Abstract
Antimicrobial peptides (AMPs) are attractive candidates to combat antibiotic resistance for their capability to target bio-membranes and restrict a wide range of pathogens. It is a daunting challenge to discover novel AMPs due to their sparse distributions in a vast peptide universe, especially for peptides that demonstrate potencies for both bacterial membranes and viral envelopes. Here we establish a de novo AMP design framework by bridging a deep generative module and a graph-encoding activity regressor. The generative module learns hidden ‘grammars’ of AMP features and produces candidates sequentially pass antimicrobial predictor and antiviral classifiers. We discover three bifunctional AMPs and experimentally validated their abilities to inhibit a spectrum of pathogens in vitro and in animal models. Notably, P076 is a highly potent bactericide with the minimal inhibitory concentration of 0.21 μM against multidrug-resistant A. baumannii, while P002 broadly inhibits five enveloped viruses. Our study provides feasible means to uncover sequences that simultaneously encode antimicrobial and antiviral activities, thus bolstering the function spectra of AMPs to combat a wide range of drug-resistant infections.
Introduction
We are rapidly entering a "post-antibiotic era" where ESKAPE pathogens—a group of highly resistant bacteria—render traditional treatments increasingly ineffective. Antimicrobial peptides (AMPs) represent a promising alternative because they physically disrupt microbial membranes, a mechanism that makes it significantly harder for pathogens to develop resistance. However, the "peptide universe" is unfathomably vast, and functional AMPs are sparsely distributed within it. Most existing AI models focus solely on antibacterial activity. This research addresses the challenge of de novo design—creating entirely new sequences rather than just tweaking nature's templates—and specifically targets "bifunctional" peptides that can tackle both bacterial infections and viral threats like SARS-CoV-2 or Zika.
Methodology
The researchers developed a two-part pipeline to navigate the complex peptide space:
- The Generator (GAN): A Generative Adversarial Network was trained on known AMP sequences to learn the "grammar" of functional peptides. It uses Amino Acid Factors (AAF) to encode physicochemical properties (like charge and hydrophobicity) into a low-dimensional space, allowing it to sample and create novel sequences.
- The Regressor (AMPredictor): This "filter" uses a Graph Convolutional Network (GCN) to predict the Minimal Inhibitory Concentration (MIC) of the generated peptides. Unlike standard models that treat peptides as simple strings, AMPredictor treats them as graphs where nodes are amino acid residues (represented by ESM protein language model embeddings) and edges represent predicted physical contact maps.
- Bifunctional Filtering: Generated sequences were passed through AMPredictor for bacterial potency and then through a suite of five existing antiviral peptide (AVP) classifiers. Only those that "passed" both were synthesized for lab testing.
The training utilized 3,280 AMP sequences from the PepVAE/GRAMPA repository (inhibiting E. coli) and approximately 2,000 sequences from AVPdb for antiviral validation.
Key Insights
A core finding is the effectiveness of leveraging structural similarities between bacterial membranes and viral envelopes. By designing for high-affinity membrane binding, the model naturally identifies sequences effective against both pathogen types. Furthermore, representing peptides as 3D-aware graphs provides superior predictive power compared to 1D sequence models. Despite a relatively small training set, the GAN generated highly novel peptides with only ~35% similarity to known sequences, demonstrating that the model learned underlying biological "rules" rather than memorizing data.
Results
The framework successfully identified two standout bifunctional peptides:
- P076 (The Bactericide): Showed a staggering MIC of 0.21 μM against multidrug-resistant A. baumannii. In lethal infection mouse models, P076 provided a 50% survival rate and proved less nephrotoxic than Polymyxin B, a current "last-resort" antibiotic.
- P002 (The Antiviral): Demonstrated broad-spectrum activity, neutralizing five enveloped viruses: CHIKV, HTNV, DENV-2, HSV-1, and SARS-CoV-2.
AMPredictor achieved a Pearson correlation coefficient of 0.7072, significantly outperforming standard deep learning baselines.
Conclusion and Future Work
While the results are promising, the study is limited by a small experimental sample size. Future work aims to incorporate antiviral sequences directly into the generative training and expand testing to a larger library of ~1,000 new designs with even lower similarity to known AMPs. This approach provides a feasible path to uncovering sequences that bolster the functional spectra of AMPs to combat diverse drug-resistant infections.
Summary updated