Multimodal Deep Learning for Generating Potential Anti-Dengue Peptides.

Area of Science:

• Computational biology
• Drug discovery
• Virology

Background:

• Dengue virus poses a significant global health threat, necessitating effective antiviral therapies.
• Limited data on anti-Dengue peptides (ADPs) hinders computational drug discovery efforts.
• There is a critical need for novel therapeutic peptides targeting the Dengue virus.

Purpose of the Study:

• To develop a multimodal computational framework for predicting and identifying novel potent anti-Dengue peptides (ADPs).
• To leverage high-performance predictive modeling and generative learning for ADP discovery.
• To create a publicly accessible tool to aid in the development of Dengue virus therapeutics.

Main Methods:

• Constructed a predictive model using a multimodal combination of bidirectional long short-term memory (BiLSTM) and a stacking ensemble of neural networks with diverse sequence representations.
• Employed a Wasserstein generative adversarial network with a gradient penalty to generate novel ADP candidates.
• Utilized random forest regression analysis to predict the inhibitory potency (IC50) of candidate peptides.

Main Results:

• Predictive models achieved high performance, with balanced accuracy, AUC-ROC, and AUC-PR exceeding 90%, and MCC over 80%.
• Identified glycine (G), phenylalanine (F), and tryptophan (W) as key residues influencing ADP inhibitory potency.
• Discovered 33 novel ADP sequences with high predictive probabilities and identified three candidate peptides with predicted IC50 values below 10 μM, targeting the Dengue virus envelope protein.

Conclusions:

• The multimodal framework effectively models ADP activity, enhancing the discovery pipeline for peptide-based antivirals.
• The study provides promising therapeutic candidates against the Dengue virus.
• A publicly available web server (https://antidengue-peptide-predictor.streamlit.app) facilitates practical application and further research.