Structure-Guided Prioritization of Amphipathic Antimicrobial Peptide Candidates against ESKAPE Pathogens via Generative Design and Physicochemical Filtering

Technical Summary: Structure-Guided Prioritization of Amphipathic Antimicrobial Peptide Candidates

Problem Statement
Antimicrobial resistance (AMR), particularly among ESKAPE pathogens (Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter spp.), poses a critical threat to global health, with conventional antibiotic pipelines failing to keep pace. While Antimicrobial Peptides (AMPs) offer a promising alternative due to their broad-spectrum, membrane-targeted mechanisms and low propensity for inducing resistance, traditional discovery methods (e.g., phage display, high-throughput screening) are often slow, costly, and structurally limited. Although deep generative models can produce novel peptide sequences, a persistent bottleneck exists in the systematic, multi-criteria prioritization of these large candidate libraries. Many generative frameworks yield hundreds to thousands of sequences, yet only a small subset is mechanistically plausible for downstream structural or experimental validation.

Methodology
The authors present a systematic in silico pipeline designed to bridge the gap between generative design and experimental feasibility. The workflow integrates the following stages:

1. Generative Design: The study utilizes BioAMPify, a conditional long short-term memory (LSTM) generative model trained on pathogen-specific AMP distributions. This model generated an initial library of 872 peptide sequences, conditioned on the six ESKAPE pathogens.
2. Physicochemical Filtering: The initial library underwent sequential filtering based on established AMP design principles. Criteria included:
   • Sequence length: 10–40 residues.
   • Net charge (at pH 7.4): +2 to +6.
   • Hydrophobic content: 30–60%.
   • Hydrophobic moment: 0.55–0.80.
   • This step reduced the pool to 85 candidates.
3. Redundancy and Protein-Binding Reduction: Sequences were clustered using CD-HIT (90% identity threshold) to remove redundancy. The Boman index was applied to filter for low protein-binding propensity (≤ 1.0), prioritizing membrane-targeting behavior over non-specific protein interactions, resulting in 47 candidates.
4. Toxicity Prediction: The ToxinPred3 platform (hybrid ML/motif-based model) was used to predict toxicity, retaining only non-toxic peptides. This reduced the dataset to 31 candidates.
5. Structural Analysis:
   • AlphaFold3 was used to predict 3D structures. The authors note that due to the intrinsic flexibility of AMPs, high disorder fractions were expected; structures were interpreted as approximations of membrane-bound conformations.
   • DSSP (v3.0.0) was applied to determine secondary structure composition (α-helix, β-sheet, coil).
   • Principal Component Analysis (PCA) was performed on physicochemical descriptors (charge, hydrophobicity, hydrophobic moment, GRAVY, Boman index) to characterize the design space.
6. Final Prioritization: Candidates were filtered for structural plausibility (helix content ≥ 60%, PC1 > −1, hydrophobic moment ≥ 0.55) and stratified into priority classes. High-priority candidates were defined by helix content ≥ 80% and PC1 > 0.

Key Results

• Pipeline Attrition: The multi-stage filtering process progressively narrowed the library from 872 to 85 (physicochemical), 47 (protein-binding), 31 (toxicity), 22 (structural/PCA), and finally 10 high-priority candidates.
• Filtering Bottlenecks: The hydrophobic moment constraint (0.55–0.80) was the most restrictive single criterion, satisfied by only 21.7% of the initial library, highlighting amphipathic organization as a primary filter.
• Structural Characteristics: The 31 non-toxic candidates exhibited a strong bias toward α-helical conformations (ranging from ~60% to 90% helix content). PCA revealed that the design space is organized along axes of membrane affinity (driven by hydrophobicity/GRAVY) and amphipathic structure (driven by net charge and hydrophobic moment).
• High-Priority Shortlist: The final 10 peptides (targeting S. aureus, K. pneumoniae, E. faecium, P. aeruginosa, and E. cloacae) displayed high helical content (≥ 80%), favorable amphipathic organization, and low protein-binding propensity. Notably, no high-priority candidate was identified for A. baumannii under the applied criteria.
• Feature Analysis: Correlation analysis confirmed a strong negative correlation between helix and coil content (r = −0.82) and a positive relationship between hydrophobic moment and helical content, supporting the structural stabilization of amphipathic motifs.

Significance and Claims
The paper claims to demonstrate that a pathogen-conditioned generative model, when coupled with a rigorous, multi-stage computational filtering pipeline, can efficiently enrich a large sequence space for structurally plausible, membrane-active AMP candidates. The study asserts that this approach:

• Moves beyond passive screening to active, structure-aware generation.
• Provides a resource-efficient shortlist (10 peptides) suitable for future molecular dynamics (MD) validation and experimental testing.
• Validates the utility of combining generative design with physicochemical and structural constraints to address the "generation vs. prioritization" bottleneck in AMP discovery.

The authors explicitly state that the work is limited to in silico characterization. They do not claim experimental validation of the peptides' efficacy but position the 10 high-priority candidates as a defined starting point for subsequent MD simulations and experimental assays (e.g., MIC testing) to establish their mechanistic and therapeutic potential. The study concludes that this framework offers a scalable pathway for early-stage AMP discovery against multidrug-resistant ESKAPE pathogens.