Investigation of Multi-task Deep Neural Networks in Automated Protein Function Prediction

Rifaioğlu, Ahmet Süreyya
Martin, Maria Jesus
Atalay, Rengül
Atalay, Mehmet Volkan
Doğan, Tunca
Functional annotation of proteins is a crucial research field for understanding molecular mechanisms of living-beings and for biomedical purposes (e.g. identification of disease-causing functional changes in genes and for discovering novel drugs). Several Gene Ontology (GO) based protein function prediction methods have been proposed in the last decade to annotate proteins. However, considering the prediction performances of the proposed methods, it can be stated that there is still room for significant improvements in protein function prediction area (1). Deep learning techniques became popular in recent years and turned out to be an industry standard in several areas such as computer vision and speech recognition. To the best of our knowledge, as of today, deep learning algorithms have not been applied to the large-scale protein function prediction problem. Here, we propose a hierarchical multi-task deep neural network architecture, DEEPred, as a solution to protein function prediction problem. First of all, we investigated the potential of employing deep learning methods for protein function prediction. For this purpose, we measured the performance of our models at different parameter settings. Furthermore, we examined the relationship between the performance of the system and the size of the training datasets, since the training set size has been reported in the literature to be significantly affecting the performance of deep learning models.
Citation Formats
A. S. Rifaioğlu, M. J. Martin, R. Atalay, M. V. Atalay, and T. Doğan, “Investigation of Multi-task Deep Neural Networks in Automated Protein Function Prediction,” presented at the ISMB/ECCB 2017: 25th Annual International Conference on Intelligent Systems for Molecular Biology, (21 - 25 Temmuz 2017), Prag, Çek Cumhuriyeti, 2017, Accessed: 00, 2021. [Online]. Available: