Query-Adaptive Ranking with Support Vector Machines for Protein Homology Prediction

摘要

Protein homology prediction is a crucial step in template-based protein structure prediction. The functions that rank the proteins in a database according to their homologies to a query protein is the key to the success of protein structure prediction. In terms of information retrieval, such functions are called ranking functions, and are often constructed by machine learning approaches. Different from traditional machine learning problems, the feature vectors in the ranking-function learning problem are not identically and independently distributed, since they are calculated with regard to queries and may vary greatly in statistical characteristics from query to query. At present, few existing algorithms make use of the query-dependence to improve ranking performance. This paper proposes a query-adaptive ranking-function learning algorithm for protein homology prediction. Experiments with the support vector machine (SVM) used as the benchmark learner demonstrate that the proposed algorithm can significantly improve the ranking performance of SVMs in the protein homology prediction task.