Discovery of Diagnostic Patterns from Protein Sequence Databases

We show how prior domain knowledge can be used in a system for mining databases of biological data. Our system performs automated discovery of diagnostic patterns from a database of protein sequences. Such patterns are used for classiication of new sequences, and identiication of biologically interesting positions in the proteins. The patterns have a simple syntax and can be translated into regular expressions , which can be used for rapid scanning of databases. Current pattern libraries are built semi-manually, since the correctness of the pattern depends on the incorporation of domain knowledge. Due to the dramatic growth of the databases it is desirable to automate this process. Our results show that the patterns derived by our fully automated system compete well with the semi-manually constructed patterns.