Incorporating Homology Using Multi-Instance Kernel for Protein Subcelluar Localization

Kernel method has witnessed many successful applications in computational biology in recent years, and thus kernel design is a key step to define the similarity between two protein sequences. This paper aims at designing a kernel to derive more accurate similarity between two protein sequences by incorporating homology. Here a homologous sequence is viewed as one evolutionary instance of the target sequence and all homologous sequences constitute one homology bag. K-mer based spectrum kernel is used to define the similarity between any two instances and multi-instance kernel is as the sum of instance-wise spectrum kernels, called homology-based multi-instance kernel (HoMIKernel). By varying k-mer size and compressing 20 amino acids, we can derive several HoMIKernels, which are combined as HoMIKernel+ to capture more contextual information and cover size-varying motifs. We evaluate HoMIKernel+ on three unbalanced eukaryotic benchmark dataset. The experiments show that HoMIKernel+ achieves better predictive performance than the baseline models; and the incorporation of homologous sequences does increase the predictive accuracy.