Structure-Based Feature Extraction from Protein Databases

In this paper, we study the performance of the feature extraction method we developed for complex object databases, called Complex Object Feature Extraction (Cofe), with respect to protein datasets, using distance measures based on structural similarity between proteins. We rst perform an assessment of the accuracy of six automatic protein comparison methods against the manually constructed classiication of proteins, called SCOP. We then compare the quality of the feature spaces resulting from applying our developed feature extraction method against those obtained when a previously proposed method is used. The results on the considered dataset for ve diierent structure-based distance spaces show that Cofe provides signiicantly higher quality embeddings for four of them. We conclude that Cofe proves to be a practical method for extracting high quality features from protein databases.