Spectral Clustering on Neighborhood Kernels with Modified Symmetry for Remote Homology Detection

Remote homology detection among proteins in an unsupervised approach from sequences is an important problem in computational biology. The existing neighborhood cluster kernel methods and Markov clustering algorithms are most efficient for homolog detection. Yet they deviate from random walks with inflation or similarity depending on hard thresholds. Our spectral clustering approach with new combined local alignment kernels more effectively exploits state-of-the-art neighborhood vectors globally. This approach combined with Markov clustering similarity after modified symmetry based corrections outperforms other six cluster kernels for unsupervised remote homolog detection even in multi-domain and promiscuous proteins from Genolevures database with better biological relevance. Source code available upon request.