Vaidya Hiteshi, Gautam Sakshi, Kumar Manoj
Virology Unit and Bioinformatics Centre, Institute of Microbial Technology, Council of Scientific and Industrial Research (CSIR), Sector 39A, Chandigarh 160036, India.
Academy of Scientific and Innovative Research (AcSIR), Ghaziabad 201002, India.
Comput Struct Biotechnol J. 2025 May 1;27:1784-1799. doi: 10.1016/j.csbj.2025.04.042. eCollection 2025.
Epstein-Barr virus (EBV) is linked to various cancers like gastric carcinoma, nasopharyngeal carcinoma, and Burkitt's lymphoma, leading to around 200,000 deaths annually. Despite efforts, FDA-approved drugs to combat EBV infection are lacking. In this endeavor, we have developed an AI/ML based predictive algorithm "Anti-EBV" to find potential antivirals against EBV. We utilized small molecules from the ChEMBL database, which were experimentally tested for antiviral activity against EBV in lytic phase, in terms of IC /EC values. 17,968 molecular fingerprints and descriptors were computed for each molecule. Further, the best-performing 150 descriptors were used in the predictive model development. The molecules were then split into training/testing (T) and independent validation (V) datasets, followed by 10-fold cross validation to develop robust models. Various machine-learning techniques (MLTs) namely SVM, KNN, ANN, DNN, RF and XGBoost were used for predictive models development. SVM model achieved the best performance with Pearson's correlation coefficient (PCC) of 0.91 on T dataset and 0.95 on V dataset, respectively. These models were found to be robust by applicability domain, decoy dataset and chemical clustering analyses. The top-performing model was used to screen approved drugs from DrugBank, identifying potential repurposed drugs namely arzoxifene, succimer, abemaciclib and many more. To further validate these findings, top compounds were docked against key lytic proteins BZLF1 and BHRF1, demonstrating strong binding affinities for compounds like fluspirilene and suvorexant. This model is accessible as the "Anti-EBV" web server http://bioinfo.imtech.res.in/manojk/antiebv/ for antiviral prediction, making it the first AI/ML-based study for antiviral identification against EBV in lytic phase.
爱泼斯坦-巴尔病毒（EBV）与多种癌症相关，如胃癌、鼻咽癌和伯基特淋巴瘤，每年导致约20万人死亡。尽管人们做出了努力，但目前仍缺乏美国食品药品监督管理局（FDA）批准的用于对抗EBV感染的药物。在这项工作中，我们开发了一种基于人工智能/机器学习的预测算法“Anti-EBV”，以寻找针对EBV的潜在抗病毒药物。我们利用了ChEMBL数据库中的小分子，这些小分子针对裂解期EBV的抗病毒活性进行了实验测试，测试指标为IC/EC值。为每个分子计算了17968个分子指纹和描述符。此外，在预测模型开发中使用了表现最佳的150个描述符。然后将这些分子分为训练/测试（T）数据集和独立验证（V）数据集，接着进行10折交叉验证以开发稳健的模型。使用了各种机器学习技术（MLT），即支持向量机（SVM）、k近邻算法（KNN）、人工神经网络（ANN）、深度神经网络（DNN）、随机森林（RF）和极端梯度提升（XGBoost）来开发预测模型。SVM模型表现最佳，在T数据集上的皮尔逊相关系数（PCC）为0.91，在V数据集上为0.95。通过适用域、诱饵数据集和化学聚类分析发现这些模型是稳健的。使用表现最佳的模型从DrugBank中筛选已批准的药物，确定了潜在的重新利用药物，如阿佐昔芬、二巯基丁二酸、阿贝西利等。为了进一步验证这些发现，将顶级化合物与关键裂解蛋白BZLF1和BHRF1进行对接，结果表明氟司必林和苏沃雷生等化合物具有很强的结合亲和力。该模型可作为“Anti-EBV”网络服务器http://bioinfo.imtech.res.in/manojk/antiebv/访问，用于抗病毒预测，这使其成为第一项基于人工智能/机器学习的针对裂解期EBV进行抗病毒鉴定的研究。
