Protein functional annotation refinement based on graph regularized l^sub 1^-norm PCA

• Published in: Pattern recognition letters Vol. 87; p. 212
• Main Authors: Sun, Dengdi, Liang, Huadong, Ge, Meiling, Ding, Zhuanlian, Cai, Wanting, Luo, Bin
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Science Ltd 01.02.2017
• Subjects: Annotations; Convexity; Genomes; Graph theory; Iterative methods; Missing data; Optimization; Polymerase chain reaction; Protein interaction; Proteins; Sparsity; Yeast
• ISSN: 0167-8655; 1872-7344

As various high-throughput experimental techniques are developed to characterize biological systems at the genome scale, vast protein function data have been collected. However, these functional tags are often imprecise and/or incomplete, resulting in unsatisfactory performances in function-related experiments and applications. Therefore, the refinement of protein functional annotation has been one of the fundamental challenges in post-genomic era. In this paper, the refinement problem is formulated as a decomposition of the available functional tag matrix X with noise and missing data into latent protein features U and function representations V. Specifically, we aim to minimize the sparse error between original tag matrix X and the recovery matrix, and incorporate the protein-protein interactions and the function-function correlations to regularize two latent feature matrices U and V respectively. Finally, all these components are integrated into a convex optimization model, named graph regularized ℓ1-norm PCA, and an efficient convergence iterative procedure is proposed for the model based on augmented lagrangian multiplier method. We conduct extensive experiments on yeast database and the results indicate that our new refinement method outperforms consistently other related state-of-the-art approaches by the average precision and average F1-score.