Document Type : Original Article
Authors
1
Information Systems Department, Bioinformatics Program, Faculty of Computer and Information Sciences, Ain Shams University.
2
Information Systems, Faculty of Computer and Information Sciences, Ain Shams University
3
Vice Dean for Postgraduate Studies & Research, Faculty of Computer and Information Sciences, Ain Shams University
4
Department of Information Systems, Faculty of Computer and Information Sciences, Ain Shams University, Cairo, 11566, Egypt
Abstract
Bioinformaticians face an issue in analyzing and studying protein similarity as the number of proteins grows. Protein sequence analysis helps in the prediction of protein functions. It is critical for the analysis process to be able to appropriately categorize proteins based on their sequences. The extraction of features from protein sequences is done using a variety of methods. The goal of this study is to investigate the different variations of data on the classification performance of a deep learning model employing 3D data. First, few research questions were formulated regarding the impact of the following criteria: dataset size, IMF importance, feature size, and preprocessing on the proposed deep learning classification process. Second, comprehensive experiments were conducted to answer the research questions. Six feature extraction methods were utilized to create 3D features with two sizes (7x7x7 and 9x9x9), which were then fed into a convolutional neural network. Three datasets different in their sorts, sizes, and balance state were used. Accuracy, precision, recall and F1-score are the standard assessment metrics used. Experimental results draw significant conclusions. First, the 7x7x7 feature matrix has a positive correlation between its dimensions, which improved the results. Second, using the sum of the first three IMF components had better impact than using the first IMF component. Third, the classification process did not benefit from the normalization of features for small datasets unlike the large dataset. Finally, the dataset size had a significant impact on training the CNN model, with a training accuracy reaching 84.03%.
Keywords
)
View on SCiNiTO