DLFea4AMPGen de novo design of antimicrobial peptides by integrating features learned from deep learning models

Han Gao, Feifei Guan, Boyu Luo, Dongdong Zhang, Wei Liu, Yuying Shen, Lingxi Fan, Guoshun Xu, Wang Yuan, Tao Tu, Ningfeng Wu, Bin Yao, Huiying Luo, Yue Teng, Jian Tian, Huoqing HuangNature Communications 2025 · DOI

Abstract

Deep learning models show promise in accelerating the design and optimization of antimicrobial peptides (AMPs), but current methods face challenges, such as low success rates or large virtual library scales. In this study, we introduce DLFea4AMPGen, a bioactive peptide design strategy that leverages deep learning models to identify and extract key features associated with antimicrobial peptide activity. This approach enables the generation of peptide sequences with potential bioactivities. Using the SHapley Additive exPlanations (SHAP) method, we quantify the contribution of each amino acid in multifunctional peptides with potential antibacterial, antifungal, and antioxidant activities. Key feature fragments (KFFs) with the highest average contributions are extracted and classified into four subfamilies based on amino acid frequency. These high-frequency amino acids are systematically arranged to generate a plausible sequence subspace for candidate peptides, from which 16 representative sequences were selected for experimental validation. The results show that 75% (12/16) of the sequences exhibited at least two types of activity. Notably, D1 exhibits broad-spectrum antimicrobial activity, including efficacy against multidrug-resistant clinical pathogenic isolates both in vitro and in vivo. This proof-of-concept study underscores the potential of the DLFea4AMPGen platform for efficient design and screening of bioactive peptides, showcasing its value in AMP research.

Introduction

The rise of multidrug-resistant (MDR) "superbugs" has made the search for new antibiotics urgent. Antimicrobial peptides (AMPs) are a promising solution because they physically disrupt bacterial membranes, making it harder for microbes to develop resistance. However, designing AMPs de novo is a massive "needle-in-a-haystack" problem. Traditional AI models often act as "black boxes"—they can predict if a sequence might work, but they don't explain why. This leads to low success rates in the lab and unmanageable virtual library scales. Furthermore, most designs ignore the fact that bacterial infections often cause secondary damage through oxidative stress and fungal co-infections. This study aims to design multifunctional peptides that tackle these issues simultaneously using an interpretable deep learning framework.

Methodology

The researchers developed DLFea4AMPGen, a multi-stage pipeline designed to move from simple prediction to feature extraction and targeted generation. The process utilized 20 bioactive peptide datasets totaling approximately 23,346 sequences to fine-tune specialized models.

  1. Transfer Learning: The team utilized MP-BERT, a large-scale protein language model, and fine-tuned it into specialized models to predict antibacterial, antifungal, and antioxidant activities.
  2. Feature Extraction (SHAP): Using SHapley Additive exPlanations (SHAP), the model quantified the contribution of each amino acid at every position. This allowed the identification of Key Feature Fragments (KFFs)—13-amino acid windows with the highest potency.
  3. Subfamily Clustering and Subspace Generation: 3,400 KFFs were clustered into four subfamilies (SF-A through SF-D) based on chemical properties and structural motifs (e.g., $\alpha$-helices vs. $\beta$-sheets). By arranging the most frequent high-contribution amino acids from these subfamilies, the team created a "plausible sequence subspace," significantly narrowing the search for candidate peptides.
The DLFea4AMPGen design pipeline, showing model fine-tuning and SHAP-based feature extraction.

Results and Discussion

The DLFea4AMPGen platform demonstrated a remarkably high validation rate. Out of 16 de novo designed peptides, 12 (75%) exhibited at least two types of bioactivity. Notably, 15 of the 16 peptides showed negligible hemolytic activity, indicating high safety for potential clinical use.

In Vitro and In Vivo Efficacy

Lead candidates D1 and D2 showed powerful broad-spectrum activity against a wide range of pathogens, including clinical MDR isolates of S. aureus and E. coli. In a mouse sepsis model, peptide D1 significantly reduced bacterial loads across multiple organs (liver, spleen, lungs, and kidneys). Beyond its antimicrobial properties, D1 exhibited an anti-inflammatory effect by lowering levels of pro-inflammatory cytokines such as TNF-$\alpha$, IL-1$\beta$, and IL-6, suggesting its utility in managing the host's immune response during infection.

Experimental validation workflow, from primary screening to in vivo mouse sepsis models.

Conclusion

This study demonstrates that integrating interpretability tools like SHAP into deep learning workflows can transform AI from a simple filter into a powerful design blueprint. By extracting the essential "features" of successful peptides, DLFea4AMPGen enables the efficient creation of multifunctional therapeutics. Future work will focus on transitioning from classification to regression models to predict exact Minimum Inhibitory Concentrations (MIC) and further exploring the internal cellular mechanisms of these novel peptides.

Summary updated

Abstract

Deep learning models show promise in accelerating the design and optimization of antimicrobial peptides (AMPs), but current methods face challenges, such as low success rates, or large virtual library scales. In this study, we introduce DLFea4AMPGen, a bioactive peptide design strategy that leverages deep learning models to identify and extract key features associated with antimicrobial peptide activity. This approach enables the generation of peptide sequences with potential bioactivities. Using the SHapley Additive exPlanations (SHAP) method, we quantify the contribution of each amino acid in multifunctional peptides with potential antibacterial, antifungal, and antioxidant activities. Key feature fragments (KFFs) with the highest average contributions are extracted and classified into four subfamilies based on amino acid frequency. These high-frequency amino acids are systematically arranged to generate a plausible sequence subspace for candidate peptides, from which 16 representative sequences were selected for experimental validation. The results show that 75% (12/16) of the sequences exhibited at least two types of activity. Notably, D1 exhibits broad-spectrum antimicrobial activity, including efficacy against multidrug-resistant clinical pathogenic isolates both in vitro and in vivo. This proof-of-concept study underscores the potential of the DLFea4AMPGen platform for efficient design and screening of bioactive peptides, showcasing its value in AMP research.
BibTeX
@inproceedings{gao2025dlfea,
  author = {Han Gao and Feifei Guan and Boyu Luo and Dongdong Zhang and Wei Liu and Yuying Shen and Lingxi Fan and Guoshun Xu and Wang Yuan and Tao Tu and Ningfeng Wu and Bin Yao and Huiying Luo and Yue Teng and Jian Tian and Huoqing Huang},
  title = {{DLFea4AMPGen de novo design of antimicrobial peptides by integrating features learned from deep learning models}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-64378-y},
}
Figure