A hybrid framework of generative deep learning for antiviral peptide discovery.

Abstract

Antiviral peptides (AVPs) hold great potential for combating viral infections, yet their discovery and development remain challenging. In this study, we present a hybrid model combining Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) and Bidirectional Long Short-Term Memory (BiLSTM) networks to address these challenges. The BiLSTM model was thoroughly constructed and validated, showing reliable performance in identifying AVPs. Additionally, analyses such as applicability domain and feature importance from BiLSTM provided valuable insights into the generated peptides. The performance of the WGAN-GP model was comprehensively evaluated, confirming its ability to produce diverse and functional peptide sequences. The model successfully generated and identified 815 novel AVPs, demonstrating its effectiveness in peptide generation and classification. Novel antiviral peptides (AVPs) were successfully identified across all viral endpoints tested. However, their abundance exhibited significant variability, with the highest levels observed in influenza A virus and the lowest levels detected in human parainfluenza virus type 3. These findings highlight the potential of our hybrid approach as a powerful tool for antiviral peptide discovery and contribute to advancing peptide-based therapeutic research. To maximize its impact on AVPs discovery, we have deployed our predictive models as a publicly accessible platform at https://avp-predictor.streamlit.app, offering researchers a practical tool for prioritizing candidate peptides.