Document Type

Technical Report

Publication Date






Technical Report Number



Finding all gene structures is a crucial step in obtaining valuable information from genomic sequences. It is still a challenging problem, especially for vertebrate genomes, such as the human genome. Expressed Sequence Tags (ESTs) provide a tremendous resource for determining intron-exon structures. However, they are short and error prone, which prevents existing methods from exploiting EST information efficiently. This dissertation addresses three aspects of using ESTs for gene structure annotation. The first aspect is using ESTs to improve de novo gene prediction. Probability models are introduced for EST alignments to genomic sequence in exons, introns, interknit regions, splice sites and UTRs, representing the EST alignment patterns in these regions. New gene prediction systems were developed by combining the EST alignments with comparative genomics gene prediction systems, such as TWINSCAN and N-SCAN, so that they can predict gene structures more accurately where EST alignments exist without compromising their ability to predict gene structures where no EST exists. The accuracy of TWINSCAN_EST and NSCAN_EST is shown to be substantially better than any existing methods without using full-length cDNA or protein similarity information. The second aspect is using ESTs and de novo gene prediction to guide biology experiments, such as finding full ORF-containing-cDNA clones, which provide the most direct experimental evidence for gene structures. A probability model was introduced to guide experiments by summing over gene structure models consistent with EST alignments. The last aspect is a novel EST-to-genome alignment program called QPAIRAGON to improve the alignment accuracy by using EST sequencing quality values. Gene prediction accuracy can be improved by using this new EST-to-genome alignment program. It can also be used for many other bioinformatics applications, such as SNP finding and alternative splicing site prediction.


Permanent URL: