Based on the collected phosphosites with functional annotations, the ranked first three molecular functions include molecular association, activity regulation and intracellular localization. As in the Venn diagram, many phosphosites possess the two and even three functions, indicating the function biased predictive model is not easily accessible. Herein, a novel integrated deep neural network model named FuncPhos-SEQ is proposed for functional assignment, including general function, activity and molecular association, of human proteome-level phosphosites. The FuncPhos-SEQ incorporates SeqNet module to extract phosphosite motif information from protein sequence using multiple convolutional neural networks (CNNs) channels, and SPNet module to extract network features from protein-protein interactions (PPIs) using network embedding and deep neural network (DNN) channels. These concatenated features were jointly fed into a heterogeneous feature CoNet module to prioritize functional phosphosites.

FuncPhos-SEQ requires only the raw protein sequence and concerned phosphosites as input, avoiding the problem of losing information for a large amount of phosphosites due to the limited number of protein crystal structures. It is computationally efficient and has a broader scope of application for biologists.