Abstract
Motivation Rapid and accurate identification of transmembrane (TM) topology is well suited for the annotation of the entire membrane proteome. It is the initial step of predicting the structure and function of membrane proteins. However, existing methods that utilize only amino acid sequence information suffer from low prediction accuracy, whereas methods that exploit sequence profile or consensus need too much computational time.
Method Here we propose a deep learning framework DeepCNF that predicts TM topology from amino acid sequence only. Compared to previous sequence-based approaches that use hidden Markov models or dynamic Bayesian networks, DeepCNF is able to incorporate much more contextual information by a hierarchical deep neural network, while simultaneously modeling the interdependency between adjacent topology labels.
Result Experimental results show that PureseqTM not only outperforms existing sequence-based methods, but also reaches or even surpasses the profile/consensus methods. On the 39 newly released membrane proteins, our approach successfully identifies the correct TM segments and boundaries for at least 3 cases while all existing methods fail to do so. When applied to the entire human proteome, our method can identify the incorrect annotations of TM regions by UniProt and discover the membrane-related proteins that are not manually curated as membrane proteins.
Availability http://pureseqtm.predmp.com/
Footnotes
Add Figure 10: The runtime analysis of PureseqTM, Phobius and Philius with different protein lengths.