AVPpred-BWR: antiviral peptides prediction via biological words representation

Zhuoyu Wei, Yongqi Shen, Xiang Tang, Jian Wen, Youyi Song, Mingqiang Wei, Jing Cheng, Xiaolei ZhuBioinformatics 2025 · DOI

TL;DR

This paper introduces AVPpred-BWR, a novel deep learning model for predicting antiviral peptides (AVPs) by representing peptide sequences as "biological words" derived from their secondary structures. This approach significantly improves prediction accuracy compared to existing methods, offering a more efficient way to discover potential antiviral drugs.

What Problem Does This Paper Address?

The discovery of effective antiviral drugs is a critical challenge, especially in the face of emerging viral threats like COVID-19. Traditional experimental methods for identifying antiviral peptides (AVPs) are slow and resource-intensive. While computational methods have advanced, they often struggle with accuracy or overlook crucial biological information. Specifically, existing machine learning models for AVP prediction frequently rely on hand-crafted features or general protein language models, potentially missing the nuanced information encoded in the secondary structures of peptides, which are known to be important for their function. This paper aims to bridge this gap by developing a more accurate and biologically informed computational approach for AVP prediction.

Methods

The core innovation of AVPpred-BWR lies in its "biological words" representation. Recognizing that AVPs primarily consist of alpha-helices and loops, the authors segment peptide sequences into "words" of length 1 (representing loops) and length 4 (representing alpha-helices). These biological words are then used to train Word2Vec models, generating embeddings that capture sequential information.

These embeddings are fed into a hybrid neural network architecture. This architecture features two parallel channels:

  1. Channel 1: A combination of 1D Convolutional Neural Networks (CNNs) and Transformer encoders.
  2. Channel 2: Exclusively Transformer encoders.

The outputs from these two channels are concatenated and passed through a fully connected layer for the final prediction. This multi-scale, multi-channel approach aims to leverage both local patterns (CNNs) and long-range dependencies (Transformers) from the biological word embeddings.

Datasets

The study utilized several datasets for training and evaluation:

  • Primary Training/Testing Dataset: Derived from the Pinacho-Castellanos et al. dataset, this dataset was curated to contain 2944 positive AVPs and 2944 negative samples. 80% was used for training, and the remaining 20% for independent testing.
  • External Test Sets: To assess generalization, three external test sets were used:
  • A set of 475 AVPs collected from updated databases (DRAMP 4.0, dbAMP 3.0, APD3, DBAASP v3).
  • A dataset created by Xu et al. (2024), containing 16 positive and 16 negative samples, specifically focused on anti-coronavirus peptides (ACVPs).
  • An imbalanced dataset with 1230 positive and 10,771 negative samples.
  • Transferability Datasets: The pipeline was also extended to predict anti-cancer peptides (ACPs) using a dataset of 970 positive and negative samples from AntiCP 2.0, and DPP-IV inhibitory peptides using a dataset of 665 positive and negative samples from iDPPIV-SCM.

Key Insights and Ideas

  • Biological Words as a Novel Representation: The concept of segmenting peptides into "biological words" based on secondary structure motifs (1-mers for loops, 4-mers for alpha-helices) is a novel approach that captures biologically relevant information more effectively than generic sequence representations.
  • Complementarity of 1-mer and 4-mer: Combining embeddings from both 1-mer and 4-mer representations proved superior to using other k-mer combinations, suggesting that these specific granularities capture complementary information crucial for AVP prediction.
  • CNN-Transformer Synergy: The hybrid CNN-Transformer architecture effectively leverages the strengths of both models – CNNs for local feature extraction and Transformers for capturing long-range dependencies – leading to improved performance.
  • Secondary Structure Importance: The analysis of secondary structure composition in the datasets (a-helix and loop being dominant) provides strong biological justification for the chosen "biological words" approach.

Key Results

  • Improved Prediction Accuracy: AVPpred-BWR achieved state-of-the-art performance on the independent test set, with improvements of 4.6% in AUROC and 11.0% in MCC compared to the next best competitor (UniDL4BioPep).
  • Robust Generalization: The model demonstrated strong performance across multiple external test sets, indicating its robustness and ability to generalize to diverse AVP datasets. For instance, it achieved an AUC of 95.3% on the primary independent test set and 97.0% on one external test set.
  • Effective Transferability: The AVPpred-BWR framework was successfully adapted to predict anti-cancer peptides and DPP-IV inhibitory peptides, outperforming existing models on these tasks as well.
  • Model Interpretation: T-SNE plots showed that while initial embeddings for 1-mer and 4-mer were mixed, the features extracted by the fully connected layer clearly separated positive and negative samples, indicating effective feature learning. Attention heatmaps also revealed specific positions that the model focuses on.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that while comprehensive datasets were used, the size of the datasets could be larger, and future work could incorporate more experimental data for model refinement.
  • Feature Integration: The current model primarily relies on the proposed biological word embeddings. Future work could explore incorporating other sequential features and structural features to further enhance model performance.
  • Multi-label Classification: The current model focuses on binary classification (AVP vs. non-AVP). Expanding to multi-label classification for different types of AVPs or other bioactive peptides is a promising future direction.
  • Model Interpretability: While some interpretation was performed, further work could delve deeper into understanding the model's decision-making process, especially for misclassified examples.

Summary updated

BibTeX
@article{wei2025avppred,
  author = {Zhuoyu Wei and Yongqi Shen and Xiang Tang and Jian Wen and Youyi Song and Mingqiang Wei and Jing Cheng and Xiaolei Zhu},
  title = {{AVPpred-BWR: antiviral peptides prediction via biological words representation}},
  journal = {Bioinformatics},
  year = {2025},
  doi = {10.1093/bioinformatics/btaf126},
}
Figure