Whole‐genome sequence annotation: ‘Going wrong with confidence’

Sir, The seventeenth complete genome, that of Chlamydia trachomatis, has been published recently, infusing the public database with another 894 protein sequences (Stephens et al., 1998, Science282: 754–759). The annotation of this genome is different from the previous 16 analyses, in that it introduces a number of spurious functional assignments based on questionable predictions (Table 1). First, new terms have been used that are semantically meaningless, for example ‘predicted’ or ‘possible’ functions (e.g. CT149, see Table 1 for identifiers), without explicit reference to the prediction methods used. Secondly, the presence of motifs and/or domains has been widely used as a substitute for function assignments (e.g. CT555). Thirdly, there are a number of overpredictions, i.e. overly specific assignments with no, or very remote, similarity to proteins of known function (e.g. CT775). All the above cases have the potential for resulting in a significant error propagation effect, especially when they refer to large hypothetical families from previous and ongoing genome projects. In other words, the Chlamydia genome is an actual snapshot of error propagation in public sequence databases (see also Karp, 1998, Bioinformatics14: 753–754). Thus, some caution should be exercised when assigning new functions based exclusively on the annotations of this genome. Overambitious annotation projects have appeared ever since the first genome sequences started covering unknown territory with novel protein families (Casari et al., 1995, Nature376: 647–648). For elusive cases, conflicting assignments are common and are inherent in the process of function prediction. Yet, there appears to exist a limit, above which predictions do more harm than good. A unique example of annotation abuse is the study of the Methanococcus jannaschii genome by Koonin and colleagues (http://www.ncbi.nlm.nih.gov/Complete_Genomes/MJan2/mjtable.html) (Koonin et al., 1997, Mol Microbiol25: 619–637). While three independent analyses have not surpassed a 50% level of function prediction (with 95% agreement; Andrade et al., 1997, Comput Appl Biosci13: 481–483), the study by Koonin et al. claimed that ‘about 70% of the archaeal proteins were predicted with varying precision’ (Koonin et al., 1997, Mol Microbiol25: 619–637). In this additional 20%, there are cases such as MJ0539, which was predicted to be a cysteinyl-tRNA synthetase, later found to be the lysyl-tRNA synthetase (Ibba et al., 1997, Science278: 1119–1122) (more examples in Table 1). Further to the above-mentioned three sources of error, there also exist cases apparently inconsistent with the clusters of orthologous genes (COGs), described previously by the same group (Tatusov et al., 1997, Science278: 631–637). Putative assignments based on ‘twilight-zone’ similarities and/or the mere presence of motifs should only serve as working hypotheses in search of a function. As such, they can reflect the current status and wealth of our collective knowledge and should therefore be continuously analysed within species-specific genome databases before their final admission to the public database. A concerted effort using continuous curation by experts and clear demonstration of evidence is therefore necessary to avoid this serious error propagation effect in computational genomics.