De Novo Design and In Silico Validation of a Cationic Antimicrobial Peptide Using an AI-Guided Framework for Membrane Thermodynamics and Hemolytic Toxicity

De Novo Design and In Silico Validation of a Cationic Antimicrobial Peptide Using an AI-Guided Framework for Membrane Thermodynamics and Hemolytic Toxicity | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article De Novo Design and In Silico Validation of a Cationic Antimicrobial Peptide Using an AI-Guided Framework for Membrane Thermodynamics and Hemolytic Toxicity Mateen Ur Rehman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9615735/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The growing epidemic of antimicrobial resistance (AMR) due to multidrug-resistant (MDR) Gram-negative pathogens like Escherichia coli challenges traditional antibiotic treatment with fast evolutionary shifts and declining drug development pipelines. The use of antimicrobial peptides (AMPs) provides a potentially game-changing paradigm shift, using innate cationic-amphipathic structures to permeabilize bacterial membranes through either toroidal pores, carpet mechanisms, or barrel-stave models- multifaceted effects resistant to single-target mechanisms.Although broad-spectrum potent, AMPs encounter translational challenges: host cytotoxicity, serum instability, immunogenicity, and synthetic cost, highlighting the need to be able to engineer with precision de novo. Recent advances in AI-enhanced computational biology presently enable rational optimization of AMP, combining sequence-to-structure prediction (AlphaFold2/ColabFold), physicochemical profiling (modlAMP/Biopython), interfacial energetics (Wimley-White scales), and toxicity prediction by machine learning. A new 16-mer cationic AMP (WKKIWKDPGIKKWIKR) by template-guided design, with a proline-induced hinge in a Trp/Lys scaffold that is modified with C-terminal Arg to enable electrostatic selectivity against Gram-negative. Its efficacy and safety in targeting the E. coli membrane is confirmed by comprehensive in silico studies of helical amphipathicity (high segmental µH), topology (is an 3D α-helix core via ColabFold/ChimeraX), membrane partitioning (ΔG vs. LL-37/magainin benchmarks), and hemolytic risk (Random Forest + A net charge (+ 6 ) and Boman index and instability measures also confirm stability and binding potential. This study proves a complete AI-computational pipeline of AMP discovery that is faster to discover viable therapeutics against MDR threats by reducing wet-lab iterations. Our framework combines biophysics, structural modeling, and explainable ML to become a blueprint to the next generation of antimicrobials. Bioinformatics E. Coli AMP De Novo Alphfold/Colab Fold XAI MDR Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1.0 Introduction Antimicrobial resistance (AMR) is an urgent issue in the global health agenda, especially considering the growing number of multidrug-resistant (MDR) Gram-negative organisms like Escherichia coli [ 1 ]. The traditional antibiotics are gradually losing their effectiveness with the extensive abuse, genetic adaptation of the bacteria and the lack of discovery of new types of antimicrobial agents [ 2 , 3 ]. This pattern has created an urgent need to research other therapeutic approaches that can overcome the resistance mechanisms without compromising safety and specificity [ 4 ]. Antimicrobial peptides (AMPs), a group of short, naturally occurring or synthetic-designed peptides, with broad-spectrum antimicrobial activity and a lower likelihood of triggering resistance are among the most promising candidates [ 5 ]. AMPs usually work via membrane-targeting processes by exploiting their cationic and amphipathic characteristics to react with negatively charged bacterial membranes [ 6 , 7 ]. This communication usually results in membrane disruption, pore formation or intracellular targeting and eventual cell death in bacteria [ 8 , 9 ]. The AMPs have multifaceted and rapid mechanisms of action as compared to the traditional antibiotics, which are specific to certain metabolic pathways, thus less prone to the development of resistance [ 10 ]. Nevertheless, their clinical translation has been hampered by certain issues like toxicity, stability, and cost of production, thus the necessity to adopt rational design strategy to maximize their therapeutic use [ 11 – 14 ]. Recent breakthroughs in computational biology and artificial intelligence (AI) have introduced possibilities of de novo design and optimization of AMPs [ 15 ]. With the combination of sequence-based descriptors, structural prediction software and machine learning models, it has become possible to design peptides with customized physicochemical and functional characteristics [ 16 , 17 ]. Using these methods, it is possible to predict the antimicrobial efficacy, the potential of membrane interaction, and toxicity profiles before experimental validation, which will greatly shorten the drug discovery pipeline [ 18 , 19 ]. de novo design and in silico validation of a new cationic antimicrobial peptide against E. coli is the subject of the current research [ 20 ]. A rational, template-directed strategy was employed to prepare the peptide in order to optimize certain critical properties including net positive charge, hydrophobicity, and amphipathicity-features that are crucial to membrane activity [21,]. Advanced computational methods were used in the analysis of physicochemical properties, conformational structure, membrane partitioning energetics and hemolytic toxicity [ 22 ]. Also, machine learning models were used to forecast safety profiles and offer mechanistic understanding of peptides [ 23 ]. This study seeks to add to the next generation of antimicrobial therapeutics by integrating principles of peptide chemistry, membrane biophysics, and AI-driven analysis. The results do not only indicate the practicability of computational AMP design, but also offer a scalable platform of developing viable and secure alternatives to the traditional antibiotics in combating antimicrobial resistance. 2.0 Methodology 2.1 Sequence Design and Candidate Selection A novel antimicrobial peptide (AMP) candidate was designed de novo using a template-guided rational design strategy. The initial scaffold was a 15-residue peptide (WKKIWKDPGIKKWIK), selected based on its enrichment in lysine (K) and tryptophan (W) residues, a compositional pattern commonly associated with membrane-active AMP scaffolds and STAMP-like motifs. This design strategy aimed to optimize cationic charge and hydrophobicity, key determinants for selective interaction with negatively charged bacterial membranes, particularly in Gram-negative organisms such as Escherichia coli . A central proline residue was retained to introduce a conformational disruption within the peptide backbone. This proline-induced hinge was incorporated to promote structural flexibility and enable segmental adaptation during membrane interaction. To further enhance electrostatic attraction toward anionic membrane components, a terminal arginine residue was appended to the C-terminus, resulting in the final 16-residue sequence: WKKIWKDPGIKKWIKR. The designed peptide was treated as a computational candidate and was not derived from a naturally occurring sequence. Sequence composition and motif design were guided by established structure–activity relationships reported for antimicrobial peptides. 2.2 Physicochemical Property Calculation Physicochemical properties of the designed peptide (WKKIWKDPGIKKWIKR) were calculated using the modlAMP and Biopython libraries in a Python-based environment. The peptide sequence was provided as input and analyzed using the GlobalDescriptor module of modlAMP to compute key parameters relevant to antimicrobial activity. The calculated molecular weight of the peptide was approximately 2110.6 Da, consistent with the typical size range of short antimicrobial peptides (< 3000 Da). The net charge at physiological pH (7.4) was determined to be + 5.99, reflecting a strongly cationic nature favorable for interaction with negatively charged bacterial membranes. The isoelectric point (pI) was calculated as 11.31, further confirming the peptide’s basic character. The Boman index, which estimates protein-binding potential and membrane interaction capability, was calculated to be 2.141, indicating moderate to high affinity for lipid interfaces. These physicochemical properties collectively suggest that the designed peptide possesses characteristics consistent with membrane-active antimicrobial peptides. 2.3 Amphipathicity Analysis The amphipathic character of the designed peptide was evaluated using hydrophobic moment (µH), calculated based on the Eisenberg consensus hydrophobicity scale using the modlAMP Python library. Hydrophobic moment quantifies the spatial segregation of hydrophobic and hydrophilic residues along a helical axis and is widely used to assess membrane-active potential in antimicrobial peptides. Hydrophobic moment was computed across the full peptide sequence (WKKIWKDPGIKKWIKR) using a large window size (window = 1000) to capture global amphipathic behavior. The calculated global hydrophobic moment (µH ≈ 0.143) indicated a low apparent amphipathicity when averaged across the entire sequence. To account for structural discontinuity introduced by the central proline residue, a segmental analysis was performed by dividing the sequence into N-terminal (WKKIWKD) and C-terminal (GIKKWIKR) regions. Hydrophobic moment was calculated independently for each segment using the same Eisenberg-based approach. The N-terminal segment exhibited a hydrophobic moment of approximately µH ≈ 0.815, while the C-terminal segment showed µH ≈ 0.858, both values exceeding the typical threshold (> 0.4) associated with amphipathic α-helices. This segmental disparity suggests that the peptide does not behave as a single continuous amphipathic helix but rather as two amphipathic subdomains separated by a proline-induced structural hinge. This configuration supports a flexible, hinge-driven topology commonly observed in membrane-active antimicrobial peptides. 2.4 Secondary Structure Visualization Helical wheel projections were generated to visualize residue distribution around the helical axis and to qualitatively evaluate amphipathic patterning. Helical wheel projections were generated using the modlAMP helical wheel function. These plots were used to identify whether hydrophobic residues such as tryptophan and isoleucine clustered on one face of the helix, while cationic or polar residues such as lysine, arginine, and aspartate occupied the opposite face. This visualization was particularly useful for interpreting the structural consequences of the central proline residue, which was expected to induce a local bend or kink. The resulting topology was evaluated in the context of common AMP architectures, including hinge-containing and boomerang-like membrane-active peptides. 2.5 Three-Dimensional Structure Prediction Protein structure prediction was performed using ColabFold (v1.6.1), which integrates the deep learning–based AlphaFold2 with rapid multiple sequence alignment (MSA) generation via MMseqs2. The peptide sequence (WKKIWKDPGIKKWIKR) was submitted as a single-chain input for monomeric structure prediction. MSAs were generated using the mmseqs2_uniref_env mode, incorporating both UniRef and environmental sequence databases, with a pairing strategy set to unpaired_paired to include both paired and unpaired alignments. Template-based modeling was disabled (template_mode = none), ensuring that predictions relied solely on de novo inference. Model parameters were maintained at default initialized settings, including model_type = auto, which applies AlphaFold2-ptm for monomer prediction, num_recycles = 3, and recycle_early_stop_tolerance = auto, with a greedy pairing strategy and additional calculation of pairwise ipTM/actifpTM scores enabled. Sampling parameters were kept unchanged (max_msa = auto, num_seeds = 1, and dropout disabled). Structural relaxation using AMBER was not performed (num_relax = 0, relax_max_iterations = 200) to reduce computational overhead. All predictions were executed within the ColabFold notebook environment, and resulting models were ranked based on internal confidence metrics, including predicted Local Distance Difference Test (pLDDT) and predicted TM-score (pTM). The top-ranked structure (rank 1) was selected for further analysis and visualization using pLDDT-based coloring. All output files, including predicted structures, alignment data, and confidence scores, were retained with save_all and save_recycles enabled at a resolution of 200 dpi for downstream analysis. 2.6 Structural Visualization in ChimeraX Structural visualization and analysis were performed using UCSF ChimeraX (version 1.11.1, released January 23, 2026). The predicted peptide structure obtained from ColabFold was imported in PDB format and processed within the ChimeraX environment. Secondary structure elements were automatically assigned using built-in algorithms, and the structure was visualized using a cartoon (ribbon) representation to assess overall folding and α-helical content. Residue-level confidence was visualized by coloring the structure according to B-factor values corresponding to AlphaFold-derived pLDDT scores using the “alphafold” color palette. Cartoon thickness was adjusted to enhance structural clarity, and additional stick representations were briefly used to inspect atomic-level interactions before reverting to ribbon visualization. The solvent background was set to white to improve image contrast for publication-quality rendering. Structural analysis focused on identifying helix formation, residue distribution, and conformational features such as the central proline-induced hinge. No structural refinement or energy minimization was performed within ChimeraX. High-resolution images were generated using supersampling to enhance visual quality, and final structures were exported as PNG files for figure preparation. 2.7 Membrane Partitioning Analysis The thermodynamic propensity of the peptide to associate with lipid membranes was estimated using the Wimley–White interfacial hydrophobicity scale. Residue-specific free-energy contributions were assigned based on the peptide sequence, and the total membrane partitioning energy ( \(\:{\Delta\:}G\) ) was calculated by summing the individual values. In addition to residue-level analysis, a smoothed local energy profile was generated using a moving-window average to approximate local helical environments and reduce single-residue noise. The resulting thermodynamic profile was used to estimate the peptide’s membrane insertion potential and compare its behavior with known AMPs and non-AMP control peptides. 2.8 Benchmarking Against Reference Peptides Thermodynamic benchmarking of membrane interaction propensity was performed using the Wimley–White interfacial hydrophobicity scale, which provides experimentally derived free energy values (ΔG, kcal/mol) for amino acid partitioning at membrane interfaces. Residue-specific ΔG values were assigned to each amino acid, and total membrane anchoring energy was calculated by summing all energetically favorable (negative) contributions. For the designed peptide (WKKIWKDPGIKKWIKR), this analysis yielded a total anchoring energy of approximately − 6.48 kcal/mol, indicating a strong intrinsic propensity for membrane insertion driven primarily by hydrophobic residues such as tryptophan (W) and isoleucine (I). To contextualize this value, comparative benchmarking was performed against a panel of experimentally validated antimicrobial peptides (AMPs) and non-antimicrobial human protein fragments. The AMP reference set included melittin (ΔG ≈ − 5.02 kcal/mol), magainin 2 (ΔG ≈ − 4.80 kcal/mol), LL-37 (ΔG ≈ − 7.69 kcal/mol), and indolicidin (ΔG ≈ − 10.12 kcal/mol), representing peptides with well-characterized membrane-disruptive activity. Buforin II (ΔG ≈ − 2.81 kcal/mol) was also included as a lower membrane-active AMP for comparison. The negative control group consisted of fragments from human serum albumin (ΔG ≈ − 3.38 kcal/mol), hemoglobin beta (ΔG ≈ − 2.97 kcal/mol), and actin alpha (ΔG ≈ − 2.49 kcal/mol), which are not expected to spontaneously insert into lipid bilayers. All sequences were analyzed using a custom Python-based implementation of the Wimley–White scale, and resulting ΔG values were compiled into a structured dataset and ranked from most negative (strongest membrane affinity) to least negative. A theoretical membrane insertion threshold of approximately − 3.0 kcal/mol was used as a reference point to distinguish membrane-active peptides from non-interacting or weakly interacting sequences. Comparative visualization was performed using horizontal bar plots, where the designed peptide, reference AMPs, and non-AMP controls were color-coded and displayed according to their respective ΔG values. All computations and visualizations were performed using Python (NumPy, Pandas, and Matplotlib) within a Google Colab environment. 2.9 Toxicity Prediction Using Machine Learning Hemolytic toxicity of the designed peptide was predicted using a supervised machine learning approach implemented in Python. A Random Forest classifier was trained on a curated dataset of 14 experimentally characterized antimicrobial peptides annotated as hemolytic (n = 7) or non-hemolytic (n = 7). The dataset included representative peptides such as melittin, LL-37, mastoparan, aurein 1.2, indolicidin, temporin L, and ovispirin (hemolytic class), as well as magainin 2, buforin II, pexiganan, alyteserin, rana-box, dermaseptin, and esculentin-2 (non-hemolytic class). Each peptide sequence was transformed into a numerical feature vector using sequence-derived physicochemical descriptors. The extracted features included: peptide length, net charge (calculated as the difference between positively charged residues [Lys, Arg] and negatively charged residues [Asp, Glu]), membrane anchoring energy derived from the Wimley–White hydrophobicity scale (sum of negative ΔG contributions), hydrophobic residue ratio (fraction of residues belonging to A, V, I, L, M, F, Y, W), and aromatic residue ratio (fraction of W, F, and Y residues). Feature scaling was performed using standard normalization (StandardScaler) to ensure comparable feature distributions. The Random Forest model was trained using 100 decision trees (n_estimators = 100) with a fixed random seed (random_state = 42) to ensure reproducibility. The trained model was then used to predict the hemolytic potential of the designed peptide (WKKIWKDPGIKKWIKR), yielding a probability score between 0 and 1. A threshold of 0.5 was applied to classify the peptide as hemolytic or non-hemolytic. Feature importance analysis was conducted using Gini impurity-based importance scores derived from the trained Random Forest model, providing insight into the relative contribution of each physicochemical feature to the toxicity prediction. 2.9.1 Explainable AI Analysis Explainable artificial intelligence (XAI) analysis was performed using SHAP (SHapley Additive exPlanations) with the TreeExplainer algorithm, which is specifically optimized for tree-based models such as Random Forest. This approach enabled quantitative decomposition of the model output into feature-level contributions, facilitating mechanistic interpretation of hemolytic toxicity predictions.” A SHAP waterfall plot was used to visualize baseline prediction, feature-level contributions, and final model output. This helped identify the major biophysical drivers of toxicity and supported mechanistic interpretation of the classifier’s prediction. 2.9.2 Computational Environment and Implementation Details All computational analyses were performed in a cloud-based environment using Google Colab, ensuring reproducibility and accessibility of the workflow. The entire computational pipeline was implemented in Python (version 3.10), utilizing open-source libraries for sequence analysis, machine learning, and data visualization. Specifically, Biopython (v1.81) was employed for sequence-based physicochemical calculations and proteolytic digestion analysis, while modlAMP (v4.3) was used for hydrophobic moment (µH) calculations and amphipathicity profiling. Numerical computations and data handling were conducted using NumPy (v1.26) and Pandas (v2.0), and graphical representations were generated using Matplotlib (v3.7) and Seaborn (v0.12). Machine learning implementation was performed using scikit-learn (v1.3), and model interpretability was achieved using SHAP (v0.43) for feature attribution analysis. Structural predictions were carried out using ColabFold, while molecular visualization and structural analysis were conducted using ChimeraX. All computations were executed on Google Colab’s GPU-enabled infrastructure, utilizing NVIDIA Tesla T4 or V100 GPUs depending on session availability, with approximately 12–16 GB of RAM allocated for runtime operations. CPU-based execution was used for lightweight preprocessing and feature extraction steps. For machine learning analysis, a Random Forest classifier was implemented using scikit-learn with 100 estimators, Gini impurity as the splitting criterion, and default unrestricted tree depth. The dataset was partitioned into training and testing sets using an 80:20 split ratio. Feature vectors were constructed from sequence-derived descriptors, including peptide length, net charge, hydrophobic residue ratio, aromatic residue ratio, and membrane anchoring energy (ΔG). Model performance was evaluated using accuracy metrics, and feature importance was derived from Gini-based impurity reduction. To enhance interpretability, SHAP (SHapley Additive exPlanations) analysis was applied using the TreeExplainer method optimized for tree-based models. SHAP values were computed to quantify the contribution of individual features to the predicted hemolytic toxicity of the peptide. Waterfall plots were generated to visualize the baseline prediction, feature-wise contributions, and final model output. Positive SHAP values indicated features contributing to increased toxicity, whereas negative values indicated protective or stabilizing effects, enabling precise identification of structural determinants influencing hemolytic potential. All computational steps, including sequence design, structural prediction, physicochemical analysis, thermodynamic modeling, and machine learning, were integrated into a unified workflow within the Colab environment. Random seeds were fixed where applicable to ensure reproducibility of machine learning outputs. All scripts, intermediate data, and generated outputs (e.g., .pdb files, feature matrices, and plots) were retained within the runtime environment and are available for export to support external validation and reproducibility. 3.0 Results 3.1 Physicochemical Profiling and Amphipathic Character The rationally designed 16-mer candidate peptide (WKKIWKDPGIKKWIKR) was first evaluated for its foundational physicochemical properties. Sequence analysis confirmed a highly cationic nature, yielding a net charge of + 6 at physiological pH (7.4), driven by the enrichment of lysine (K) and the C-terminal arginine (R) residues. This pronounced positive charge provides a strong electrostatic driving force for initial attraction to the anionic lipopolysaccharides (LPS) of the E. coli outer membrane. Calculations of the Boman index indicated a high potential for protein and membrane interaction, typical of active AMPs, while the instability index suggested that the peptide maintains structural viability within biological parameters. Furthermore, helical wheel projections and calculations of the hydrophobic moment (µH) confirmed a strong amphipathic character. The central proline residue (P8) effectively partitioned the sequence, generating distinct segmental amphipathicity. The strategic placement of hydrophobic residues (Tryptophan and Isoleucine) on one face and cationic residues on the opposing face established an optimal topological profile for membrane insertion. 3.2 Three-Dimensional Structure and Topology Structural modeling was performed using ColabFold under the parameters described in Section 2.5 , providing high-confidence predictions of the peptide’s conformational tendencies. Visualized using ChimeraX, the predicted 3D structure of the designed antimicrobial peptide exhibits a dominant alpha-helical core accompanied by a flexible terminal region (Fig. 1 ). This structural architecture aligns well with known membrane-active AMPs. The alpha-helical core facilitates deep insertion into the hydrophobic tail region of the bacterial lipid bilayer, while the central proline acts as a structural hinge. This localized flexibility likely allows the peptide to dynamically adapt its conformation upon encountering the membrane interface, maximizing contact area and promoting disruptive mechanisms such as toroidal pore or carpet-like formation. The predicted structure shows a mostly alpha-helical antimicrobial peptide with a flexible looped tail. This type of fold is common in AMPs because the helix can help the peptide interact with the negatively charged bacterial membrane, while the flexible region may help it adapt during binding 3.3 Membrane Partitioning Thermodynamics and Residue-Level Energy Profile The membrane interaction propensity of the designed antimicrobial peptide (WKKIWKDPGIKKWIKR) was evaluated using the Wimley–White interfacial hydrophobicity scale to estimate residue-specific free energies of partitioning (ΔG). The resulting thermodynamic profile (Fig. 2 ) provides insight into the energetic landscape governing peptide–membrane interactions. The analysis revealed a distinct amphipathic energetic pattern, characterized by alternating favorable and unfavorable contributions along the peptide sequence. Hydrophobic residues, particularly tryptophan (W) and isoleucine (I), exhibited strongly negative ΔG values (reaching approximately − 1.8 kcal/mol), indicating a high propensity for spontaneous insertion into the lipid bilayer. These residues are likely to function as membrane-anchoring elements, facilitating penetration into the hydrophobic core of the membrane. In contrast, cationic residues such as lysine (K) and arginine (R) displayed positive ΔG values, reflecting energetically unfavorable insertion into the membrane interior. However, this is consistent with their expected role in electrostatic interactions with negatively charged lipid headgroups, particularly in Gram-negative bacterial membranes. This complementary distribution of hydrophobic and charged residues supports a functionally optimized amphipathic architecture. The smoothed local energy profile further highlighted cooperative membrane-binding regions, with favorable segments concentrated in the N-terminal and central regions of the peptide. Notably, the central proline residue appears to introduce a disruption in the energetic continuity, supporting its proposed role as a structural hinge that may facilitate conformational adaptability during membrane interaction Residue-level free energy of partitioning (ΔG, kcal/mol) for WKKIWKDPGIKKWIKR calculated using the Wimley–White interfacial hydrophobicity scale. Bars represent individual residue contributions, where negative ΔG (blue) indicates favorable membrane insertion and positive ΔG (red) indicates unfavorable insertion. The black line shows the smoothed local membrane affinity profile (moving average). Hydrophobic residues (W, I) display strongly favorable energies, consistent with membrane anchoring, while cationic residues (K, R) show unfavorable insertion but contribute to interfacial electrostatic interactions. The overall pattern reflects a segmented amphipathic profile characteristic of membrane-active peptides. 3.4 Thermodynamic Benchmarking Against Reference Peptides The membrane partitioning behavior of the designed peptide was further evaluated through comparative thermodynamic benchmarking against a panel of experimentally characterized antimicrobial peptides (AMPs) and non-antimicrobial human protein fragments (Fig. 3 ). Total membrane anchoring energies (ΔG) were calculated using the Wimley–White interfacial hydrophobicity scale and used as a quantitative measure of membrane affinity. The results demonstrate that the designed peptide exhibits a strongly favorable total free energy of membrane partitioning (ΔG ≈ − 6 kcal/mol), positioning it well within the range of known membrane-active AMPs. Notably, its thermodynamic profile is comparable to established peptides such as magainin 2 and melittin, both of which are known for their effective membrane-disruptive properties. While the designed peptide shows slightly less favorable energetics than highly potent AMPs such as indolicidin and LL-37, it remains significantly more membrane-active than non-AMP controls. In contrast, protein fragments derived from human serum albumin, hemoglobin beta, and actin alpha exhibited relatively weak or near-neutral partitioning energies (ΔG > − 3 kcal/mol), indicating limited intrinsic membrane affinity. This clear separation between AMP and non-AMP groups supports the validity of the thermodynamic framework used in this study. Importantly, the designed peptide falls well below the empirical threshold (ΔG ≈ − 3 kcal/mol) associated with effective membrane insertion, reinforcing its classification as a membrane-active candidate. At the same time, its intermediate positioning relative to highly cytotoxic peptides such as melittin suggests a potential balance between antimicrobial efficacy and reduced host toxicity, which is a critical consideration in AMP design. Comparison of total membrane anchoring energies (ΔG, kcal/mol) calculated using the Wimley–White interfacial hydrophobicity scale for the designed peptide and a panel of reference sequences. The dataset includes known antimicrobial peptides (blue), non-antimicrobial human protein fragments (gray), and the designed peptide (red). More negative ΔG values indicate stronger membrane affinity and a higher propensity for membrane insertion. 3.5 Biophysical Determinants of Predicted Hemolytic Toxicity To identify the key physicochemical features governing the predicted hemolytic toxicity of the designed peptide, a Random Forest classifier was employed and feature importance was evaluated using Gini impurity-based metrics (Fig. 4 ). The analysis revealed that peptide length is the most influential factor, contributing the highest importance (~ 0.28) to the model’s decision. This suggests that peptide size plays a critical role in modulating toxicity, likely by influencing membrane coverage, pore formation potential, and overall structural stability. The second most significant contributor was membrane anchoring energy (ΔG) (~ 0.21), highlighting the importance of thermodynamic membrane affinity in determining cytotoxic behavior. Peptides with stronger membrane insertion capabilities are more likely to disrupt not only bacterial membranes but also host cell membranes, thereby increasing hemolytic risk. Aromatic residue content (~ 0.19), primarily driven by tryptophan residues, also showed a strong contribution. Aromatic residues are known to facilitate membrane insertion by interacting with lipid headgroup regions, which can enhance both antimicrobial activity and toxicity. Similarly, the hydrophobic residue ratio (~ 0.18) was identified as a key determinant, reflecting the role of hydrophobic interactions in membrane permeabilization. In contrast, net charge (~ 0.14) exhibited the lowest relative importance among the evaluated features. While cationic charge is essential for selective binding to negatively charged bacterial membranes, this result suggests that charge alone is not the primary driver of toxicity, but rather acts in combination with hydrophobic and structural properties. Feature importance analysis derived from a Random Forest classifier illustrating the relative contribution of key physicochemical properties to the prediction of hemolytic toxicity. Importance scores are based on Gini impurity reduction, where higher values indicate greater influence on the model’s decision. Peptide length emerged as the most influential feature, followed by membrane anchoring energy (ΔG), highlighting the importance of structural size and membrane affinity in toxicity prediction. Aromatic residue ratio and hydrophobic residue ratio also contributed substantially, reflecting the role of membrane-interacting residues in modulating cytotoxic effects. In contrast, net charge showed comparatively lower importance, suggesting that electrostatic interactions alone are insufficient to explain hemolytic behavior. 3.6 Explainable AI Analysis of Hemolytic Toxicity (SHAP Interpretation) To further interpret the machine learning-based toxicity prediction, SHAP (SHapley Additive exPlanations) analysis was applied to quantify the contribution of individual physicochemical features to the predicted hemolytic risk of the designed peptide (Fig. 5 ). The model yielded a predicted toxicity score of f(x) = 0.54, relative to a baseline expectation of E[f(x)] = 0.486, indicating a moderate but controlled hemolytic potential. Feature-level contributions revealed a nuanced interplay between factors that increase and decrease toxicity. Among the positive contributors, the aromatic residue ratio exhibited the strongest effect (+ 0.08), suggesting that the presence of tryptophan-rich regions enhances membrane interaction and contributes to increased cytotoxicity. Similarly, membrane anchoring energy (ΔG) contributed positively (+ 0.05), indicating that stronger membrane affinity is associated with a higher likelihood of disrupting host cell membranes. In contrast, several features acted to reduce predicted toxicity. The hydrophobic residue ratio (− 0.04) and peptide length (− 0.04) both contributed negatively, suggesting that the specific balance of hydrophobic content and the relatively short peptide length may help limit excessive membrane disruption. These findings imply that while hydrophobicity is necessary for antimicrobial function, its controlled distribution plays a protective role against toxicity. The net charge showed only a minimal positive contribution (+ 0.01), reinforcing earlier observations that electrostatic interactions alone are not the primary drivers of hemolytic activity, but instead function in coordination with other structural and thermodynamic factors. SHAP (SHapley Additive exPlanations) analysis showing feature-wise contributions to the predicted hemolytic toxicity of the designed peptide. The baseline prediction (E[f(x)] = 0.486) is shifted to the final output (f(x) = 0.54) by individual feature contributions. Positive values (red) increase predicted toxicity, while negative values (blue) decrease it. The aromatic residue ratio is the dominant positive contributor, followed by membrane anchoring energy (ΔG), indicating that increased aromatic content and membrane affinity elevate toxicity. In contrast, hydrophobic residue ratio and peptide length contribute negatively, moderating the prediction. Net charge shows a minimal positive effect, suggesting a limited role in toxicity determination. 4.0 Discussion The present study demonstrates a rational and integrative framework for the de novo design of antimicrobial peptides by combining biophysical modeling with explainable machine learning. The designed peptide (WKKIWKDPGIKKWIKR) exhibits key hallmarks of membrane-active AMPs, including a strong cationic charge, pronounced amphipathicity, and a structurally adaptive topology driven by a central proline hinge. Thermodynamic profiling revealed a favorable membrane partitioning landscape, with hydrophobic residues driving insertion and cationic residues stabilizing interfacial interactions [ 24 ]. The residue-level ΔG distribution and smoothed energy profile support a cooperative and segmented membrane interaction mechanism, consistent with known AMP behaviors such as toroidal pore formation or carpet-like disruption [ 25 ]. Importantly, benchmarking against experimentally validated peptides positioned the candidate within the functional AMP range, with a total anchoring energy comparable to magainin 2 and melittin, yet less extreme than highly cytotoxic peptides such as indolicidin [ 26 ]. Machine learning analysis further provided insight into the determinants of hemolytic toxicity. Feature importance indicated that toxicity is governed by a combination of peptide length, membrane affinity, and aromatic content rather than charge alone [ 27 ]. SHAP-based interpretation refined this understanding by demonstrating that toxicity-enhancing features, such as aromatic residue enrichment and strong membrane anchoring, are partially offset by stabilizing factors including controlled hydrophobicity and optimized peptide length [ 28 ]. Together, these findings highlight the importance of achieving a balanced biophysical profile. The designed peptide appears to occupy a favorable intermediate space, combining effective membrane activity with moderated toxicity, thereby supporting its potential as a promising candidate for further experimental validation against multidrug-resistant pathogens. 5.0 Conclusion This study is a comprehensive AI-guided framework for the de novo design and in silico validation of antimicrobial peptides targeting multidrug-resistant pathogens. The designed 16-mer peptide (WKKIWKDPGIKKWIKR) highlights a favorable combination of physicochemical and biophysical properties, including strong cationic charge, amphipathic organization, and a structurally adaptive topology facilitated by a central proline hinge. Thermodynamic analysis confirmed a favorable membrane partitioning profile, with residue-level contributions supporting efficient membrane insertion and interaction. Comparative benchmarking against known antimicrobial peptides further validated that the designed sequence falls within the functional range of membrane-active peptides while maintaining a more balanced energetic profile than highly cytotoxic counterparts. Importantly, machine learning-based toxicity prediction, complemented by SHAP interpretability, revealed that hemolytic risk is governed by a combination of membrane affinity, aromatic content, and structural features. The designed peptide achieves a controlled balance between these factors, suggesting the potential to retain antimicrobial efficacy while minimizing host toxicity. References Salam MA, Al-Amin MY, Salam MT, Pawar JS, Akhter N, Rabaan AA, Alqumber MA (2023), July Antimicrobial resistance: a growing serious threat for global public health. In Healthcare (Vol. 11, No. 13, p. 1946). MDPI Muteeb G, Rehman MT, Shahwan M, Aatif M (2023) Origin of antibiotics and antibiotic resistance, and their impacts on drug development: A narrative review. Pharmaceuticals 16(11):1615 Schneider YK (2021) Bacterial natural product drug discovery for new antibiotics: strategies for tackling the problem of antibiotic resistance by efficient bioprospecting. Antibiotics 10(7):842 MacNair CR, Rutherford ST, Tan MW (2024) Alternative therapeutic strategies to treat antibiotic-resistant pathogens. Nat Rev Microbiol 22(5):262–275 Mookherjee N, Anderson MA, Haagsman HP, Davidson DJ (2020) Antimicrobial host defence peptides: functions and clinical potential. Nat Rev Drug Discovery 19(5):311–332 Huang X, Li G (2023) Antimicrobial peptides and cell-penetrating peptides: non-antibiotic membrane-targeting strategies against bacterial infections. Infect Drug Resist, 1203–1219 Pirtskhalava M, Vishnepolsky B, Grigolava M, Managadze G (2021) Physicochemical Features and Peculiarities of Interaction of AMP with the Membrane. Pharmaceuticals 14(5):471 Duong L, Gross SP, Siryaporn A (2021) Developing antimicrobial synergy with AMPs. Front Med Technol 3:640981 Seyfi R, Kahaki FA, Ebrahimi T, Montazersaheb S, Eyvazi S, Babaeipour V, Tarhriz V (2020) Antimicrobial peptides (AMPs): roles, functions and mechanism of action. Int J Pept Res Ther 26(3):1451–1463 Oliveira Júnior NG, Souza CM, Buccini DF, Cardoso MH, Franco OL (2025) Antimicrobial peptides: structure, functions and translational applications. Nat Rev Microbiol 23(11):687–700 Taheri-Araghi S (2024) Synergistic action of antimicrobial peptides and antibiotics: current understanding and future directions. Front Microbiol 15:1390765 Zhang C, Yang M (2022) Antimicrobial peptides: from design to clinical application. Antibiotics 11(3):349 Zheng B, Wang X, Guo M, Tzeng CM (2025) Therapeutic peptides: recent advances in discovery, synthesis, and clinical translation. Int J Mol Sci 26(11):5131 Costa F, Teixeira C, Gomes P, Martins MCL (2019) Clinical application of AMPs. Antimicrobial peptides: basics for clinical application. Springer Singapore, Singapore, pp 281–298 Meng H (2025) AI-driven discovery and design of antimicrobial peptides: Progress, challenges, and opportunities. Probiotics Antimicrob Proteins, 1–23 Chen X, Li C, Bernards MT, Shi Y, Shao Q, He Y (2021) Sequence-based peptide identification, generation, and property prediction with deep learning: a review. Mol Syst Des Eng 6(6):406–428 Choudhury PR, Mishra SK, Yadav S, Singh S, Mathur P (2025) In Silico Peptide Design: Methods, Resources, and Role of AI. J Pept Sci, 31(12), e70063 Brown ED, Wright GD (2005) New targets and screening approaches in antimicrobial drug discovery. Chem Rev 105(2):759–774 Amorim AM, Piochi LF, Gaspar AT, Preto AJ, Rosario-Ferreira N, Moreira IS (2024) Advancing drug safety in drug development: bridging computational predictions for enhanced toxicity prediction. Chem Res Toxicol 37(6):827–849 Lin S, Wade JD, Liu S (2020) De novo design of flavonoid-based mimetics of cationic antimicrobial peptides: Discovery, development, and applications. Acc Chem Res 54(1):104–119 Yuhao Z, Chen Y, Jiecheng L, Yan L, Yufen Z (2025) Prebiotic template-directed peptide formation mediated by polyphosphates and peptide promoters. Org Biomol Chem 23(36):8205–8211 Robles-Loaiza, A. A., Pinos-Tamayo, E. A., Mendes, B., Ortega-Pila, J. A., Proaño-Bolaños,C., Plisson, F., … Almeida, J. R. (2022). Traditional and computational screening of non-toxic peptides and approaches to improving selectivity. Pharmaceuticals, 15(3), 323. Soto-Garcia N, Davari MD, Medina‐Ortiz D (2026) Machine Learning to Accelerate the Discovery of Therapeutic Peptides. Mach Learn Big Data‐enabled Biotechnol, 183–217 Zhang N, Dai S, Qi R, Han Y (2025) Thermodynamic studies on partition of glutamic amphiphiles into phospholipid membranes with different compositions. Colloids Surf A 708:135953 Kyriazis V, Gadhe L, Konstantoulea K, Louros N, Schymkowitz J, Rousseau F (2025) Amyloid Explorer: a global atlas of amyloid fibril structures and thermodynamic principles. bioRxiv, 2025–2010 Oliveira Júnior NG, Souza CM, Buccini DF, Cardoso MH, Franco OL (2025) Antimicrobial peptides: structure, functions and translational applications. Nat Rev Microbiol 23(11):687–700 Bhatnagar P, Khandelwal Y, Mishra S, Dutta A, Mitra D, Biswas S (2024) Predicting antibacterial activity, efficacy, and hemotoxicity of peptides using an explainable machine learning framework. Process Biochem 145:163–174 Yogarathinam LT, Abba SI, Usman J, Ramamoorthy M, Aljundi IH (2025) Interpretable SHAP-based machine learning-assisted design for selecting ultrafiltration membranes in protein-laden phosphate wastewater. Clean Chem Eng 11:100187 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9615735","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":634624363,"identity":"9fcb434a-e9f1-4c96-99c4-c8b52d5bbc41","order_by":0,"name":"Mateen Ur Rehman","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9ElEQVRIiWNgGAWjYDADAyBm/FABJJmZGwioZUZoYZY4AyIZSdDCwNsGIglo0W0/f/DBxx335MzFDj97IDmvNpq/HajlR8U2nFrMziQzG848U2xsOTvN3KBw2/HcGYcZGxh7ztzGreVAMps0b1tC4obbCWYSktuO5TYAtTAztuHRcv4x+++/bQn1G26nf5PgnXMsdz5BLTeS2YAKEhIMbueYSfA21ORuIKzlsbFkb1uC4YbbOWXSEscO5G4EajmI1y/nEx9++NmWIG9wO32b5Ieautx55w8ffPCjArcWdHAYTB4gWj0Q1JGieBSMglEwCkYIAADhR15ebJ0iKAAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0000-4253-6442","institution":"The University of Lahore","correspondingAuthor":true,"prefix":"","firstName":"Mateen","middleName":"Ur","lastName":"Rehman","suffix":""}],"badges":[],"createdAt":"2026-05-05 08:21:10","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9615735/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9615735/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108570074,"identity":"3ec28ec7-0bd0-49e9-bdaf-852c1baf655d","added_by":"auto","created_at":"2026-05-06 06:05:25","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":33943,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredicted 3D structure of the designed antimicrobial peptide showing a dominant alpha-helical core and a flexible terminal region\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/64b9632cfd59a36ece99b1ea.png"},{"id":108804388,"identity":"ed322328-6bf1-4f39-acf9-b73265965bde","added_by":"auto","created_at":"2026-05-08 15:20:05","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79922,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThermodynamic membrane insertion profile of the designed antimicrobial peptide\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/bc933d5601b1c582d0f2d173.png"},{"id":108570076,"identity":"ed702d00-657f-409a-87af-5639c4f15aab","added_by":"auto","created_at":"2026-05-06 06:05:25","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":81823,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThermodynamic benchmarking of the designed peptide against reference antimicrobial and non-antimicrobial peptides.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/9c70c90401606518a75b9233.png"},{"id":108570077,"identity":"c84b6ee8-6235-4c3f-a2be-c4b6ca7d7cab","added_by":"auto","created_at":"2026-05-06 06:05:25","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":42155,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBiophysical determinants of hemolytic toxicity predicted by Random Forest analysis.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/44a92ebb8e60120b01dd0963.png"},{"id":108570078,"identity":"1bc8bb2e-fc19-4e92-a234-e76856cd2cd1","added_by":"auto","created_at":"2026-05-06 06:05:25","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":41294,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP-based interpretation of biophysical drivers of hemolytic toxicity\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/c1320a17cadf2e817a7b5f37.png"},{"id":108811783,"identity":"f8cbf892-100b-450e-8266-0cb93a565bf9","added_by":"auto","created_at":"2026-05-08 16:07:05","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":454665,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9615735/v1/1fe004ad-cf9d-459b-9975-aac83c65ff2a.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eDe Novo Design and In Silico Validation of a Cationic Antimicrobial Peptide Using an AI-Guided Framework for Membrane Thermodynamics and Hemolytic Toxicity\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1.0 Introduction","content":"\u003cp\u003eAntimicrobial resistance (AMR) is an urgent issue in the global health agenda, especially considering the growing number of multidrug-resistant (MDR) Gram-negative organisms like Escherichia coli [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The traditional antibiotics are gradually losing their effectiveness with the extensive abuse, genetic adaptation of the bacteria and the lack of discovery of new types of antimicrobial agents [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. This pattern has created an urgent need to research other therapeutic approaches that can overcome the resistance mechanisms without compromising safety and specificity [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Antimicrobial peptides (AMPs), a group of short, naturally occurring or synthetic-designed peptides, with broad-spectrum antimicrobial activity and a lower likelihood of triggering resistance are among the most promising candidates [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAMPs usually work via membrane-targeting processes by exploiting their cationic and amphipathic characteristics to react with negatively charged bacterial membranes [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. This communication usually results in membrane disruption, pore formation or intracellular targeting and eventual cell death in bacteria [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The AMPs have multifaceted and rapid mechanisms of action as compared to the traditional antibiotics, which are specific to certain metabolic pathways, thus less prone to the development of resistance [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Nevertheless, their clinical translation has been hampered by certain issues like toxicity, stability, and cost of production, thus the necessity to adopt rational design strategy to maximize their therapeutic use [\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecent breakthroughs in computational biology and artificial intelligence (AI) have introduced possibilities of de novo design and optimization of AMPs [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. With the combination of sequence-based descriptors, structural prediction software and machine learning models, it has become possible to design peptides with customized physicochemical and functional characteristics [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Using these methods, it is possible to predict the antimicrobial efficacy, the potential of membrane interaction, and toxicity profiles before experimental validation, which will greatly shorten the drug discovery pipeline [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003ede novo design and in silico validation of a new cationic antimicrobial peptide against E. coli is the subject of the current research [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. A rational, template-directed strategy was employed to prepare the peptide in order to optimize certain critical properties including net positive charge, hydrophobicity, and amphipathicity-features that are crucial to membrane activity [21,]. Advanced computational methods were used in the analysis of physicochemical properties, conformational structure, membrane partitioning energetics and hemolytic toxicity [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Also, machine learning models were used to forecast safety profiles and offer mechanistic understanding of peptides [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study seeks to add to the next generation of antimicrobial therapeutics by integrating principles of peptide chemistry, membrane biophysics, and AI-driven analysis. The results do not only indicate the practicability of computational AMP design, but also offer a scalable platform of developing viable and secure alternatives to the traditional antibiotics in combating antimicrobial resistance.\u003c/p\u003e"},{"header":"2.0 Methodology","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Sequence Design and Candidate Selection\u003c/h2\u003e \u003cp\u003eA novel antimicrobial peptide (AMP) candidate was designed de novo using a template-guided rational design strategy. The initial scaffold was a 15-residue peptide (WKKIWKDPGIKKWIK), selected based on its enrichment in lysine (K) and tryptophan (W) residues, a compositional pattern commonly associated with membrane-active AMP scaffolds and STAMP-like motifs. This design strategy aimed to optimize cationic charge and hydrophobicity, key determinants for selective interaction with negatively charged bacterial membranes, particularly in Gram-negative organisms such as \u003cem\u003eEscherichia coli\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eA central proline residue was retained to introduce a conformational disruption within the peptide backbone. This proline-induced hinge was incorporated to promote structural flexibility and enable segmental adaptation during membrane interaction. To further enhance electrostatic attraction toward anionic membrane components, a terminal arginine residue was appended to the C-terminus, resulting in the final 16-residue sequence: WKKIWKDPGIKKWIKR.\u003c/p\u003e \u003cp\u003eThe designed peptide was treated as a computational candidate and was not derived from a naturally occurring sequence. Sequence composition and motif design were guided by established structure\u0026ndash;activity relationships reported for antimicrobial peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Physicochemical Property Calculation\u003c/h2\u003e \u003cp\u003ePhysicochemical properties of the designed peptide (WKKIWKDPGIKKWIKR) were calculated using the modlAMP and Biopython libraries in a Python-based environment. The peptide sequence was provided as input and analyzed using the GlobalDescriptor module of modlAMP to compute key parameters relevant to antimicrobial activity.\u003c/p\u003e \u003cp\u003eThe calculated molecular weight of the peptide was approximately 2110.6 Da, consistent with the typical size range of short antimicrobial peptides (\u0026lt;\u0026thinsp;3000 Da). The net charge at physiological pH (7.4) was determined to be +\u0026thinsp;5.99, reflecting a strongly cationic nature favorable for interaction with negatively charged bacterial membranes. The isoelectric point (pI) was calculated as 11.31, further confirming the peptide\u0026rsquo;s basic character.\u003c/p\u003e \u003cp\u003eThe Boman index, which estimates protein-binding potential and membrane interaction capability, was calculated to be 2.141, indicating moderate to high affinity for lipid interfaces. These physicochemical properties collectively suggest that the designed peptide possesses characteristics consistent with membrane-active antimicrobial peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Amphipathicity Analysis\u003c/h2\u003e \u003cp\u003eThe amphipathic character of the designed peptide was evaluated using hydrophobic moment (\u0026micro;H), calculated based on the Eisenberg consensus hydrophobicity scale using the modlAMP Python library. Hydrophobic moment quantifies the spatial segregation of hydrophobic and hydrophilic residues along a helical axis and is widely used to assess membrane-active potential in antimicrobial peptides.\u003c/p\u003e \u003cp\u003eHydrophobic moment was computed across the full peptide sequence (WKKIWKDPGIKKWIKR) using a large window size (window\u0026thinsp;=\u0026thinsp;1000) to capture global amphipathic behavior. The calculated global hydrophobic moment (\u0026micro;H\u0026thinsp;\u0026asymp;\u0026thinsp;0.143) indicated a low apparent amphipathicity when averaged across the entire sequence.\u003c/p\u003e \u003cp\u003eTo account for structural discontinuity introduced by the central proline residue, a segmental analysis was performed by dividing the sequence into N-terminal (WKKIWKD) and C-terminal (GIKKWIKR) regions. Hydrophobic moment was calculated independently for each segment using the same Eisenberg-based approach. The N-terminal segment exhibited a hydrophobic moment of approximately \u0026micro;H\u0026thinsp;\u0026asymp;\u0026thinsp;0.815, while the C-terminal segment showed \u0026micro;H\u0026thinsp;\u0026asymp;\u0026thinsp;0.858, both values exceeding the typical threshold (\u0026gt;\u0026thinsp;0.4) associated with amphipathic α-helices.\u003c/p\u003e \u003cp\u003eThis segmental disparity suggests that the peptide does not behave as a single continuous amphipathic helix but rather as two amphipathic subdomains separated by a proline-induced structural hinge. This configuration supports a flexible, hinge-driven topology commonly observed in membrane-active antimicrobial peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Secondary Structure Visualization\u003c/h2\u003e \u003cp\u003eHelical wheel projections were generated to visualize residue distribution around the helical axis and to qualitatively evaluate amphipathic patterning. Helical wheel projections were generated using the modlAMP helical wheel function. These plots were used to identify whether hydrophobic residues such as tryptophan and isoleucine clustered on one face of the helix, while cationic or polar residues such as lysine, arginine, and aspartate occupied the opposite face.\u003c/p\u003e \u003cp\u003eThis visualization was particularly useful for interpreting the structural consequences of the central proline residue, which was expected to induce a local bend or kink. The resulting topology was evaluated in the context of common AMP architectures, including hinge-containing and boomerang-like membrane-active peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Three-Dimensional Structure Prediction\u003c/h2\u003e \u003cp\u003eProtein structure prediction was performed using ColabFold (v1.6.1), which integrates the deep learning\u0026ndash;based AlphaFold2 with rapid multiple sequence alignment (MSA) generation via MMseqs2. The peptide sequence (WKKIWKDPGIKKWIKR) was submitted as a single-chain input for monomeric structure prediction. MSAs were generated using the mmseqs2_uniref_env mode, incorporating both UniRef and environmental sequence databases, with a pairing strategy set to unpaired_paired to include both paired and unpaired alignments. Template-based modeling was disabled (template_mode\u0026thinsp;=\u0026thinsp;none), ensuring that predictions relied solely on de novo inference. Model parameters were maintained at default initialized settings, including model_type\u0026thinsp;=\u0026thinsp;auto, which applies AlphaFold2-ptm for monomer prediction, num_recycles\u0026thinsp;=\u0026thinsp;3, and recycle_early_stop_tolerance\u0026thinsp;=\u0026thinsp;auto, with a greedy pairing strategy and additional calculation of pairwise ipTM/actifpTM scores enabled. Sampling parameters were kept unchanged (max_msa\u0026thinsp;=\u0026thinsp;auto, num_seeds\u0026thinsp;=\u0026thinsp;1, and dropout disabled). Structural relaxation using AMBER was not performed (num_relax\u0026thinsp;=\u0026thinsp;0, relax_max_iterations\u0026thinsp;=\u0026thinsp;200) to reduce computational overhead. All predictions were executed within the ColabFold notebook environment, and resulting models were ranked based on internal confidence metrics, including predicted Local Distance Difference Test (pLDDT) and predicted TM-score (pTM). The top-ranked structure (rank 1) was selected for further analysis and visualization using pLDDT-based coloring. All output files, including predicted structures, alignment data, and confidence scores, were retained with save_all and save_recycles enabled at a resolution of 200 dpi for downstream analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Structural Visualization in ChimeraX\u003c/h2\u003e \u003cp\u003eStructural visualization and analysis were performed using UCSF ChimeraX (version 1.11.1, released January 23, 2026). The predicted peptide structure obtained from ColabFold was imported in PDB format and processed within the ChimeraX environment. Secondary structure elements were automatically assigned using built-in algorithms, and the structure was visualized using a cartoon (ribbon) representation to assess overall folding and α-helical content.\u003c/p\u003e \u003cp\u003eResidue-level confidence was visualized by coloring the structure according to B-factor values corresponding to AlphaFold-derived pLDDT scores using the \u0026ldquo;alphafold\u0026rdquo; color palette. Cartoon thickness was adjusted to enhance structural clarity, and additional stick representations were briefly used to inspect atomic-level interactions before reverting to ribbon visualization. The solvent background was set to white to improve image contrast for publication-quality rendering.\u003c/p\u003e \u003cp\u003eStructural analysis focused on identifying helix formation, residue distribution, and conformational features such as the central proline-induced hinge. No structural refinement or energy minimization was performed within ChimeraX. High-resolution images were generated using supersampling to enhance visual quality, and final structures were exported as PNG files for figure preparation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Membrane Partitioning Analysis\u003c/h2\u003e \u003cp\u003eThe thermodynamic propensity of the peptide to associate with lipid membranes was estimated using the Wimley\u0026ndash;White interfacial hydrophobicity scale. Residue-specific free-energy contributions were assigned based on the peptide sequence, and the total membrane partitioning energy (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\Delta\\:}G\\)\u003c/span\u003e\u003c/span\u003e) was calculated by summing the individual values.\u003c/p\u003e \u003cp\u003eIn addition to residue-level analysis, a smoothed local energy profile was generated using a moving-window average to approximate local helical environments and reduce single-residue noise. The resulting thermodynamic profile was used to estimate the peptide\u0026rsquo;s membrane insertion potential and compare its behavior with known AMPs and non-AMP control peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.8 Benchmarking Against Reference Peptides\u003c/h2\u003e \u003cp\u003eThermodynamic benchmarking of membrane interaction propensity was performed using the Wimley\u0026ndash;White interfacial hydrophobicity scale, which provides experimentally derived free energy values (ΔG, kcal/mol) for amino acid partitioning at membrane interfaces. Residue-specific ΔG values were assigned to each amino acid, and total membrane anchoring energy was calculated by summing all energetically favorable (negative) contributions. For the designed peptide (WKKIWKDPGIKKWIKR), this analysis yielded a total anchoring energy of approximately\u0026thinsp;\u0026minus;\u0026thinsp;6.48 kcal/mol, indicating a strong intrinsic propensity for membrane insertion driven primarily by hydrophobic residues such as tryptophan (W) and isoleucine (I).\u003c/p\u003e \u003cp\u003eTo contextualize this value, comparative benchmarking was performed against a panel of experimentally validated antimicrobial peptides (AMPs) and non-antimicrobial human protein fragments. The AMP reference set included melittin (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;5.02 kcal/mol), magainin 2 (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;4.80 kcal/mol), LL-37 (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;7.69 kcal/mol), and indolicidin (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;10.12 kcal/mol), representing peptides with well-characterized membrane-disruptive activity. Buforin II (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;2.81 kcal/mol) was also included as a lower membrane-active AMP for comparison. The negative control group consisted of fragments from human serum albumin (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;3.38 kcal/mol), hemoglobin beta (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;2.97 kcal/mol), and actin alpha (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;2.49 kcal/mol), which are not expected to spontaneously insert into lipid bilayers.\u003c/p\u003e \u003cp\u003eAll sequences were analyzed using a custom Python-based implementation of the Wimley\u0026ndash;White scale, and resulting ΔG values were compiled into a structured dataset and ranked from most negative (strongest membrane affinity) to least negative. A theoretical membrane insertion threshold of approximately\u0026thinsp;\u0026minus;\u0026thinsp;3.0 kcal/mol was used as a reference point to distinguish membrane-active peptides from non-interacting or weakly interacting sequences. Comparative visualization was performed using horizontal bar plots, where the designed peptide, reference AMPs, and non-AMP controls were color-coded and displayed according to their respective ΔG values.\u003c/p\u003e \u003cp\u003eAll computations and visualizations were performed using Python (NumPy, Pandas, and Matplotlib) within a Google Colab environment.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.9 Toxicity Prediction Using Machine Learning\u003c/h2\u003e \u003cp\u003eHemolytic toxicity of the designed peptide was predicted using a supervised machine learning approach implemented in Python. A Random Forest classifier was trained on a curated dataset of 14 experimentally characterized antimicrobial peptides annotated as hemolytic (n\u0026thinsp;=\u0026thinsp;7) or non-hemolytic (n\u0026thinsp;=\u0026thinsp;7). The dataset included representative peptides such as melittin, LL-37, mastoparan, aurein 1.2, indolicidin, temporin L, and ovispirin (hemolytic class), as well as magainin 2, buforin II, pexiganan, alyteserin, rana-box, dermaseptin, and esculentin-2 (non-hemolytic class).\u003c/p\u003e \u003cp\u003eEach peptide sequence was transformed into a numerical feature vector using sequence-derived physicochemical descriptors. The extracted features included: peptide length, net charge (calculated as the difference between positively charged residues [Lys, Arg] and negatively charged residues [Asp, Glu]), membrane anchoring energy derived from the Wimley\u0026ndash;White hydrophobicity scale (sum of negative ΔG contributions), hydrophobic residue ratio (fraction of residues belonging to A, V, I, L, M, F, Y, W), and aromatic residue ratio (fraction of W, F, and Y residues).\u003c/p\u003e \u003cp\u003eFeature scaling was performed using standard normalization (StandardScaler) to ensure comparable feature distributions. The Random Forest model was trained using 100 decision trees (n_estimators\u0026thinsp;=\u0026thinsp;100) with a fixed random seed (random_state\u0026thinsp;=\u0026thinsp;42) to ensure reproducibility. The trained model was then used to predict the hemolytic potential of the designed peptide (WKKIWKDPGIKKWIKR), yielding a probability score between 0 and 1. A threshold of 0.5 was applied to classify the peptide as hemolytic or non-hemolytic.\u003c/p\u003e \u003cp\u003eFeature importance analysis was conducted using Gini impurity-based importance scores derived from the trained Random Forest model, providing insight into the relative contribution of each physicochemical feature to the toxicity prediction.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e2.9.1 Explainable AI Analysis\u003c/h2\u003e \u003cp\u003eExplainable artificial intelligence (XAI) analysis was performed using SHAP (SHapley Additive exPlanations) with the TreeExplainer algorithm, which is specifically optimized for tree-based models such as Random Forest. This approach enabled quantitative decomposition of the model output into feature-level contributions, facilitating mechanistic interpretation of hemolytic toxicity predictions.\u0026rdquo;\u003c/p\u003e \u003cp\u003eA SHAP waterfall plot was used to visualize baseline prediction, feature-level contributions, and final model output. This helped identify the major biophysical drivers of toxicity and supported mechanistic interpretation of the classifier\u0026rsquo;s prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e2.9.2 Computational Environment and Implementation Details\u003c/h2\u003e \u003cp\u003eAll computational analyses were performed in a cloud-based environment using Google Colab, ensuring reproducibility and accessibility of the workflow. The entire computational pipeline was implemented in Python (version 3.10), utilizing open-source libraries for sequence analysis, machine learning, and data visualization. Specifically, Biopython (v1.81) was employed for sequence-based physicochemical calculations and proteolytic digestion analysis, while modlAMP (v4.3) was used for hydrophobic moment (\u0026micro;H) calculations and amphipathicity profiling. Numerical computations and data handling were conducted using NumPy (v1.26) and Pandas (v2.0), and graphical representations were generated using Matplotlib (v3.7) and Seaborn (v0.12). Machine learning implementation was performed using scikit-learn (v1.3), and model interpretability was achieved using SHAP (v0.43) for feature attribution analysis. Structural predictions were carried out using ColabFold, while molecular visualization and structural analysis were conducted using ChimeraX.\u003c/p\u003e \u003cp\u003eAll computations were executed on Google Colab\u0026rsquo;s GPU-enabled infrastructure, utilizing NVIDIA Tesla T4 or V100 GPUs depending on session availability, with approximately 12\u0026ndash;16 GB of RAM allocated for runtime operations. CPU-based execution was used for lightweight preprocessing and feature extraction steps.\u003c/p\u003e \u003cp\u003eFor machine learning analysis, a Random Forest classifier was implemented using scikit-learn with 100 estimators, Gini impurity as the splitting criterion, and default unrestricted tree depth. The dataset was partitioned into training and testing sets using an 80:20 split ratio. Feature vectors were constructed from sequence-derived descriptors, including peptide length, net charge, hydrophobic residue ratio, aromatic residue ratio, and membrane anchoring energy (ΔG). Model performance was evaluated using accuracy metrics, and feature importance was derived from Gini-based impurity reduction.\u003c/p\u003e \u003cp\u003eTo enhance interpretability, SHAP (SHapley Additive exPlanations) analysis was applied using the TreeExplainer method optimized for tree-based models. SHAP values were computed to quantify the contribution of individual features to the predicted hemolytic toxicity of the peptide. Waterfall plots were generated to visualize the baseline prediction, feature-wise contributions, and final model output. Positive SHAP values indicated features contributing to increased toxicity, whereas negative values indicated protective or stabilizing effects, enabling precise identification of structural determinants influencing hemolytic potential.\u003c/p\u003e \u003cp\u003eAll computational steps, including sequence design, structural prediction, physicochemical analysis, thermodynamic modeling, and machine learning, were integrated into a unified workflow within the Colab environment. Random seeds were fixed where applicable to ensure reproducibility of machine learning outputs. All scripts, intermediate data, and generated outputs (e.g., .pdb files, feature matrices, and plots) were retained within the runtime environment and are available for export to support external validation and reproducibility.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3.0 Results","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Physicochemical Profiling and Amphipathic Character\u003c/h2\u003e \u003cp\u003eThe rationally designed 16-mer candidate peptide (WKKIWKDPGIKKWIKR) was first evaluated for its foundational physicochemical properties. Sequence analysis confirmed a highly cationic nature, yielding a net charge of \u003cb\u003e+\u0026thinsp;6\u003c/b\u003e at physiological pH (7.4), driven by the enrichment of lysine (K) and the C-terminal arginine (R) residues. This pronounced positive charge provides a strong electrostatic driving force for initial attraction to the anionic lipopolysaccharides (LPS) of the \u003cem\u003eE. coli\u003c/em\u003e outer membrane.\u003c/p\u003e \u003cp\u003eCalculations of the Boman index indicated a high potential for protein and membrane interaction, typical of active AMPs, while the instability index suggested that the peptide maintains structural viability within biological parameters. Furthermore, helical wheel projections and calculations of the hydrophobic moment (\u0026micro;H) confirmed a strong amphipathic character. The central proline residue (P8) effectively partitioned the sequence, generating distinct segmental amphipathicity. The strategic placement of hydrophobic residues (Tryptophan and Isoleucine) on one face and cationic residues on the opposing face established an optimal topological profile for membrane insertion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Three-Dimensional Structure and Topology\u003c/h2\u003e \u003cp\u003eStructural modeling was performed using ColabFold under the parameters described in Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e2.5\u003c/span\u003e, providing high-confidence predictions of the peptide\u0026rsquo;s conformational tendencies. Visualized using ChimeraX, the predicted 3D structure of the designed antimicrobial peptide exhibits a dominant alpha-helical core accompanied by a flexible terminal region (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis structural architecture aligns well with known membrane-active AMPs. The alpha-helical core facilitates deep insertion into the hydrophobic tail region of the bacterial lipid bilayer, while the central proline acts as a structural hinge. This localized flexibility likely allows the peptide to dynamically adapt its conformation upon encountering the membrane interface, maximizing contact area and promoting disruptive mechanisms such as toroidal pore or carpet-like formation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe predicted structure shows a mostly alpha-helical antimicrobial peptide with a flexible looped tail. This type of fold is common in AMPs because the helix can help the peptide interact with the negatively charged bacterial membrane, while the flexible region may help it adapt during binding\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Membrane Partitioning Thermodynamics and Residue-Level Energy Profile\u003c/h2\u003e \u003cp\u003eThe membrane interaction propensity of the designed antimicrobial peptide (WKKIWKDPGIKKWIKR) was evaluated using the Wimley\u0026ndash;White interfacial hydrophobicity scale to estimate residue-specific free energies of partitioning (ΔG). The resulting thermodynamic profile (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) provides insight into the energetic landscape governing peptide\u0026ndash;membrane interactions.\u003c/p\u003e \u003cp\u003eThe analysis revealed a distinct amphipathic energetic pattern, characterized by alternating favorable and unfavorable contributions along the peptide sequence. Hydrophobic residues, particularly tryptophan (W) and isoleucine (I), exhibited strongly negative ΔG values (reaching approximately\u0026thinsp;\u0026minus;\u0026thinsp;1.8 kcal/mol), indicating a high propensity for spontaneous insertion into the lipid bilayer. These residues are likely to function as membrane-anchoring elements, facilitating penetration into the hydrophobic core of the membrane.\u003c/p\u003e \u003cp\u003eIn contrast, cationic residues such as lysine (K) and arginine (R) displayed positive ΔG values, reflecting energetically unfavorable insertion into the membrane interior. However, this is consistent with their expected role in electrostatic interactions with negatively charged lipid headgroups, particularly in Gram-negative bacterial membranes. This complementary distribution of hydrophobic and charged residues supports a functionally optimized amphipathic architecture.\u003c/p\u003e \u003cp\u003eThe smoothed local energy profile further highlighted cooperative membrane-binding regions, with favorable segments concentrated in the N-terminal and central regions of the peptide. Notably, the central proline residue appears to introduce a disruption in the energetic continuity, supporting its proposed role as a structural hinge that may facilitate conformational adaptability during membrane interaction\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eResidue-level free energy of partitioning (ΔG, kcal/mol) for WKKIWKDPGIKKWIKR calculated using the Wimley\u0026ndash;White interfacial hydrophobicity scale. Bars represent individual residue contributions, where negative ΔG (blue) indicates favorable membrane insertion and positive ΔG (red) indicates unfavorable insertion. The black line shows the smoothed local membrane affinity profile (moving average).\u003c/p\u003e \u003cp\u003eHydrophobic residues (W, I) display strongly favorable energies, consistent with membrane anchoring, while cationic residues (K, R) show unfavorable insertion but contribute to interfacial electrostatic interactions. The overall pattern reflects a segmented amphipathic profile characteristic of membrane-active peptides.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Thermodynamic Benchmarking Against Reference Peptides\u003c/h2\u003e \u003cp\u003eThe membrane partitioning behavior of the designed peptide was further evaluated through comparative thermodynamic benchmarking against a panel of experimentally characterized antimicrobial peptides (AMPs) and non-antimicrobial human protein fragments (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Total membrane anchoring energies (ΔG) were calculated using the Wimley\u0026ndash;White interfacial hydrophobicity scale and used as a quantitative measure of membrane affinity.\u003c/p\u003e \u003cp\u003eThe results demonstrate that the designed peptide exhibits a strongly favorable total free energy of membrane partitioning (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;6 kcal/mol), positioning it well within the range of known membrane-active AMPs. Notably, its thermodynamic profile is comparable to established peptides such as magainin 2 and melittin, both of which are known for their effective membrane-disruptive properties. While the designed peptide shows slightly less favorable energetics than highly potent AMPs such as indolicidin and LL-37, it remains significantly more membrane-active than non-AMP controls.\u003c/p\u003e \u003cp\u003eIn contrast, protein fragments derived from human serum albumin, hemoglobin beta, and actin alpha exhibited relatively weak or near-neutral partitioning energies (ΔG\u0026thinsp;\u0026gt;\u0026thinsp;\u0026minus;\u0026thinsp;3 kcal/mol), indicating limited intrinsic membrane affinity. This clear separation between AMP and non-AMP groups supports the validity of the thermodynamic framework used in this study.\u003c/p\u003e \u003cp\u003eImportantly, the designed peptide falls well below the empirical threshold (ΔG\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;3 kcal/mol) associated with effective membrane insertion, reinforcing its classification as a membrane-active candidate. At the same time, its intermediate positioning relative to highly cytotoxic peptides such as melittin suggests a potential balance between antimicrobial efficacy and reduced host toxicity, which is a critical consideration in AMP design.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eComparison of total membrane anchoring energies (ΔG, kcal/mol) calculated using the Wimley\u0026ndash;White interfacial hydrophobicity scale for the designed peptide and a panel of reference sequences. The dataset includes known antimicrobial peptides (blue), non-antimicrobial human protein fragments (gray), and the designed peptide (red). More negative ΔG values indicate stronger membrane affinity and a higher propensity for membrane insertion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Biophysical Determinants of Predicted Hemolytic Toxicity\u003c/h2\u003e \u003cp\u003eTo identify the key physicochemical features governing the predicted hemolytic toxicity of the designed peptide, a Random Forest classifier was employed and feature importance was evaluated using Gini impurity-based metrics (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe analysis revealed that peptide length is the most influential factor, contributing the highest importance (~\u0026thinsp;0.28) to the model\u0026rsquo;s decision. This suggests that peptide size plays a critical role in modulating toxicity, likely by influencing membrane coverage, pore formation potential, and overall structural stability.\u003c/p\u003e \u003cp\u003eThe second most significant contributor was membrane anchoring energy (ΔG) (~\u0026thinsp;0.21), highlighting the importance of thermodynamic membrane affinity in determining cytotoxic behavior. Peptides with stronger membrane insertion capabilities are more likely to disrupt not only bacterial membranes but also host cell membranes, thereby increasing hemolytic risk.\u003c/p\u003e \u003cp\u003eAromatic residue content (~\u0026thinsp;0.19), primarily driven by tryptophan residues, also showed a strong contribution. Aromatic residues are known to facilitate membrane insertion by interacting with lipid headgroup regions, which can enhance both antimicrobial activity and toxicity. Similarly, the hydrophobic residue ratio (~\u0026thinsp;0.18) was identified as a key determinant, reflecting the role of hydrophobic interactions in membrane permeabilization.\u003c/p\u003e \u003cp\u003eIn contrast, net charge (~\u0026thinsp;0.14) exhibited the lowest relative importance among the evaluated features. While cationic charge is essential for selective binding to negatively charged bacterial membranes, this result suggests that charge alone is not the primary driver of toxicity, but rather acts in combination with hydrophobic and structural properties.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFeature importance analysis derived from a Random Forest classifier illustrating the relative contribution of key physicochemical properties to the prediction of hemolytic toxicity. Importance scores are based on Gini impurity reduction, where higher values indicate greater influence on the model\u0026rsquo;s decision.\u003c/p\u003e \u003cp\u003ePeptide length emerged as the most influential feature, followed by membrane anchoring energy (ΔG), highlighting the importance of structural size and membrane affinity in toxicity prediction. Aromatic residue ratio and hydrophobic residue ratio also contributed substantially, reflecting the role of membrane-interacting residues in modulating cytotoxic effects. In contrast, net charge showed comparatively lower importance, suggesting that electrostatic interactions alone are insufficient to explain hemolytic behavior.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Explainable AI Analysis of Hemolytic Toxicity (SHAP Interpretation)\u003c/h2\u003e \u003cp\u003eTo further interpret the machine learning-based toxicity prediction, SHAP (SHapley Additive exPlanations) analysis was applied to quantify the contribution of individual physicochemical features to the predicted hemolytic risk of the designed peptide (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe model yielded a predicted toxicity score of f(x)\u0026thinsp;=\u0026thinsp;0.54, relative to a baseline expectation of E[f(x)]\u0026thinsp;=\u0026thinsp;0.486, indicating a moderate but controlled hemolytic potential. Feature-level contributions revealed a nuanced interplay between factors that increase and decrease toxicity.\u003c/p\u003e \u003cp\u003eAmong the positive contributors, the aromatic residue ratio exhibited the strongest effect (+\u0026thinsp;0.08), suggesting that the presence of tryptophan-rich regions enhances membrane interaction and contributes to increased cytotoxicity. Similarly, membrane anchoring energy (ΔG) contributed positively (+\u0026thinsp;0.05), indicating that stronger membrane affinity is associated with a higher likelihood of disrupting host cell membranes.\u003c/p\u003e \u003cp\u003eIn contrast, several features acted to reduce predicted toxicity. The hydrophobic residue ratio (\u0026minus;\u0026thinsp;0.04) and peptide length (\u0026minus;\u0026thinsp;0.04) both contributed negatively, suggesting that the specific balance of hydrophobic content and the relatively short peptide length may help limit excessive membrane disruption. These findings imply that while hydrophobicity is necessary for antimicrobial function, its controlled distribution plays a protective role against toxicity.\u003c/p\u003e \u003cp\u003eThe net charge showed only a minimal positive contribution (+\u0026thinsp;0.01), reinforcing earlier observations that electrostatic interactions alone are not the primary drivers of hemolytic activity, but instead function in coordination with other structural and thermodynamic factors.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSHAP (SHapley Additive exPlanations) analysis showing feature-wise contributions to the predicted hemolytic toxicity of the designed peptide. The baseline prediction (E[f(x)]\u0026thinsp;=\u0026thinsp;0.486) is shifted to the final output (f(x)\u0026thinsp;=\u0026thinsp;0.54) by individual feature contributions. Positive values (red) increase predicted toxicity, while negative values (blue) decrease it.\u003c/p\u003e \u003cp\u003eThe aromatic residue ratio is the dominant positive contributor, followed by membrane anchoring energy (ΔG), indicating that increased aromatic content and membrane affinity elevate toxicity. In contrast, hydrophobic residue ratio and peptide length contribute negatively, moderating the prediction. Net charge shows a minimal positive effect, suggesting a limited role in toxicity determination.\u003c/p\u003e \u003c/div\u003e"},{"header":"4.0 Discussion","content":"\u003cp\u003eThe present study demonstrates a rational and integrative framework for the de novo design of antimicrobial peptides by combining biophysical modeling with explainable machine learning. The designed peptide (WKKIWKDPGIKKWIKR) exhibits key hallmarks of membrane-active AMPs, including a strong cationic charge, pronounced amphipathicity, and a structurally adaptive topology driven by a central proline hinge.\u003c/p\u003e \u003cp\u003eThermodynamic profiling revealed a favorable membrane partitioning landscape, with hydrophobic residues driving insertion and cationic residues stabilizing interfacial interactions [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The residue-level ΔG distribution and smoothed energy profile support a cooperative and segmented membrane interaction mechanism, consistent with known AMP behaviors such as toroidal pore formation or carpet-like disruption [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Importantly, benchmarking against experimentally validated peptides positioned the candidate within the functional AMP range, with a total anchoring energy comparable to magainin 2 and melittin, yet less extreme than highly cytotoxic peptides such as indolicidin [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning analysis further provided insight into the determinants of hemolytic toxicity. Feature importance indicated that toxicity is governed by a combination of peptide length, membrane affinity, and aromatic content rather than charge alone [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. SHAP-based interpretation refined this understanding by demonstrating that toxicity-enhancing features, such as aromatic residue enrichment and strong membrane anchoring, are partially offset by stabilizing factors including controlled hydrophobicity and optimized peptide length [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTogether, these findings highlight the importance of achieving a balanced biophysical profile. The designed peptide appears to occupy a favorable intermediate space, combining effective membrane activity with moderated toxicity, thereby supporting its potential as a promising candidate for further experimental validation against multidrug-resistant pathogens.\u003c/p\u003e"},{"header":"5.0 Conclusion","content":"\u003cp\u003eThis study is a comprehensive AI-guided framework for the de novo design and in silico validation of antimicrobial peptides targeting multidrug-resistant pathogens. The designed 16-mer peptide (WKKIWKDPGIKKWIKR) highlights a favorable combination of physicochemical and biophysical properties, including strong cationic charge, amphipathic organization, and a structurally adaptive topology facilitated by a central proline hinge.\u003c/p\u003e \u003cp\u003eThermodynamic analysis confirmed a favorable membrane partitioning profile, with residue-level contributions supporting efficient membrane insertion and interaction. Comparative benchmarking against known antimicrobial peptides further validated that the designed sequence falls within the functional range of membrane-active peptides while maintaining a more balanced energetic profile than highly cytotoxic counterparts.\u003c/p\u003e \u003cp\u003eImportantly, machine learning-based toxicity prediction, complemented by SHAP interpretability, revealed that hemolytic risk is governed by a combination of membrane affinity, aromatic content, and structural features. The designed peptide achieves a controlled balance between these factors, suggesting the potential to retain antimicrobial efficacy while minimizing host toxicity.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSalam MA, Al-Amin MY, Salam MT, Pawar JS, Akhter N, Rabaan AA, Alqumber MA (2023), July Antimicrobial resistance: a growing serious threat for global public health. In \u003cem\u003eHealthcare\u003c/em\u003e (Vol. 11, No. 13, p. 1946). MDPI\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMuteeb G, Rehman MT, Shahwan M, Aatif M (2023) Origin of antibiotics and antibiotic resistance, and their impacts on drug development: A narrative review. Pharmaceuticals 16(11):1615\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchneider YK (2021) Bacterial natural product drug discovery for new antibiotics: strategies for tackling the problem of antibiotic resistance by efficient bioprospecting. Antibiotics 10(7):842\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMacNair CR, Rutherford ST, Tan MW (2024) Alternative therapeutic strategies to treat antibiotic-resistant pathogens. Nat Rev Microbiol 22(5):262\u0026ndash;275\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMookherjee N, Anderson MA, Haagsman HP, Davidson DJ (2020) Antimicrobial host defence peptides: functions and clinical potential. Nat Rev Drug Discovery 19(5):311\u0026ndash;332\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang X, Li G (2023) Antimicrobial peptides and cell-penetrating peptides: non-antibiotic membrane-targeting strategies against bacterial infections. Infect Drug Resist, 1203\u0026ndash;1219\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePirtskhalava M, Vishnepolsky B, Grigolava M, Managadze G (2021) Physicochemical Features and Peculiarities of Interaction of AMP with the Membrane. Pharmaceuticals 14(5):471\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuong L, Gross SP, Siryaporn A (2021) Developing antimicrobial synergy with AMPs. Front Med Technol 3:640981\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeyfi R, Kahaki FA, Ebrahimi T, Montazersaheb S, Eyvazi S, Babaeipour V, Tarhriz V (2020) Antimicrobial peptides (AMPs): roles, functions and mechanism of action. Int J Pept Res Ther 26(3):1451\u0026ndash;1463\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOliveira J\u0026uacute;nior NG, Souza CM, Buccini DF, Cardoso MH, Franco OL (2025) Antimicrobial peptides: structure, functions and translational applications. Nat Rev Microbiol 23(11):687\u0026ndash;700\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTaheri-Araghi S (2024) Synergistic action of antimicrobial peptides and antibiotics: current understanding and future directions. Front Microbiol 15:1390765\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang C, Yang M (2022) Antimicrobial peptides: from design to clinical application. Antibiotics 11(3):349\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng B, Wang X, Guo M, Tzeng CM (2025) Therapeutic peptides: recent advances in discovery, synthesis, and clinical translation. Int J Mol Sci 26(11):5131\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCosta F, Teixeira C, Gomes P, Martins MCL (2019) Clinical application of AMPs. Antimicrobial peptides: basics for clinical application. Springer Singapore, Singapore, pp 281\u0026ndash;298\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeng H (2025) AI-driven discovery and design of antimicrobial peptides: Progress, challenges, and opportunities. Probiotics Antimicrob Proteins, 1\u0026ndash;23\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X, Li C, Bernards MT, Shi Y, Shao Q, He Y (2021) Sequence-based peptide identification, generation, and property prediction with deep learning: a review. Mol Syst Des Eng 6(6):406\u0026ndash;428\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoudhury PR, Mishra SK, Yadav S, Singh S, Mathur P (2025) In Silico Peptide Design: Methods, Resources, and Role of AI. J Pept Sci, 31(12), e70063\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown ED, Wright GD (2005) New targets and screening approaches in antimicrobial drug discovery. Chem Rev 105(2):759\u0026ndash;774\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmorim AM, Piochi LF, Gaspar AT, Preto AJ, Rosario-Ferreira N, Moreira IS (2024) Advancing drug safety in drug development: bridging computational predictions for enhanced toxicity prediction. Chem Res Toxicol 37(6):827\u0026ndash;849\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin S, Wade JD, Liu S (2020) De novo design of flavonoid-based mimetics of cationic antimicrobial peptides: Discovery, development, and applications. Acc Chem Res 54(1):104\u0026ndash;119\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYuhao Z, Chen Y, Jiecheng L, Yan L, Yufen Z (2025) Prebiotic template-directed peptide formation mediated by polyphosphates and peptide promoters. Org Biomol Chem 23(36):8205\u0026ndash;8211\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobles-Loaiza, A. A., Pinos-Tamayo, E. A., Mendes, B., Ortega-Pila, J. A., Proa\u0026ntilde;o-Bola\u0026ntilde;os,C., Plisson, F., \u0026hellip; Almeida, J. R. (2022). Traditional and computational screening of non-toxic peptides and approaches to improving selectivity. Pharmaceuticals, 15(3), 323.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoto-Garcia N, Davari MD, Medina‐Ortiz D (2026) Machine Learning to Accelerate the Discovery of Therapeutic Peptides. Mach Learn Big Data‐enabled Biotechnol, 183\u0026ndash;217\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang N, Dai S, Qi R, Han Y (2025) Thermodynamic studies on partition of glutamic amphiphiles into phospholipid membranes with different compositions. Colloids Surf A 708:135953\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKyriazis V, Gadhe L, Konstantoulea K, Louros N, Schymkowitz J, Rousseau F (2025) Amyloid Explorer: a global atlas of amyloid fibril structures and thermodynamic principles. bioRxiv, 2025\u0026ndash;2010\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOliveira J\u0026uacute;nior NG, Souza CM, Buccini DF, Cardoso MH, Franco OL (2025) Antimicrobial peptides: structure, functions and translational applications. Nat Rev Microbiol 23(11):687\u0026ndash;700\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhatnagar P, Khandelwal Y, Mishra S, Dutta A, Mitra D, Biswas S (2024) Predicting antibacterial activity, efficacy, and hemotoxicity of peptides using an explainable machine learning framework. Process Biochem 145:163\u0026ndash;174\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYogarathinam LT, Abba SI, Usman J, Ramamoorthy M, Aljundi IH (2025) Interpretable SHAP-based machine learning-assisted design for selecting ultrafiltration membranes in protein-laden phosphate wastewater. Clean Chem Eng 11:100187\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"The University of Lahore","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"E. Coli, AMP, De Novo , Alphfold/Colab Fold, XAI, MDR ","lastPublishedDoi":"10.21203/rs.3.rs-9615735/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9615735/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe growing epidemic of antimicrobial resistance (AMR) due to multidrug-resistant (MDR) Gram-negative pathogens like Escherichia coli challenges traditional antibiotic treatment with fast evolutionary shifts and declining drug development pipelines.\u003c/p\u003e \u003cp\u003eThe use of antimicrobial peptides (AMPs) provides a potentially game-changing paradigm shift, using innate cationic-amphipathic structures to permeabilize bacterial membranes through either toroidal pores, carpet mechanisms, or barrel-stave models- multifaceted effects resistant to single-target mechanisms.Although broad-spectrum potent, AMPs encounter translational challenges: host cytotoxicity, serum instability, immunogenicity, and synthetic cost, highlighting the need to be able to engineer with precision de novo.\u003c/p\u003e \u003cp\u003eRecent advances in AI-enhanced computational biology presently enable rational optimization of AMP, combining sequence-to-structure prediction (AlphaFold2/ColabFold), physicochemical profiling (modlAMP/Biopython), interfacial energetics (Wimley-White scales), and toxicity prediction by machine learning.\u003c/p\u003e \u003cp\u003eA new 16-mer cationic AMP (WKKIWKDPGIKKWIKR) by template-guided design, with a proline-induced hinge in a Trp/Lys scaffold that is modified with C-terminal Arg to enable electrostatic selectivity against Gram-negative.\u003c/p\u003e \u003cp\u003eIts efficacy and safety in targeting the E. coli membrane is confirmed by comprehensive in silico studies of helical amphipathicity (high segmental \u0026micro;H), topology (is an 3D α-helix core via ColabFold/ChimeraX), membrane partitioning (ΔG vs. LL-37/magainin benchmarks), and hemolytic risk (Random Forest\u0026thinsp;+\u0026thinsp;A net charge (+\u0026thinsp;6 ) and Boman index and instability measures also confirm stability and binding potential.\u003c/p\u003e \u003cp\u003eThis study proves a complete AI-computational pipeline of AMP discovery that is faster to discover viable therapeutics against MDR threats by reducing wet-lab iterations. Our framework combines biophysics, structural modeling, and explainable ML to become a blueprint to the next generation of antimicrobials.\u003c/p\u003e","manuscriptTitle":"De Novo Design and In Silico Validation of a Cationic Antimicrobial Peptide Using an AI-Guided Framework for Membrane Thermodynamics and Hemolytic Toxicity","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-06 06:05:21","doi":"10.21203/rs.3.rs-9615735/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"6bc4b737-ab77-427c-a7d3-a05a62189b0c","owner":[],"postedDate":"May 6th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67534875,"name":"Bioinformatics"}],"tags":[],"updatedAt":"2026-05-06T06:05:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-06 06:05:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9615735","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9615735","identity":"rs-9615735","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}