Recent Progress in the Discovery and Design of Antimicrobial Peptides Using Traditional Machine Learning and Deep Learning

Jielu Yan, Jianxiu Cai, Bob Zhang, Yapeng Wang, Derek F. Wong, Shirley W. I. SiuAntibiotics 2022 · DOI

TL;DR

The rise of antimicrobial resistance necessitates new therapeutic strategies, and antimicrobial peptides (AMPs) show great promise. This review surveys recent advancements in using machine learning (ML) and deep learning (DL) to discover and design AMPs, highlighting DL's growing importance in overcoming the limitations of traditional experimental methods.

What Problem Does This Paper Address?

The escalating threat of antimicrobial resistance (AMR) demands novel antibiotics. Antimicrobial peptides (AMPs) are a promising class of natural compounds with broad-spectrum activity and unique mechanisms of action that make resistance development difficult. However, discovering and designing new AMPs through traditional experimental methods is costly and time-consuming. This paper addresses the need for efficient in silico approaches, specifically focusing on the recent progress and methodologies in applying traditional ML and, more importantly, deep learning (DL) techniques for AMP discovery and design.

Methods

This review provides a comprehensive overview of ML and DL methods applied to AMP discovery and design. It covers:

  • Feature Encoding: Discusses methods to represent peptide sequences numerically, categorized into peptide-level (sequence-based and structure-based) and amino acid-level features. Techniques like one-hot encoding, amino acid composition, word embeddings (Word2vec, GloVe), and contextual embeddings (BERT) are explained.
  • Traditional ML: Briefly touches upon established methods like Support Vector Machines (SVM), Random Forests (RF), k-Nearest Neighbors (kNN), and ensemble learning.
  • Deep Learning Architectures: Details various DL approaches, including:
  • Deep Neural Networks (DNNs): Fully connected layers for learning complex patterns.
  • Convolutional Neural Networks (CNNs): Effective for extracting local features and reducing dimensionality.
  • Recurrent Neural Networks (RNNs) and variants (LSTM, Bi-LSTM, GRU): Suitable for sequential data, capturing long-term dependencies.
  • Hybrid Models: Combinations of CNNs and RNNs, or DL with traditional ML, to leverage the strengths of different architectures.
  • Attention Mechanisms: Enhancing models by focusing on important parts of the sequence.
  • Generative Models (VAEs, GANs): Used for de novo design of novel peptide sequences.
  • AMP Design Strategies: Reviews methods for optimizing known AMPs (template-based) and generating entirely new sequences (de novo design).

Datasets

The paper references several databases and datasets used for training and evaluation, including:

  • Antimicrobial Peptide Database (APD/APD3): A manually curated database of AMPs with experimentally validated activity.
  • dbAMP and DBAASP: Databases containing validated AMPs and artificial AMPs.
  • UniProt: A large protein sequence database used for pre-training language models.
  • UniRef100 and UniRef50: Protein sequence databases used for pre-training language models.
  • Various benchmark datasets: Specific datasets are mentioned in the context of different prediction tasks (e.g., anticancer peptides, antiviral peptides).

The paper does not introduce a new benchmark dataset but rather reviews existing work that utilizes these established resources.

Key Insights and Ideas

  • DL's Power in Feature Learning: Deep learning models excel at automatically learning complex, high-level features from raw peptide sequences, reducing the reliance on manual feature engineering required by traditional ML.
  • Sequence as Language: The analogy between peptide sequences and natural language has paved the way for applying powerful Natural Language Processing (NLP) techniques (like word embeddings and transformers) to AMP research, leading to more effective representations.
  • Hybrid Approaches Offer Synergy: Combining different ML/DL architectures (e.g., CNNs for feature extraction and LSTMs for sequence modeling) or integrating traditional ML with DL can often yield superior performance by leveraging complementary strengths.
  • De Novo Design is the Frontier: While predicting existing AMPs is crucial, the ultimate goal is to design novel peptides with enhanced properties. Generative models like VAEs and GANs are showing significant promise in this area.
  • Data Scarcity is a Major Hurdle: Despite advancements, the limited size and potential inconsistencies of AMP datasets remain a significant challenge for training robust DL models.

Key Results

The paper summarizes results from various studies, highlighting the performance of DL models:

  • High Prediction Accuracy: Many DL models achieve accuracies exceeding 90% for AMP classification tasks, often outperforming traditional ML methods. For example, studies report accuracies of 98.3% for anticancer peptide prediction (ENNAACT) and 92.2% for general AMP prediction (Su et al.).
  • Improved De Novo Design: Generative models have successfully designed novel peptides with significant antimicrobial activity, sometimes showing a 160-fold increase in activity compared to initial templates.
  • Transfer Learning Efficacy: Pre-trained language models and transfer learning approaches have demonstrated strong performance, especially when dealing with limited specific AMP data.
  • Multi-Activity Prediction: Efforts are being made to develop models that can predict multiple bioactivity classes (e.g., antibacterial, antifungal, antiviral) simultaneously.

Limitations and Future Work

The authors identify several key limitations and suggest future directions:

  • Data Insufficiency: The relatively small size of validated AMP datasets compared to other domains (like computer vision or NLP) is a major bottleneck for training complex DL models. Future work should focus on semi-supervised and unsupervised learning, as well as innovative data augmentation techniques.
  • Limited Modeling Beyond Binary Classification: Most current models focus on binary AMP/non-AMP classification. There's a need for more sophisticated models that can predict specific activities (multi-class/multi-label) and regression for precise activity quantification.
  • Drug-Likeness and Toxicity Prediction: Predicting pharmacokinetic properties (like half-life) and toxicity remains challenging due to limited data and non-standardized reporting. Developing in silico tools for these aspects is crucial for therapeutic development.
  • Explainable AI (XAI): Many DL models are "black boxes." Developing XAI methods to understand why a model makes a prediction is essential for gaining biological insights and building trust in the predictions. This is particularly challenging due to the complex nature of peptide representations.
  • Reproducibility: Ensuring the reproducibility of DL models through open-source code and data sharing is critical for scientific progress.
  • Integration of Diverse Data: Future research could benefit from integrating various data types (sequence, structure, experimental assays, biological context) to build more comprehensive models.

Summary updated

Abstract

Antimicrobial resistance has become a critical global health problem due to the abuse of conventional antibiotics and the rise of multi-drug-resistant microbes. Antimicrobial peptides (AMPs) are a group of natural peptides that show promise as next-generation antibiotics due to their low toxicity to the host, broad spectrum of biological activity, including antibacterial, antifungal, antiviral, and anti-parasitic activities, and great therapeutic potential, such as anticancer, anti-inflammatory, etc. Most importantly, AMPs kill bacteria by damaging cell membranes using multiple mechanisms of action rather than targeting a single molecule or pathway, making it difficult for bacterial drug resistance to develop. However, experimental approaches used to discover and design new AMPs are very expensive and time-consuming. In recent years, there has been considerable interest in using in silico methods, including traditional machine learning (ML) and deep learning (DL) approaches, to drug discovery. While there are a few papers summarizing computational AMP prediction methods, none of them focused on DL methods. In this review, we aim to survey the latest AMP prediction methods achieved by DL approaches. First, the biology background of AMP is introduced, then various feature encoding methods used to represent the features of peptide sequences are presented. We explain the most popular DL techniques and highlight the recent works based on them to classify AMPs and design novel peptide sequences. Finally, we discuss the limitations and challenges of AMP prediction.
BibTeX
@inproceedings{yan2022recent,
  author = {Jielu Yan and Jianxiu Cai and Bob Zhang and Yapeng Wang and Derek F. Wong and Shirley W. I. Siu},
  title = {{Recent Progress in the Discovery and Design of Antimicrobial Peptides Using Traditional Machine Learning and Deep Learning}},
  booktitle = {Antibiotics},
  year = {2022},
  doi = {10.3390/antibiotics11101451},
}
Figure