GLMFD: An Attention-Based CNN-LSTM Model for Transmembrane Domains Localization

Abstract

The transmembrane domains (TMDs) are involved in many significant protein-protein interactions. Structural information of TMDs is necessary for increasing our understanding of such biological processes. However, experimental determination of TMDs position was laborious and inefficient for mass integral membrane proteins. In the past two decades, many statistical algorithms were proposed to predict TMDs and achieved excellent results. These algorithms were both limited in large amounts of detailed protein topology data. In this paper, we proposed an attention-based global-local model to locate TMDs called GLMFD. TMD as a functional domain has its particular hydrophobicity patterns that the right-sized local window can capture. Different from these traditional TMD prediction models, the position information of TMDs was an extra ‘byproduct’ of our localization model. Moreover, our model combined with proposed penalization terms successful located TMDs by learning the quantity distribution of TMDs. Experimental results show that our model and strategy perform well in TMD localization.