Hybrid Model Design For Protein Function Prediction

Shenghao Zhao, Xiaoyu Zhang, Ziyuan Zhao, Peisheng Qian, Weide Liu, Zeng Zeng, Bharadwaj Veeravalli, Lingyun Dai, Pär Nordlund, Nayana Prabhu, Tam Wai Leong, Xulei Yang

Abstract: Considering the significant role of protein function probes in medicine development and health monitoring, we design a hybrid model based on traditional and deep learning methods to predict protein functions with desirable accuracy. Our work aims to better utilize the protein sequence information in our hybrid prediction model. Firstly, we introduce the high-efficiency sequence alignment tool DIAMOND to obtain function prediction reference based on sequence homology since "similar" proteins have similar protein functions. Secondly, we adopt deep learning methods to extract features from encoded protein sequences, then combine sequence features with domain features and protein-protein interaction (PPI) features in the deep neural network. Finally, we determine the best weight parameter between prediction results from DIAMOND and deep neural network. The experimental results show our proposed hybrid model outperforms traditional and state-of-the-art deep learning methods for protein function prediction.