Framework annotation of genomic DNA sequence

As increasing amounts of genomic sequence from many organisms become available, and as DNA sequences become a primary reagent in biologic investigations, the role of annotation as a prospective guide for laboratory experiments will expand rapidly. We describe here a process of high-throughput, reliable, automated sequence annotation called framework annotation, which is designed to provide a foundation for initial biologic characterization of previously unexamined sequence. This specification includes structured representations of annotation designed to facilitate consistency and queriability. Information describing the process by which results were obtained is also tracked, in order to assess both methods and results. A series of annotation policy statements describes the methods by which automated annotation is generated. We have constructed the GAIA (Genome Annotation and Information Analysis) prototype software architecture as a proof of concept for this process, and used it to analyze several methods for generating potential annotation. Using test sets of human genomic sequences, we find nucleotide similarity search against large EST datasets to be a generally sensitive, and, at high stringency, reliable probe for the existence of genes in a query sequence. These results may also provide initial information about alternative mRNA structures and expression profiles. GRAIL was somewhat more sensitive overall, and superior to EST-driven search in detecting exons away from the 3$\sp\prime$ end of a gene, but was much less specific than similarity search. Similarity search against amino acid sequence datasets was also apparently sensitive, and reliable at high stringency. Interestingly, little advantage was gained by considering conservative substitutions in addition to identities between sequences. We propose that an annotation process which grounds gene models in high-stringency similarity results, extends them with lower-stringency and pattern-based data, and uses information from amino acid alignments for initial functional analysis may serve as an effective means of first-pass gene-related annotation. The refinements in data definition and visualization presented here may also be extended to management of genome-scale information beyond framework annotation