De novo Prediction of Moonlighting Proteins Using Multimodal Deep Ensemble Learning

Ying Li; Jianing Zhao; Zhaoqian Liu; Cankun Wang; Lizheng Wei; Siyu Han; Wei Du

doi:10.3389/fgene.2021.630379

De novo Prediction of Moonlighting Proteins Using Multimodal Deep Ensemble Learning

Front Genet. 2021 Mar 22:12:630379. doi: 10.3389/fgene.2021.630379. eCollection 2021.

Authors

Ying Li¹, Jianing Zhao¹, Zhaoqian Liu², Cankun Wang², Lizheng Wei², Siyu Han³, Wei Du¹

Affiliations

¹ Key Laboratory of Symbol Computation and Knowledge Engineering of Ministry of Education, College of Computer Science and Technology, Jilin University, Changchun, China.
² Department of Biomedical Informatics, College of Medicine, Ohiostate University, Columbus, OH, United States.
³ Department of Computer Science, Faculty of Engineering University of Bristol, Bristol, United Kingdom.

Abstract

Moonlighting proteins (MPs) are a special type of protein with multiple independent functions. MPs play vital roles in cellular regulation, diseases, and biological pathways. At present, very few MPs have been discovered by biological experiments. Due to the lack of data sample, computation-based methods to identify MPs are limited. Currently, there is no de-novo prediction method for MPs. Therefore, systematic research and identification of MPs are urgently required. In this paper, we propose a multimodal deep ensemble learning architecture, named MEL-MP, which is the first de novo computation model for predicting MPs. First, we extract four sequence-based features: primary protein sequence information, evolutionary information, physical and chemical properties, and secondary protein structure information. Second, we select specific classifiers for each kind of feature. Finally, we apply the stacked ensemble to integrate the output of each classifier. Through comprehensive model selection and cross-validation experiments, it is shown that specific classifiers for specific feature types can achieve superior performance. For validating the effectiveness of the fusion-based stacked ensemble, different feature fusion strategies including direct combination and a multimodal deep auto-encoder are used for comparative purposes. MEL-MP is shown to exhibit superior prediction performance (F-score = 0.891), surpassing the existing machine learning model, MPFit (F-score = 0.784). In addition, MEL-MP is leveraged to predict the potential MPs among all human proteins. Furthermore, the distribution of predicted MPs on different chromosomes, the evolution of MPs, the association of MPs with diseases, and the functional enrichment of MPs are also explored. Finally, for maximum convenience, a user-friendly web server is available at: http://ml.csbg-jlu.site/mel-mp/.

Keywords: deep learning; ensemble learning; multimodal; prediction model; protein moonlighting.