Despite considerable efforts in the Bioinformatics community, the performance of existing gene prediction tools is still not satisfactory. In current genome projects, a common approach for gene finding is the following. Several sets of gene predictions are compiled, usually from different gene finders trained specifically for the species at hand. Further, alignments from ESTs and proteins to the genome are constructed. Finally, the predictions and the alignments are combined to find plausible gene structures, either manually or by using meta tools that combine several predictions and alignments [e.g. (1)].

AUGUSTUS is a gene finder based on a Generalized Hidden Markov Model (GHMM) (2,3). The original version of the program was a purely ab initio method, i.e. its prediction was based on information contained in the genomic sequence to be analyzed. An extended version of the program is able to use additional extrinsic information, for example matches to protein databases or alignments of genomic sequences, to improve the prediction accuracy (4). At the recent EGASP workshop in Cambridge, UK, a systematic evaluation of existing gene finders for the human genome has been carried out based on a large set of well-annotated parts of the human genome (5). At this workshop, AUGUSTUS turned out to be the best program in the category of ab initio gene prediction. Its performance could be further improved by using BLAST (6) hits to EST or protein sequences and alignments of syntenic genomic sequences using DIALIGN (7,8); in this category, however, the program was outperformed by N-Scan (9), a new program based on multiple alignments of genomic sequences. Compared to more traditional approaches, gene-finding methods based on genomic sequence alignments have a considerable advantage since they do not depend on EST or protein sequences or statistical models of gene structures (10–13). On the other hand, alignment-based methods work only if genome sequences at an appropriate evolutionary distance are available. Although the performance of ab initio gene-prediction methods is usually improved if information from comparative sequence analysis is added, ab initio gene prediction remains highly important since for many newly sequenced genomes, few EST or related genomic sequences sequences are available and comparison to protein sequences can find only those genes that have close relatives in existing databases.

To make AUGUSTUS available to the research community, we set up a WWW server at Göttingen Bioinformatics Compute Server (GOBICS) (14,15). Like most gene-prediction methods that are currently available, earlier versions of AUGUSTUS predicted exactly one transcript per gene and ignored the fact that one gene often yields more than one distinct mRNA product. It has been estimated that 40–60% of all human genes have alternative splice forms. Of those genes 70–88% of alternative splices change the protein product; the remaining splice variants differ in the untranslated regions only (16). Thus, it is important to have gene-finding tools that are able to deal with this phenomenon. The program SLAM (17), for example, predicts alternative splice variants. This program, however, is based on alignments of genomic sequences, and it requires two syntenic genomic sequences as input data. We recently installed a new version of AUGUSTUS at our server that can predict multiple transcripts for predicted genes. To our knowledge, this is the first ab initio gene finder that can predict multiple transcripts, and our web server is the only gene prediction web server with this option.

With our new alternative-transcripts option, the user can control the number of predicted splice variants per gene. This way, it is possible to influence sensitivity and specificity of the program output. If predicted genes or transcripts are automatically evaluated and post-processed, high prediction sensitivity may be desirable to increase the number of candidate genes that are to be analyzed, even if this increases the number of false-positive predictions. In contrast, if expensive experiments are carried out based on computationally predicted genes, it is preferable to have highly specific tools that minimize the risk of false-positive predictions. Thus, a good gene-finding method should allow the user to chose between high sensitivity and high specificity. At our server, this can be done by specifying the maximum number of predicted splice variants. In addition, we implemented a motif-searching option at our server where predicted genes can be searched for user-specified regular expressions, e.g. PROSITE patterns.

To analyze putative protein products of predicted genes, our server offers a pattern-searching option. Here, the user can enter arbitrary patterns as regular expressions; these patterns are then searched against all predicted splice variants from all genes predicted in the input sequence. This can be helpful if the user is looking for members of a particular protein family with conserved positions that are already described by a pattern, for example as an entry in the PROSITE database. In particular, in connection with the above described option for alternative splice forms, our pattern-searching option can be applied to filter out those splice variants that are consistent with the described pattern.

Following the PROSITE syntax, the conserved positions of a protein family may be specified as in the following example describing the hydrophobin family: <{C}(17,85)-C-{C}(5,10)-C-C-{C}(11,44)-C-{C}(8,23)-C-{C}(5,9)-C-C-{C}(6,18)-C-{C}(2,13)>. This pattern consists of eight conserved cystein residues with certain restrictions on the space between two subsequent cysteins. If a genome is to be searched for members of a protein family characterized by a regular expression, it is advisable to let AUGUSTUS predict a large number of splice variants to increase the prediction sensitivity, possibly at the expense of its specificity. With this strategy, the number of false positive splice variants may be increased, but a large number of predicted transcripts also increases the probability of finding all instances of the specified motif in the genome sequence under study.

We tested our method on a large set of test data from the recent EGASP workshop that was organized in the context of the ENCODE project (5). This dataset comprises a total of 296 genes with an average of 2.2 transcripts per gene. Table 1 shows the relation between the posterior probability of exons as defined above and the specificity at the exon level. As one may expect, the posterior probability is a good indicator of exon specificity. Predicted exons with low posterior probability are much less likely to be correct than predicted exons with high posterior probability. Therefore, the posterior probability gives a good criterion for prioritising putative exons, e.g. for experimental verification.

Table 2 summarizes the prediction accuracy of AUGUSTUS on the EGASP test data. We used AUGUSTUS with the original single-transcript option and with the new options for ‘few transcripts’, ‘medium number of transcripts’ and ‘many transcripts’. Sensitivity and specificity of these program versions are given at four different levels, namely at the base, exon, transcript and gene level. At the gene level, a gene is considered a true positive if one predicted transcript coincides with an existing transcript of this gene. As expected, the program sensitivity increases at all four levels if the number of transcripts per gene is increased. Correspondingly, the specificity at the base, exon and transcript level decreases if the number of predicted transcript increases. In contrast, at the gene level the specificity of AUGUSTUS increases if more transcripts per gene are predicted. The reason for this seemingly paradoxical result is as follows: with an increased number of predicted transcripts, it is more likely that one of the real transcripts in a gene is matched by a predicted transcript. The number of predicted genes, on the other hand, does not increase significantly if more transcripts are predicted because most of the additionally predicted transcripts belong to one of the previously predicted genes. Thus, the ratio of true positive genes to predicted genes increases if the number of predicted transcripts is increased.

Letting AUGUSTUS predict many transcripts is most useful when the focus is on finding at least one correct splice form for a gene: as shown in Table 2, the gene-level sensitivity increases from 23.3 to 34.5% if the option ‘many transcripts’ is used (on average, 5.1 transcripts per gene are predicted with this option). This is substantially more than the gene-level sensitivity of standard gene finders; e.g. for the commonly used programs GENSCAN (21), GENEID (22) and GENEZILLA (23), the gene-level sensitivity values on the EGASP dataset are 15.5, 10.5 and 19.6%, respectively.