|
Protein function prediction is one of the hottest research areas in biotech. According to different similarities characteristics of proteins, scientists construct more than one relationship between proteins, which can be called kernels. Different kernels have different domain-related information of proteins' relationship. Generally speaking, due to the complementarity of these information, multiple kernel learning (MKL) can overcome the heterogeneity between the kernels data to some extent and improve the protein function prediction precision. However, the existing domain-related kernels may contain some isolated proteins. Furthermore, the similarities between proteins may be considerably affected by noises. In this paper, we propose a fully connected kernel to enrich the available kernels. Based on the label propagation algorithm, the enriched kernels can not only propagate information to the isolated protein, but also can reduce the noises influences and improve the protein function prediction accuracy. We tested the method on the benchmark protein datasets, and the MKL algorithms with enriched kernel have a better performance. |
|
Keywords:Data Mining; Enriched Kernel; Multiple Kernels; Protein Function |
|