Research Article

Protein-protein interaction network construction for cancer using a new L1/2-penalized Net-SVM model

Published: July 25, 2016
Genet. Mol. Res. 15(3): gmr8794 DOI: https://doi.org/10.4238/gmr.15038794
Cite this Article:
H. Chai, H.H. Huang, H.K. Jiang, Y. Liang, L.Y. Xia, H. Chai, H.H. Huang, H.K. Jiang, Y. Liang, L.Y. Xia (2016). Protein-protein interaction network construction for cancer using a new L1/2-penalized Net-SVM model. Genet. Mol. Res. 15(3): gmr8794. https://doi.org/10.4238/gmr.15038794
2,014 views

Abstract

Identifying biomarker genes and characterizing interaction pathways with high-dimensional and low-sample size microarray data is a major challenge in computational biology. In this field, the construction of protein-protein interaction (PPI) networks using disease-related selected genes has garnered much attention. Support vector machines (SVMs) are commonly used to classify patients, and a number of useful tools such as lasso, elastic net, SCAD, or other regularization methods can be combined with SVM models to select genes that are related to a disease. In the current study, we propose a new Net-SVM model that is different from other SVM models as it is combined with L1/2-norm regularization, which has good performance with high-dimensional and low-sample size microarray data for cancer classification, gene selection, and PPI network construction. Both simulation studies and real data experiments demonstrated that our proposed method outperformed other regularization methods such as lasso, SCAD, and elastic net. In conclusion, our model may help to select fewer but more relevant genes, and can be used to construct simple and informative PPI networks that are highly relevant to cancer.

Identifying biomarker genes and characterizing interaction pathways with high-dimensional and low-sample size microarray data is a major challenge in computational biology. In this field, the construction of protein-protein interaction (PPI) networks using disease-related selected genes has garnered much attention. Support vector machines (SVMs) are commonly used to classify patients, and a number of useful tools such as lasso, elastic net, SCAD, or other regularization methods can be combined with SVM models to select genes that are related to a disease. In the current study, we propose a new Net-SVM model that is different from other SVM models as it is combined with L1/2-norm regularization, which has good performance with high-dimensional and low-sample size microarray data for cancer classification, gene selection, and PPI network construction. Both simulation studies and real data experiments demonstrated that our proposed method outperformed other regularization methods such as lasso, SCAD, and elastic net. In conclusion, our model may help to select fewer but more relevant genes, and can be used to construct simple and informative PPI networks that are highly relevant to cancer.