Combining LSTM Network Model and Wavelet Transform for Predicting Self-interacting Proteins

Abstract

With the explosive growth of protein sequences generated by biological experiment in the post-genomic era, more and more researchers pay particular attention to the development of approaches for the prediction of protein interactions and functions from sequences. In addition, elucidation of the self-interacting proteins (SIPs) play significant roles in the understanding of cellular process and cell functions. This work explored the use of deep learning model, Long-Short Term Memory (LSTM), for the prediction of SIPs directly from their primary sequences. More specifically, the protein sequence is firstly converted to Position Specific Scoring Matrix (PSSM) by exploiting the Position Specific Iterated BLAST method, in which the evolutionary information is contained. Then, the wavelet transform algorithm is used on PSSM to extract discriminative feature. Finally, based on the knowledge of known self-interacting and non-interacting proteins, LSTM model is trained to recognize SIPs. The prediction performance of the proposed method is evaluated on yeast dataset, which achieved an accuracy rate of 92.21%. The experimental results show that the proposed method outperforms other six existing methods for SIPs prediction. Achieved results demonstrate that the proposed model is an effective architecture with SIPs detection, and would provide a useful supplement for the proteomics research.