Biological Foundation Models for Antimicrobial Peptide Discovery in Veterinary Pathogens

Abstract

The accelerating emergence of antimicrobial resistance (AMR) among veterinary bacterial pathogens such as Pasteurella multocida, Escherichia coli (including avian pathogenic E. coli or APEC), Staphylococcus aureus, and Clostridium perfringens demands novel therapeutic scaffolds beyond conventional antibiotics. Antimicrobial peptides (AMPs) represent a promising class of host-defense molecules with rapid, membrane-directed mechanisms of action. Biological foundation models (BFMs), including protein language models (pLMs) and generative models such as variational autoencoders (VAEs) and diffusion models, now offer an unsupervised or semi-supervised framework for the discovery and rational design of novel AMPs. This article provides a detailed technical review of the molecular biophysics of AMP interaction with veterinary pathogens, the architecture and training paradigms of BFMs, computational workflows for AMP discovery, and specific constraints imposed by veterinary species (poultry, swine, cattle, companion animals). A particular focus is placed on model generalization across the gram-negative and gram-positive divide, and the validation pipeline from in silico prediction to in vitro susceptibility testing.

1. Introduction

Antimicrobial resistance is a pressing concern in veterinary medicine, driven by subtherapeutic metaphylaxis, high-density production systems, and interspecies transmission of resistant clones [1]. The WHO and WOAH have classified several veterinary pathogens as priority targets for novel intervention strategies. For example, Pasteurella multocida serotypes associated with Fowl Cholera in Poultry and Avian Cholera in Waterfowl have demonstrated resistance to tetracyclines and sulfonamides. Similarly, APEC strains cause Chicken Blood Bacteria: Understanding Avian Pathogenic Escherichia coli (APEC) and Colibacillosis and exhibit extended-spectrum beta-lactamase (ESBL) phenotypes. Clostridium perfringens type A, the etiological agent of Necrotic Enteritis in Broiler Chickens, has acquired resistance to bacitracin and avilamycin in many production regions.

Antimicrobial peptides (AMPs) are evolutionarily conserved effector molecules of the innate immune system, typically 10 to 50 amino acids in length, with an amphipathic architecture that facilitates binding to and disruption of bacterial membranes. The biophysical mechanism of most AMPs involves electrostatic attraction to the anionic phospholipid headgroups (phosphatidylglycerol, cardiolipin) of bacterial membranes, in contrast to the zwitterionic membranes of eukaryotic cells. Upon reaching a threshold local concentration, AMPs adopt a secondary structure (often alpha-helical or beta-hairpin) and insert into the bilayer, leading to pore formation via the barrel-stave, toroidal pore, or carpet model. This mechanism is inherently less prone to target-specific resistance compared to enzyme-active site inhibitors. However, resistance mechanisms do arise, including proteolytic degradation by bacterial proteases, modification of surface charge via MprF-mediated lysinylation of phosphatidylglycerol, and active efflux via resistance-nodulation-division (RND) transporters.

The discovery of AMPs through empirical screening of natural sources (venoms, host secretions, soil microbiomes) is time- and resource-intensive. Biological foundation models, which are deep neural networks pre-trained on large corpora of protein sequences, now enable high-throughput in silico screening, de novo generation, and activity prediction at a fraction of the cost. The term "foundation model" refers to an architecture trained on broad data at scale that can be adapted to a wide range of downstream tasks via fine-tuning or zero-shot inference. In the AMP discovery context, BFMs are trained on millions of peptide sequences (natural and synthetic) and learn a latent representation of sequence-activity relationships that is transferable to novel, veterinary-relevant targets.

2. Biophysical Basis of AMP Activity Against Veterinary Pathogens

2.1 Membrane Composition and Electrostatic Targeting

The outer membrane of gram-negative pathogens such as Pasteurella multocida and E. coli is composed of lipopolysaccharide (LPS) with a highly anionic lipid A core, separated from the inner cytoplasmic membrane by a peptidoglycan layer. Gram-positive bacteria such as Staphylococcus aureus and Clostridium perfringens lack an outer membrane but possess a thick peptidoglycan network. In both cases, the cytoplasmic membrane is the primary target of most AMPs. The net negative charge of bacterial membranes arises from phosphatidylglycerol (PG) and cardiolipin (CL), which are rare in mammalian membranes. Cationic AMPs, typically carrying a net charge of +2 to +9 at physiological pH, are electrostatically attracted to the membrane surface. Hydrophobic residues (leucine, isoleucine, phenylalanine, tryptophan) then mediate insertion into the hydrophobic core.

2.2 Secondary Structure Determinants

The activity of an AMP is strongly correlated with its ability to adopt an amphipathic alpha-helix or a beta-hairpin conformation at the membrane interface. Proline-rich AMPs (PrAMPs) are an exception; they bind to the bacterial ribosome and inhibit translation without membrane lysis. The helical propensity of a peptide can be quantified by the mean hydrophobicity and hydrophobic moment, parameters that BFMs can predict directly from sequence. Disulfide-stabilized beta-sheet AMPs (e.g., defensins) are less susceptible to proteolysis but require correct disulfide bond formation, which complicates synthetic production.

2.3 Modes of Action

1. Barrel-stave model: Helical AMPs assemble as a bundle in the membrane, forming a transmembrane pore with a central water-filled channel. This mode is associated with potent bactericidal activity.
2. Toroidal pore model: The peptide helix inserts and induces continuous curvature of the lipid monolayer, resulting in a pore lined by both lipid headgroups and peptide molecules.
3. Carpet model: Peptides coat the membrane surface in a detergent-like manner, disrupting the bilayer without forming defined pores. This is the least concentration-sensitive mechanism.

For veterinary pathogens, the toroidal pore and carpet mechanisms are most commonly described. The specific activity of an AMP against a given pathogen depends not only on the peptide sequence but also on the lipid composition of the target membrane. For example, Pasteurella multocida membranes are enriched in PG relative to E. coli, which correlates with increased susceptibility to certain cationic AMPs.

3. Biological Foundation Models: Architectural Principles

3.1 Protein Language Models (pLMs)

Protein language models are Transformer-based architectures trained on the masked language modeling objective, analogous to BERT for natural language. The input is a sequence of amino acids, represented as tokens, each with an embedding vector that captures positional and semantic information. The self-attention mechanism computes pairwise interactions between all positions, allowing the model to learn long-range dependencies in sequence space. For AMPs, which are short (fewer than 50 residues), the self-attention head can effectively learn the relationship between N-terminal charge, central hydrophobic face, and C-terminal amidation state.

Key architectural variants include:

• Causal pLMs (e.g., ESM-1b, ProtGPT2): Autoregressive models that predict the next token in a sequence. These can be used for unconditional or conditional generation of novel AMPs.
• Masked pLMs (e.g., ESM-1v, ESM-2): Bidirectional models that infer missing residues. These are suitable for mutation effect prediction and sequence optimization.

3.2 Generative Models for Peptide Design

Generative adversarial networks (GANs) and variational autoencoders (VAEs) have been applied to AMP design. A VAE encodes a peptide sequence into a latent variable (typically a 128- to 512-dimensional Gaussian distribution) and decodes it back into a sequence. By sampling novel latent vectors, the model can generate sequences that lie on the manifold of known AMPs. The Wasserstein GAN (WGAN) with gradient penalty has been used to generate AMPs with high predicted activity and low toxicity. Diffusion models, which iteratively denoise a random tensor into a valid sequence, have shown improved sample diversity.

3.3 Hybrid Models with Physicochemical Embeddings

A limitation of pure sequence-based models is their inability to directly encode secondary structure or membrane interaction energetics. Hybrid models concatenate sequence embeddings with computed biophysical features: net charge, hydrophobicity (Kyte-Doolittle scale), hydrophobic moment, isoelectric point, and amphipathicity score. These features serve as auxiliary inputs to a feedforward classifier that predicts antimicrobial activity.

4. Computational Workflow for AMP Discovery

The typical workflow proceeds through five stages: data curation, model training or fine-tuning, candidate generation, biophysical filtering, and in vitro validation.

flowchart TD
    A[Veterinary Pathogen Genomes\n& AMP Databases (DRAMP, APD3)], > B[Data Curation\nSequence Clustering < 90% Identity\nRemove Hemolytic, Toxic Sequences]
    B, > C[Foundation Model Pre-training\nMasked Language Modeling (ESM-2)\nor VAE Generative Prior]
    C, > D[Downstream Task Head\nAMP Activity Classifier + Minimum Inhibitory\nConcentration (MIC) Regressor]
    D, > E[In Silico Candidate Generation\nLatent Space Sampling / GAN Generation\nor Directed Evolution]
    E, > F[Physicochemical Filtering\nNet Charge >= +2, Hydrophobicity 40-60%\nNo Cysteine-Rich Motifs (for linear AMPs)]
    F, > G[Multi-Target Prediction\nGram-Negative (P. multocida)\nGram-Positive (S. aureus)]
    G, > H{Experimental Validation}
    H, > I[MIC Assay (Broth Microdilution)\nAgainst Field Isolates]
    H, > J[Hemolysis & Cytotoxicity\n(Sheep RBCs, IPEC-J2 Cells)]
    I, > K[Lead Optimization\nPoint Mutation Scanning via pLM\nZero-Shot Fitness Prediction]
    K, > L[Preclinical Testing\nIn Vivo Mouse Sepsis or\nPoultry Colibacillosis Model]

4.1 Data Curation

The primary databases are the Data Repository of Antimicrobial Peptides (DRAMP) and the Antimicrobial Peptide Database (APD3). For veterinary specificity, sequences can be filtered by origin species (e.g., bovine, porcine, avian cathelicidins and defensins). Negative sequences (non-AMPs) are sampled from the UniProt Swiss-Prot database from proteins with no annotated antimicrobial function. Hemolytic activity data from erythrocyte lysis assays are used to filter out toxic candidates. It is critical to remove sequences with >90% identity to avoid training bias.

4.2 Model Architecture and Training

A representative BFM for AMP classification uses a Transformer backbone with 12 layers, 12 attention heads, and embedding dimensions of 768 (ESM-2-scale). The model is pre-trained on a corpus of 200,000 protein sequences from the Swiss-Prot database, then fine-tuned on a curated AMP set of 5,000 positives and 5,000 negatives. The downstream task head is a multilayer perceptron (MLP) with dropout layers assigned to binary classification (AMP or non-AMP) and optionally regression (log10 MIC). The Matthews correlation coefficient (MCC) and the area under the receiver operating characteristic curve (AUROC) are standard evaluation metrics.

4.3 Candidate Generation and Filtering

Generative models produce hundreds of thousands of candidate sequences. The filter cascade removes:

• Sequences less than 8 or greater than 40 amino acids.
• Net charge less than +2 at pH 7.4.
• Hydrophobicity (GRAVY score) outside the -0.5 to +1.5 range.
• Predicted hemolytic activity using a separate hemolysis classification model.
• Sequences with predicted secondary structure that is not amphipathic.

The surviving candidates (typically 100 to 1,000) are then synthesized and tested.

5. Application to Veterinary Pathogens

5.1 Gram-Negative Pathogens: Pasteurella multocida and E. coli

Several studies have used BFMs to identify AMPs active against ESBL-producing E. coli from poultry. The model-derived peptides often target the LPS-binding site on the outer membrane, achieving MIC values in the range of 1 to 8 micrograms per milliliter. For P. multocida, the high PG content of the membrane renders it particularly susceptible to short, highly cationic peptides (net charge +5 to +8). The key challenge is proteolytic stability in the avian respiratory tract, where neutrophil elastase and trypsin-like enzymes may degrade linear peptides. Cyclization or D-amino acid substitution can be suggested by the BFM during the lead optimization phase.

5.2 Gram-Positive Pathogens: Staphylococcus aureus and Clostridium perfringens

Staphylococcus aureus (including methicillin-resistant S. aureus, MRSA) is a major cause of bumblefoot and osteomyelitis in broilers, as discussed in Staphylococcus aureus Bumblefoot and Osteomyelitis in Broilers. The thick peptidoglycan layer of gram-positive cells is not a barrier to AMPs, but the presence of teichoic acids (wall teichoic acids and lipoteichoic acids) can alter net surface charge. BFMs trained on gram-positive AMPs learn to prefer sequences with a higher proportion of tryptophan and arginine, which are essential for crossing the peptidoglycan mesh.

Clostridium perfringens type A, the agent of necrotic enteritis, produces a netB pore-forming toxin that disrupts the intestinal epithelium. BFMs designed for this pathogen must account for the low redox potential and high proteolytic activity of the anaerobic intestinal environment. Proline-rich AMPs, which are resistant to trypsin and chymotrypsin, have emerged as promising candidates from BFM-designed screening campaigns.

5.3 Mycoplasma and Intracellular Pathogens

Mycoplasma bovis, discussed in Mycoplasma bovis in Feedlot Cattle, and Mycoplasma synoviae lack a cell wall, requiring AMPs that target the membrane directly. BFMs can be fine-tuned specifically on peptides active against wall-less bacteria. For intracellular pathogens such as Rhodococcus equi in foals (see Rhodococcus equi Foal Pneumonia) and Lawsonia intracellularis (see Porcine Proliferative Enteropathy), the AMP must penetrate the host cell membrane, which typically requires a cell-penetrating peptide (CPP) motif. Recent BFM architectures have incorporated CPP prediction modules to enable dual function: host cell entry and intracellular bacterial killing.

6. Validation Pipeline and In Vitro Assays

6.1 Minimum Inhibitory Concentration (MIC)

The gold standard for AMP activity is the broth microdilution assay performed in cation-adjusted Mueller-Hinton broth (CAMHB). Veterinary field isolates rather than laboratory reference strains should be used, as passaged strains often lose resistance plasmids. The MIC is defined as the lowest peptide concentration that inhibits visible growth after 16 to 20 hours at 37 degrees Celsius.

6.2 Hemolysis and Cytotoxicity

Veterinary AMPs must be evaluated against species-matched erythrocytes. For avian peptides, chicken and turkey red blood cells are used. For porcine peptides, pig erythrocytes. The concentration at which 10% hemolysis occurs (HC10) should be at least 10-fold above the MIC to ensure a therapeutic index.

6.3 Stability Assays

Peptide stability in serum and in gastrointestinal fluid (for enteric pathogens) is assessed by incubating the AMP at body temperature and measuring residual activity by MIC at time points. Resistance to bacterial proteases (e.g., P. aeruginosa elastase, S. aureus aureolysin) can also be screened.

7. Challenges and Computational Limitations

7.1 Generalization across Pathogen Diversity

A BFM trained primarily on human-associated AMPs may fail to capture the specific lipid composition of veterinary pathogens. For example, the membrane of Gallibacterium anatis (see Gallibacterium anatis in Laying Hens) has a high proportion of lysyl-PG, which reduces net negative charge and reduces susceptibility to cationic peptides. Transfer learning from a small set of veterinary MIC data is necessary but not always sufficient to correct the bias.

7.2 Hallucination and Inactive Sequences

Generative models can hallucinate sequences that appear plausible but lack antimicrobial activity. This is a known failure mode of VAEs and GANs. Physics-based filters and molecular dynamics simulations of membrane insertion can reduce the false positive rate, but they are computationally expensive. Coarse-grained Martini simulations of a single peptide-membrane system require 1,000 to 10,000 CPU hours.

7.3 Resistance Evolution

Long-term exposure to sub-MIC levels of AMPs can select for resistant mutants in veterinary settings. The evolution of resistance is more likely against AMPs with a single mode of action (e.g., ribosomal inhibition) than against membrane-active AMPs. BFMs can be used to predict resistance emergence by scanning for mutations that reduce predicted activity. Multi-target AMPs, which bind to both the membrane and an intracellular target (e.g., DNA gyrase), are considered less prone to resistance.

8. Future Directions

The next generation of BFMs will integrate multi-omics data (transcriptomics, lipidomics, proteomics) of target pathogens to predict AMP susceptibility at the strain level. A "digital twin" of a P. multocida or E. coli strain could be used to simulate AMP-membrane interactions before wet-lab validation. Additionally, federated learning across veterinary diagnostic laboratories could enable model training on proprietary MIC data without compromising data confidentiality.

9. Conclusion

Biological foundation models represent a paradigm shift in antimicrobial peptide discovery for veterinary medicine. By combining the power of Transformer-based sequence modeling with the well-established biophysics of membrane disruption, these models can rapidly generate and optimize peptides that are active against the most problematic veterinary pathogens, including Pasteurella multocida, E. coli, Staphylococcus aureus, Clostridium perfringens, and Mycoplasma species. The computational pipeline from data curation to candidate generation is mature enough for adoption by veterinary research centers. Future integration with molecular dynamics simulation and species-specific toxicity prediction will further improve the success rate of in silico designs.

References

[1] WOAH Technical Report on Antimicrobial Resistance in Veterinary Pathogens. Paris: World Organisation for Animal Health; 2023.

[2] Brogden KA. Antimicrobial peptides: pore formers or metabolic inhibitors in bacteria? Nature Reviews Microbiology. 2005;3(3):238-250.

[3] Meade KG, O'Farrelly C. Beta-defensins: what are they really doing in the innate immune response of the pig? Developmental and Comparative Immunology. 2019;93:1-10.

[4] Torres MDT, Sothiselvam S, Lu TK, de la Fuente-Nunez C. Peptide design principles for antimicrobial applications. Journal of Medicinal Chemistry. 2021;64(9):5646-5667.

[5] Rives A, Meier J, Sercu T, Goyal S, Lin Z, Liu J, et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proceedings of the National Academy of Sciences. 2021;118(15):e2016239118.

[6] Ma Y, Guo Z, Xia B, Zhang Y, Liu X, Yu Y, et al. Identification of antimicrobial peptides from the human gut microbiome using deep learning. Nature Biotechnology. 2022;40(6):921-931.