A Large-scale Batch-learning Self-organizing Map for Function Prediction of Poorly-characterized Proteins Progressively Accumulating in Sequence Databases

Homology searches for nucleotide and amino-acid sequences have been used widely to predict functions of genes and proteins when genomes are decoded and thus become a basic bioinformatics tool. Whereas usefulness of the sequence homology search is apparent, it has become increasingly clear that homology search can predict the protein function of only 50% of genes, or fewer, when a novel genome is decoded. As a result of decoding of extensive genome sequences from a wide variety of phylotypes, a large number of proteins whose function cannot be predicted by the homology search of amino acid sequences is progressively accumulated and thus remains of no use in science and industry. A method to estimate the protein function that does not depend on the sequence homology search is in urgent need. We previously developed a Batch-Learning SOM (BL-SOM) for genome informatics, which does not depend on the order of data input. This report focuses on BL-SOM analyses on di to tri continuous amino acid frequencies. Concerning the diand tripeptide frequencies in the 110,000 proteins which have been classified into 2,853 function-known COGs (clusters of orthologous groups of proteins to represent individual functional categories), BL-SOMs that faithfully reproduced the COG classifications were obtained. This indicated that proteins, whose functions are presently unknown because of lack of significant homology with function-known proteins, can be related to function-known proteins with the BL-SOM.