The completion of a first draft for the human genome sequence represents an outstanding scientific achievement of our times. Despite the unprecedented hype in the popular media, it is evident that the completion of this phase is just the beginning of a long march towards an understanding of the structure and function of our genetic blueprint [1]. What is to be expected from the analysis of the human genome five years from now? One guide in answering this question is a thorough assessment of our capability to analyze the genomes that have already been sequenced in the past five years, starting in 1995 with the genome of Haemophilus influenzae [2,3]. Since then, 31 entire genome sequences have been made available to the public domain (as of September 2000) [4]. To obtain a snapshot of the maximum possible set of database-driven sequence annotations across species, we have used GeneQuiz, a system for automated large-scale sequence analysis [3,5].

Automated genome sequence analysis and annotation has the advantage that the analysis strategy is uniformly applied to all genome sequences against the same database, rendering results comparable [5]. The updated annotations also form a resource for the scientific community for further analysis and assessment. The set of annotations we obtained contains an additional 5,534 novel protein function assignments, a 15% increase over the original genome sequence publications (Figure 1a). Evaluation of the results has shown that the agreement of automated function prediction reaches 95%, when compared to expert manual analysis [5,6]. Through this procedure, patterns that are not discernible from individual genomes become apparent, with interesting implications for the future.

First, the detection of proteins of known structure and/or function has increased over time (Figure 1b). This is especially true for homologs of known structure, a trend that should be further enhanced during the next five years as a result of the recent initiatives in structural genomics [7,8]. The total number of homologs of known function has also increased, but at a slower pace (Figure 1b). This may be attributable to two factors: either high-throughput experimental characterization has not yet provided genome-wide functional information or that this type of functional annotation has not yet been adequately transferred into the public sequence databases (Figure 1b) [9]. It is imperative that provisions be made to start recording this type of functional information for the human genome [10].

Second, there is some degree of variability for the 31 genomes that we have analyzed. On average, function is known or can be predicted for 62% of the proteins encoded in each genome (Figure 1a). Of these, a significant proportion has homologs in the structure database (19% on average) and the remaining majority (43% on average) have homologs of known function. The total fraction of proteins of predicted structure/function varies significantly across species, ranging from 31% for Aeropyrum pernix (9% structure and 22% function) to 80% for Escherichia coli (21% and 59%, respectively; see Figure 1a). It remains to be seen how the human genome ranks within this range of functional assignments.

Third, there is a substantial number of uncharacterized proteins, including hypothetical proteins (with homologs of unknown function; 26% on average) or 'unique' proteins (without homologs; 12% on average). Hypothetical proteins range from 13% for E. coli and 14% for Rickettsia prowazekii to 40% for Pyrococcus abyssii and 47% for Plasmodium falciparum (chromosome 2). These protein families represent excellent candidate targets for functional genomics. Species with high percentages of hypothetical proteins demarcate taxa whose properties generally remain unknown. This group includes Archaea (34% on average), Mycobacterium tuberculosis (33%), Helicobacter pylori and Chlamydia pneumoniae strains (all at 30%) and yeast (24%). At the other end of the spectrum, certain species are sufficiently covered by their relatives (e.g. H. influenzae at 14%).

Finally, the percentage of 'unique' proteins delineates the number of uncharacterized protein families. This number varies widely, ranging from 0% for Mycoplasma genitalium and 4% for two Chlamydia trachomatis strains to 39% for A. pernix. A few of these genes may encode taxon-specific proteins or may represent false gene predictions. It is interesting that species with few unique proteins are bacterial (Figure 1a), an indication that many protein families in their phylogenetic neighbourhood have been detected. At the other extreme, species with many unique sequences also include Treponema pallidum, the two Neisseria meningitidis strains (all at 22%) and Deinococcus radiodurans (24%). It is expected that the human genome will also contain a high number of unique proteins, because it currently represents the only vertebrate genome that has been fully sequenced.

The annotation level for each species crucially depends on the existence of homologs of the proteins that are potentially encoded in the genome sequence. Automatic function assignment also depends on the quality of the underlying databases and the availability of proteins of known structure or function. Another factor that significantly influences annotation quality is the definition of gene structure, an exceedingly difficult task for eukaryotic genomes [11]. The difficulties of eukaryotic genome annotation have been recently reviewed [12], with specific details on the Ensembl and GadFly projects for the human and the fruitfly genomes, respectively.

The present study suggests that a great deal should be expected from the analysis of the human genome. On the one hand, using up-to-date databases and automatic approaches, a considerable proportion of the human genome will be annotated, to guide further experimentation and discovery. On the other hand, no matter how much the database increases over time, there will still be a great need for functional experiments, to detect the cellular roles of novel proteins. In spite of progress in the field of bioinformatics [13], filling this information gap is going to be the next major challenge for genomics research, where large-scale experimentation will lead the way [14].