice President of the Chinese Student Organization (CSO). Contact: ALISHA.ZHU@RUTGERS.EDU