Duy Huynh Anh, Srisongkram Tarapong
Graduate School in the Program of Research and Development in Pharmaceuticals, Faculty of Pharmaceutical Sciences, Khon Kaen University, Khon Kaen 40002, Thailand.
Department of Health Sciences, College of Natural Sciences, Can Tho University, Can Tho 900000, Vietnam.
ACS Omega. 2025 Aug 19;10(34):38653-38674. doi: 10.1021/acsomega.5c03510. eCollection 2025 Sep 2.
Dengue virus remains a significant global health threat, imposing a substantial disease burden on nearly half of the world's population. The urgent need for effective antiviral therapeutics, including therapeutic peptides targeting the Dengue virus, is critical in the current healthcare landscape. However, the availability of anti-Dengue peptides (ADPs) data remains limited in existing data sets, posing a challenge for computational modeling and discovery. This study presents a novel multimodal framework integrating high-performance predictive modeling with generative learning to accurately predict and potentially identify novel potent ADPs. Specifically, a predictive model was constructed using a multimodal combination of bidirectional long short-term memory (BiLSTM) and a stacking ensemble of neural networks, both using diverse sequence representations. Additionally, a Wasserstein generative adversarial network with a gradient penalty was employed to generate novel ADP candidates. The predictive models demonstrated robust performance, achieving balanced accuracy, area under the receiver operating characteristic curve, and area under the precision-recall curve exceeding 90%, with a Matthews correlation coefficient surpassing 80%. In addition, glycine (G), phenylalanine (F), and tryptophan (W) are the most influential residues to the inhibitory potency of ADPs. Through the proposed multimodal framework, 33 novel ADP sequences with the highest predictive probabilities were identified. Furthermore, regression analysis using a random forest model was developed to predict three candidate peptides with predicted IC values below 10 μM, specifically targeting the envelope protein of the Dengue virus. These findings underscore the effectiveness of the multimodal BiLSTM-based prediction models and stacking neural networks integrating convolutional neural networks, BiLSTM, and transformer architectures in accurately modeling ADP activity. The proposed approach may enhance the discovery pipeline for peptide-based antivirals and contribute to the development of promising therapeutic candidates against the Dengue virus. To facilitate practical application, a publicly available web server for ADP prediction has been deployed at https://antidengue-peptide-predictor.streamlit.app.
登革病毒仍然是全球重大的健康威胁，给世界近一半人口带来了沉重的疾病负担。在当前的医疗环境下，迫切需要有效的抗病毒疗法，包括针对登革病毒的治疗性肽。然而，现有数据集中抗登革肽（ADP）的数据仍然有限，这给计算建模和发现带来了挑战。本研究提出了一种新颖的多模态框架，将高性能预测建模与生成学习相结合，以准确预测并潜在地识别新型强效ADP。具体而言，使用双向长短期记忆（BiLSTM）和神经网络堆叠集成的多模态组合构建了一个预测模型，两者均使用不同的序列表示。此外，采用具有梯度惩罚的瓦瑟斯坦生成对抗网络来生成新型ADP候选物。预测模型表现出强大的性能，平衡准确率、受试者工作特征曲线下面积和精确召回率曲线下面积均超过90%，马修斯相关系数超过80%。此外，甘氨酸（G）、苯丙氨酸（F）和色氨酸（W）是对ADP抑制效力最有影响的残基。通过所提出的多模态框架，确定了33个预测概率最高的新型ADP序列。此外，开发了使用随机森林模型的回归分析来预测三种预测IC值低于10 μM的候选肽，这些肽专门针对登革病毒的包膜蛋白。这些发现强调了基于多模态BiLSTM的预测模型以及集成卷积神经网络、BiLSTM和Transformer架构的堆叠神经网络在准确模拟ADP活性方面的有效性。所提出的方法可能会加强基于肽的抗病毒药物的发现流程，并有助于开发有前景的抗登革病毒治疗候选物。为便于实际应用，已在https://antidengue-peptide-predictor.streamlit.app上部署了一个用于ADP预测的公开网络服务器。
