Diff-AMP: tailored designed antimicrobial peptide framework with all-in-one generation, identification, prediction and optimization

Rui Wang, Tao Wang, Linlin Zhuo, Jinhang Wei, Xiangzheng Fu, Quan Zou, Xiaojun YaoBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces Diff-AMP, a comprehensive deep learning framework that automates the design, identification, and optimization of antimicrobial peptides (AMPs). By integrating diffusion models, attention mechanisms, and reinforcement learning, Diff-AMP addresses key limitations in current AMP research, enabling the generation of novel, highly effective AMP candidates with tailored properties.

What Problem Does This Paper Address?

The rise of antibiotic-resistant "superbugs" has created an urgent need for new antibacterial agents. Antimicrobial peptides (AMPs) are promising candidates due to their low toxicity and ability to target bacterial membranes, making resistance less likely. While deep learning has shown promise in generating AMPs, existing methods often overlook crucial aspects:

  • Interdependencies: They fail to fully capture the complex relationships between amino acids within a peptide sequence.
  • Integrated Workflow: There's no single framework that seamlessly handles AMP generation, identification, attribute prediction, and iterative optimization.
  • Customization: Current models struggle to generate AMPs with specific, user-defined properties.

This paper aims to bridge these gaps by developing an all-in-one solution that streamlines and enhances the entire AMP discovery pipeline.

Methods

Diff-AMP is built upon a modular architecture, integrating several advanced deep learning techniques:

  • AMP Generation (RLGen): This module uses a novel approach combining a thermodynamic diffusion model with reinforcement learning. The diffusion model, guided by an attention mechanism, effectively models the interdependence between amino acids, leading to more diverse and biologically active AMP candidates.
  • AMP Identification: To accurately identify promising AMPs from generated candidates, the framework leverages transfer learning with the ESM2 pre-trained protein language model. This allows the model to learn universal features from vast protein datasets, improving its generalization performance for AMP recognition.
  • AMP Attribute Prediction: A convolutional neural network (CNN) with a multi-hybrid embedding strategy is employed to predict various AMP attributes (e.g., antibacterial, anticancer). This module uses multiple encoding methods (one-hot, BLOSUM62, AAIndex, PAAC) to capture rich sequence information. To speed up training, mixed-precision training and R-Drop regularization are utilized.
  • Iterative Optimization: A reinforcement learning-based strategy is used for multi-objective iterative optimization. This allows researchers to guide the generation process towards AMPs with specific desired attributes and lengths, enabling customized AMP design.

The framework operates in an iterative loop (C-E-F-G-H-C), where generated AMPs are evaluated, predicted for attributes, and refined based on user-defined criteria.

Datasets

The study utilizes three main datasets:

  • Dataset1: Sourced from the CAMP server, containing 8225 AMP sequences, primarily used for AMP generation and iterative optimization training.
  • Dataset2: A balanced dataset of 8268 samples (4134 AMPs and 4134 non-AMPs), used for training the AMP recognition model. This dataset highlights distinct positional preferences for specific amino acids (K in AMPs, L in non-AMPs).
  • Dataset3: Derived from a previous study, this dataset comprises 20 types of AMPs with diverse properties, used for training the AMP multi-attribute prediction module.

Key Insights and Ideas

  • Synergy of Diffusion and Attention: Integrating diffusion models with attention mechanisms is a novel approach for capturing complex amino acid interdependencies in AMP generation, leading to more biologically relevant sequences.
  • Pre-training is Key for Identification: Leveraging large pre-trained protein language models like ESM2 significantly boosts AMP identification accuracy by providing robust general feature representations.
  • Iterative Optimization for Customization: Reinforcement learning provides a powerful mechanism to iteratively refine AMPs, allowing for the generation of peptides with specific, user-defined attributes, a critical step towards personalized drug design.
  • All-in-One Framework Advantage: Consolidating generation, identification, prediction, and optimization into a single framework streamlines the AMP discovery process, saving time and resources.

Key Results

  • AMP Recognition: The Diff-AMP model achieved state-of-the-art performance in AMP recognition, outperforming several existing models across metrics like ACC (86.92%), PRE (91.48%), SEN (83.02%), SPE (91.44%), MCC (74.07%), F1-Score (86.39%), and AUC (92.22%).
  • AMP Generation: The RLGen module successfully generated diverse AMP candidates, including those with a-helical, β-helical, and extended structures, mimicking known active AMPs.
  • Attribute Prediction: The multi-attribute prediction module demonstrated the ability to predict a broad range of AMP properties, with generated AMPs showing comparable or even enhanced antibacterial properties compared to real AMPs.
  • Efficiency: The use of mixed-precision training and R-Drop significantly improved training efficiency without substantial loss in accuracy.

Limitations and Future Work

  • Multi-modal Information: The current framework primarily relies on AMP sequences. The authors acknowledge the limitation of not analyzing multi-modal information such as AMP structures and interaction networks.
  • Future Directions:
    • Integrate information from multiple modalities (sequences, structures, interaction networks) to further improve module performance.
  • Enhance the overall Diff-AMP framework through collaborative work.
  • The framework has been deployed as a web server, making it accessible for researchers.

Summary updated

BibTeX
@article{wang2024diff,
  author = {Rui Wang and Tao Wang and Linlin Zhuo and Jinhang Wei and Xiangzheng Fu and Quan Zou and Xiaojun Yao},
  title = {{Diff-AMP: tailored designed antimicrobial peptide framework with all-in-one generation, identification, prediction and optimization}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae078},
}
Figure