Empirical Comparison and Analysis of Artificial Intelligence-Based Methods for Identifying Phosphorylation Sites of SARS-CoV-2 Infection

Abstract

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is a member of the large coronavirus family with high infectivity and pathogenicity and is the primary pathogen causing the global pandemic of coronavirus disease 2019 (COVID-19). Phosphorylation is a major type of protein post-translational modification that plays an essential role in the process of SARS-CoV-2-host interactions. The precise identification of phosphorylation sites in host cells infected with SARS-CoV-2 will be of great importance to investigate potential antiviral responses and mechanisms and exploit novel targets for therapeutic development. Numerous computational tools have been developed on the basis of phosphoproteomic data generated by mass spectrometry-based experimental techniques, with which phosphorylation sites can be accurately ascertained across the whole SARS-CoV-2-infected proteomes. In this work, we have comprehensively reviewed several major aspects of the construction strategies and availability of these predictors, including benchmark dataset preparation, feature extraction and refinement methods, machine learning algorithms and deep learning architectures, model evaluation approaches and metrics, and publicly available web servers and packages. We have highlighted and compared the prediction performance of each tool on the independent serine/threonine (S/T) and tyrosine (Y) phosphorylation datasets and discussed the overall limitations of current existing predictors. In summary, this review would provide pertinent insights into the exploitation of new powerful phosphorylation site identification tools, facilitate the localization of more suitable target molecules for experimental verification, and contribute to the development of antiviral therapies.

Keywords: SARS-CoV-2; computation tool; deep learning; machine learning; phosphorylation site.

Figure 1

The workflow of existing computational tools for predicting phosphorylation sites in host cells infected with SARS-CoV-2. These tools are developed based on conventional machine learning models and end-to-end deep learning networks, mainly through the following steps: benchmark sequence data preparation, feature encoding and selection, classification model design, and prediction assessment.

Figure 2

The well-studied benchmark datasets of phosphorylation sites in host cells infected with SARS-CoV-2. (A) The originations, as well as the common collection and preprocessing procedures of benchmark datasets. (B) The detailed information of the three benchmark datasets, named A549 (human), Vero E6 (African green monkey), and combined.