GPS-PBS: A Deep Learning Framework to Predict Phosphorylation Sites that Specifically Interact with Phosphoprotein-Binding Domains

Abstract

Protein phosphorylation is essential for regulating cellular activities by modifying substrates at specific residues, which frequently interact with proteins containing phosphoprotein-binding domains (PPBDs) to propagate the phosphorylation signaling into downstream pathways. Although massive phosphorylation sites (p-sites) have been reported, most of their interacting PPBDs are unknown. Here, we collected 4458 known PPBD-specific binding p-sites (PBSs), considerably improved our previously developed group-based prediction system (GPS) algorithm, and implemented a deep learning plus transfer learning strategy for model training. Then, we developed a new online service named GPS-PBS, which can hierarchically predict PBSs of 122 single PPBD clusters belonging to two groups and 16 families. By comparison, GPS-PBS achieved a highly competitive accuracy against other existing tools. Using GPS-PBS, we predicted 371,018 mammalian p-sites that potentially interact with at least one PPBD, and revealed that various PPBD-containing proteins (PPCPs) and protein kinases (PKs) can simultaneously regulate the same p-sites to orchestrate important pathways, such as the PI3K-Akt signaling pathway. Taken together, we anticipate GPS-PBS can be a great help for further dissecting phosphorylation signaling networks.

Driver Attribute Filling for Genes in Interaction Network via Modularity Subspace-Based Concept Learning from Small Samples

“…In future work, there is still some room for improvement of our proposed model. One promising improvement is compensating more information from the data beyond interaction network and genomic mutation data, that is, integrated information from multiomics data such as transcriptome [39,40], epigenome [41], and proteome [42]. Another point is the consideration of effect from deleterious synonymous variants into the framework of our model, that is, regarding the mutation types with more…”

FSL-Kla: A few-shot learning-based multi-feature hybrid system for lactylation site prediction

“…Here, we defined a Kla site peptide KSP(m, n) as a lysine residue flanked by m residues upstream and n residues downstream. As previously described [17] , we adopted KSP(10, 10) for model training and parameter optimization in a rapid manner. For KSPs located at N- or C-terminals, we added one or multiple special characters “*” to complement the full KSP(10, 10) entries.…”

“…In the past study, these two groups of features were regarded as independent but highly complementary features. Advances of multiple feature encodings are obvious and facilitate the in silico PTM site prediction to a large extent [12] , [17] . Reproduceable and stable extraction is important for feature encodings [12] .…”

“…We comprehensively analyzed and assessed 11 types of feature encoding schemes, consisting of 8 types of sequence-derived features including amino acid composition (AAC), amino acid index (AAindex), composition of k -spaced amino acid pairs (CKSAAP), composition/transition (CTDC, CTDT), conjoint triad (CTriad), di-peptide composition (DPC) as well as position specific scoring matrix (PSSM), and the three additional types of structural features including accessible solvent accessibility (ASA), backbone torsion angles (BTA) and secondary structure (SS) ( Fig. 1 B) [15] , [16] , [17] , [18] , [19] , [20] . In this study, heterogeneous few-shot strategies were developed for balancing dataset and reducing overfitting ( Fig.…”

Biocomputing and Synthetic Biology in Cells: Cells Special Issue

“…Three articles focus on computational approaches and analyses for the identification of related binding proteins, such as phosphorylation sites in phosphoprotein-binding domains, ubiquinone-binding proteins and phage virion proteins. The article by Guo et al [ 9 ] develops an improved framework for identification of phosphorylation sites (p-sites) that specifically interact with phosphoprotein-binding domains (PPBDs), based on deep learning methods. A framework of seven-layer deep neural networks (DNNs) is implemented to train a general model to predict PPBD-specific binding p-sites (PBSs), containing an input layer, five fully connected layers (hidden layers) and an output layer.…”