基于图聚类和序列信息的蛋白质远同源性探测

Detecting protein remote homologues based on graph clustering and sequence similarity

远同源蛋白间的序列相似性很低,处于随机涨落区域边缘,很难区分通过比对获得的序列特征是进化过程中功能约束还是随机突变导致的结果。为了从具有高度噪音的比对分数中提取关于同源性的微弱信号,本文将基于模块度优化的图聚类算法应用于蛋白质相似性网络来探测远同源性,并与已有的性能表现较好的方法进行比较,实验结果表明,该方法能较好地探测到蛋白质远同源性,输出的集团在很大程度上对应着蛋白质超家族,输出的集团数目接近数据集中超家族的个数,得到的结果明显优于其他方法。

Remote homologues are hidden in the twilight zone, where their sequence similarity is too low to distinguish them from pairs of sequences with equal or even higher sequence similarity due to chance. To extract weak signals about homology from sequence similarity with high noises, we apply modularity-based graph clustering algorithm to a protein similarity network to detect remote homology. Performance is compared with existing algorithms. Numerical experiments show that the modularity-based algorithm manages to detect protein remote homology, for the communities it outputs are largely corresponding to protein supeffamilies and the number of communities is close to the number of superfamilies in the given dataset. Also, the performance is superior other algorithms.