DeepEAG: A deep learning-based hybrid framework for identifying epilepsy-associated genes using a stacking strategy.

As one of the most common neurological diseases in the world, epilepsy can be caused through the deletion or duplication of known epilepsy-associated genes. Therefore, it is crucial to accurately identify epilepsy-associated genes to gain a deeper understanding of the pathogenesis of epilepsy and to develop new therapies. Although a number of wet-lab experimental methods have been proposed, they are reliable but in general time-consuming and labor-intensive. Hence, their practical application in epilepsy-associated genes identification is quite limited. To address the existing limitations, we propose a computational method, called DeepEAG, for the identification of epilepsy-associated genes throughout the human genome. To develop DeepEAG, we first constructed a new and, to our knowledge, the most comprehensive epilepsy genomic benchmark dataset. Afterwards, we investigated four feature encoding algorithms with different perspectives and trained them using well-established traditional classifiers and deep learning classifier, resulting in 13 baseline models. Finally, the stacking strategy is effectively utilized by integrating the predicted output of the optimal baseline models and training with xgboost. In results, the DeepEAG predictor obtained excellent performance on the cross-validation with accuracy and AUC of 0.835 and 0.907, respectively. Overall, DeepEAG showed more stable and accurate predictive performance compared to baseline models, and also outperforms other ensemble strategies, demonstrating the effectiveness of our proposed hybrid framework. In addition, DeepEAG is anticipated to facilitate community-wide efforts to identify putative epilepsy-associated genes and generate new biological insights and testable hypotheses. DeepEAG is freely avaliable at https://github.com/JfXie/DeepEAG.