Search

We report on the development and application of an efficient unsupervised learning procedure for the classification of an unsegmented datastream, given a set of probabilistic binary similarity judgments between regions in the stream. The presence of noise in the similarity judgements and in the extent of similar regions is taken into account. Our method is effective on very large datastreams, and produces both a classified collection of segments and a set of frequency matrices that define patterns for the induced classes. We applied this method to the problem of finding the sequence-level building blocks of proteins. We first tested the clusterer on synthetic protein data with known evolutionary history. We then applied the method to a large protein sequence database (a datastream of more than elements) and found about 10,000 protein sequence classes which can be divided into definitions of protein families (classes of whole proteins) and definitions of protein building blocks (patterns that appear in otherwise unrelated proteins). These motifs are of significant biological interest

Publisher Summary There exists a sequence of 8 nucleotides, highly conserved at the boundary between an exon and an intron, referred to as the 5′ splice site (5′ ss). The boundary between an intron and an exon, the 3′ splice site (3′ ss), also exhibits a highly conserved sequence of 4 nucleotides, preceded by a pyrimidine-rich region. To interpret the deluge of sequence information generated by the human genome project, it is important to be able to identify genes in uncharacterized sequences. Without this capability, the sequence information is of little value. For each sequence in GenBank, the splice sites were located using the information in the annotation preceding the sequence, a FEATURES line being followed by multiple lines indicating different features present in the gene. The percentages of lariat windows with high-scoring branch point sequences in all GenBank categories, with the exception of viruses, are higher than those of random sequences. A notable observation is that even the window that contains the second highest frequency of branch points among the five windows analyzed always had a lower frequency of high scoring branch points than expected for a window with a random sequence.