First Pass Annotation of Promoters on Human Chromosome 22

RESULTS

Experimentally Verified Promoters on Human Chromosome 22

To identify experimentally verified promoters on chromosome 22, we performed extensive searches in MEDLINE and GenBank. Because promoters are often referred to by a variety of expressions, we carried out both sequence-based BLAST searches (NCBI) (Altschul et al. 1990) and keyword-oriented text searches. We used genomic fragments containing 2 kb upstream and 500 bp downstream of the annotated gene starts as query sequences for BLAST. GenBank annotation as well as ENTREZ and MEDLINE were searched for entries containing the gene names and/or chromosome 22 annotation in order to include as many promoters as possible. Lastly, we mapped all entries of the Eukaryotic Promoter Database (EPD) (Perier et al. 2000) to the sequence of chromosome 22.

This approach yielded only 20 experimentally verified promoters of known genes on chromosome 22 (Table 1). We compared the location of the 20 promoters with the gene starts annotated by Dunham et al. (1999). In 18 cases, the experimentally verified promoters agreed very well with gene starts. However, two promoters (PLA2G6 and GGT1) were found to be located at a significant distance upstream of the annotated gene starts (12 kb, PLA2G6 and 20 kb, GGT1). We were able to map the 93 bp noncoding first exon of the PLA2G6 mRNA to the genomic sequence of chromosome 22 (ExonMapper, GEMS Launcher package). This exon was not included in the original annotation by Dunham et al. (1999), and the real gene start is located 12 kb upstream of the annotated gene start due to a large first intron. In the case of the GGT1 promoter, no continuous mRNA was available but the promoter sequence matches the chromosome 22 sequence over a stretch of more than 450 bp with just a single mismatch, showing no gaps at all. In summary, experimentally verified data are available for only a very few of the promoters on chromosome 22, even considering that we might have missed a few promoters due to unusual annotation. Therefore, large-scale promoter annotation requires appropriate in silico methods.

Sequence Analysis and Promoter Region Predictions

PromoterInspector is an in silico method which is trained to predict the genomic context of polymerase II promoters. Details of the algorithm are as described earlier (Scherf et al. 2000). The application of PromoterInspector to chromosome 22 yielded 465 regions (minimum length 192 bp, maximum 2432 bp, average 555 bp, Table 2).

PromoterInspector predicts CpG island- as well as nonCpG island-associated promoter regions. Because 60% of human genes have distinctive CpG islands at their 5′ ends (Cross and Bird 1995) and the chromosome 22 sequence was found to be G + C rich (Dunham et al. 1999), we examined CpG island-associated predictions in more detail.Dunham et al. (1999) reported 553 CpG islands of which 543 were documented on the Web server of the Sanger Centre (http://www.sanger.ac.uk/cgi-bin/cwa/22cwa.pl). The minimum CpG island length is 400 bp, the maximum length is 10,000 bp and the average length is 1074 bp (Table 2).

Comparison of Promoter Region Predictions with Existing Annotation

PromoterInspector, as well as CpG islands predictions, yielded reasonable numbers of matches on chromosome 22 (PromoterInspector: 465 matches, CpG islands: 543 matches). The next step was to compare the results with existing gene annotations in order to determine whether the predictions were reliable.

The quality of the predicted regions was assessed on basis of the 5′ ends of the genes annotated by Dunham et al. (1999). We carried out a correlation analysis of all predicted promoter regions with annotated gene starts with the program packageGenomeInspector (Quandt et al. 1996). The correlation analysis was done with respect to the different quality of gene annotation. We considered three groups of genes which were introduced by Dunham et al. (1999): (1) known genes (genes which are identical to human genes or protein sequences), (2) related genes (genes homologous, or containing a region of similarity, to gene or protein sequences from human or other species), and (3) predicted genes (sequences homologous to ESTs).

Promoter regions were correlated with “known genes” and “related genes” within a region of 2 kb upstream and 0.5 kb downstream of the annotated gene starts (Fig. 1). In the case of the “predicted genes,” the correlation peak was extended up to 6 kb upstream (Fig. 1B). The correlated promoter regions were considered “annotation-supported” promoter regions.

Table 3 summarizes the results ofPromoterInspector predictions and CpG islands. The portion of regions correlated with gene is approximately the same forPromoterInspector predictions (38.7%) and CpG islands (39.4%). The numbers of the annotation-supported promoter regions and CpG islands might still be on the cautious side, especially in the case of EST-based gene annotation, where missing 5′ sequences can easily exceed 10 kb. In order to calculate the percentage of annotation-supported predictions, we set the total number of predictions obtained with each method to 100%. Dunham et al. (1999)identified a group of 134 “pseudo genes;” that is, sequences homologous to a known gene or protein sequence but with a disrupted open reading frame. Given a threshold of 2 kb, only six promoter regions predicted by PromoterInspector were correlated with a gene start in this group.

Because PromoterInspector predicts CpG island- as well as nonCpG island-associated promoters, we tried to improve the CpG island predictions by a filter approach: We considered only those CpG islands which overlapped with a PromoterInspector prediction. The filter approach resulted in 358 CpG islands, and 47.5% of them are correlated with an annotated gene start, as summarized in Table4. In light of these results, the question arose as to whether this improvement could also be reached by filtering CpG islands with other in silico promoter prediction methods. We applied Promoter 2.0 (Knudsen 1999) and NNPP 2.1 (M. Reese, in prep.) and considered only those CpG islands where a promoter was predicted. Table 5summarizes the results. NNPP 2.1 reduced the number of CpG islands without an improvement of predictions. Promoter 2.0 predicted a promoter in only 164 CpG islands, of which 52.2% are correlated with a gene start.

Finally, we considered PromoterInspector predictions which are not correlated with CpG islands. As can be seen in Table 1, 20% of the PromoterInspector regions which are correlated with an experimentally verified promoter are nonCpG island predictions. The correlation of nonCpG island PromoterInspector predictions with gene annotations is summarized in Table6.

Gene Prediction Combined with Promoter Prediction

Dunham et al. (1999) applied GenScan (Burge and Karlin 1997), a program for identification of exon/intron structures, to predict genes ab initio. A total of 817 GenScanpredictions were obtained. Although 94% of the annotated genes were at least partially detected by GenScan, all exons were predicted correctly for only 20% of annotated genes. Because of these results, Dunham et al. (1999) stated that “… ab initio gene prediction cannot be used directly to annotate genes in human sequences.”

We examined whether a combination of GenScan andPromoterInspector might improve the ab initio gene prediction. As a first step, we determined the set of composite predictions (i.e., GenScan gene predictions with a 5′ end within or at most 100 bp downstream from aPromoterInspector promoter region). A total of 92GenScan/PromoterInspector predictions fulfilled this requirement. Again we used the gene annotations ofDunham et al. (1999) to estimate the reliability of these predictions.

We found 11 composite predictions where the GenScanprediction did not overlap with an annotated gene, nor was the respective promoter region annotation-supported. Of the remaining 81 composite predictions, 49 (60.4%) had an annotation-supported promoter region and the respective GenScan predicted gene overlapped with the annotated gene. In 32 cases, the promoter regions were not annotation-supported but the respective GenScanprediction overlapped partially with an annotated gene.

From these results we concluded that composite predictions have a high chance (>50%) to correlate with true promoters. In addition, promoter regions appear to be useful markers for delineating the 5′ boundary of subsequences to be analyzed by GenScan. We could verify this for an example, the SLCRA1 gene (Heisterkamp et al. 1995), which was originally not correctly predicted byGenScan. Using the annotation-supported promoter region predicted by PromoterInspector as a 5′ boundary,GenScan correctly recognized all exons of theSLCRA1 gene.

In summary, our results suggest that the 11 additional composite predictions are more likely candidates for real genes than are isolatedGenScan predictions, because GenScan andPromoterInspector independently identify different sequence features. Composite gene predictions might thus be useful as an in silico extension of the chromosome 22 annotation.

DISCUSSION

Promoters contain vital information about gene expression and regulatory networks, including gene targets of individual cascades/signaling pathways. To date, <5% of the promoters in chromosome 22 are known from experimental analysis.

We have shown that in silico promoter annotation of large-scale chromosomal sequences is feasible with a quality that is suitable for experimental design. Every second to third prediction ofPromoterInspector and every second prediction of thePromoterInspector-filtered CpG island predictions can be shown to be correct. Although the annotation derived by our predictions is not complete (about every third annotated gene was correlated with a predicted promoter region), it is, to our knowledge, the first successful large-scale prediction of promoter regions.

As can be seen in Table 3, PromoterInspector and CpG islands led to comparable numbers in gene start correlations. However, a significant difference between these two approaches is that the length of the PromoterInspector regions is, on average, one-half the size of the minimum length of CpG islands. Therefore,PromoterInspector predictions pinpoint gene starts with much more precision. In addition, the results in Tables 1, 4, and 6show that PromoterInspector predicts CpG island- as well as nonCpG island-associated promoters. Our results show a bias ofPromoterInspector predictions towards CpG islands. One reason for this is that ∼60% of the promoter sequences which were used to train the PromoterInspector contained CpG islands. Since the training procedure focuses on the most common patterns in the training set (Scherf et al. 2000), it is clear that the prediction is biased towards C + G-rich patterns. However, our results show that the PromoterInspector approach is (to our knowledge) the only one able to predict promoter regions on the genome level with such a small sequence coverage (i.e., precision). To underline this statement, we compared PromoterInspector with the promoter prediction tools NNPP 2.1 (M. Reese, in prep.),TSSG (Solovyev and Salamov 1997), TSSW(Solovyev and Salamov 1997) and Promoter 2.0 (Knudsen 1999). These approaches focus on the detection of promoter elements like TATA and CAAT boxes rather than promoter regions. Since it is not possible to analyze whole chromosomes with these tools, we randomly extracted and analyzed 10 nonoverlapping sequences with a length of 50,000 bp from chromosome 22. From the obtained results we would expect 11,890 (TSSW), 14,963 (TSSW), 50,233 (Promoter 2.0) and 87,641 (NNPP 2.1) promoter predictions on the chromosome 22 sequence. Assuming that all promoters of the 545 annotated genes of chromosome 22 are included in these predictions, then only every 20th to 140th prediction is expected to be correct. This is certainly not very useful for subsequent experimental design.

Annotated gene starts are not always useful for the identification of promoters because gene annotation might be 5′ incomplete, as suggested by Dunham et al. (1999). We found two examples in which experimentally mapped promoters were located more than 10 kb upstream of annotated gene starts, demonstrating that our methods were able to identify promoter-containing regions with high reliability. The approach of predicting promoter regions independent of gene annotations also provides a new way toward mapping of short first exons that are most frequently missed by both cDNA mapping and gene prediction (Dunham et al. 1999). A prediction upstream of the known gene sequence is very likely to represent the correct promoter and should also be a useful addition for annotated genes.

Reliable promoter prediction can also be used in a more general way to provide a biologically meaningful “chromosomal scaffold” for a variety of further analyses. For example, gene prediction tools such asGenScan perform much better when they are used on segments containing only one gene or at least starting with a gene. Therefore, the combination of promoter region prediction with gene prediction tools like GenScan is a promising way to enhance the specificity of de novo gene prediction. Our data already show a dramatic improvement in the amount of verified gene predictions obtained by simply combining the results of independent gene predictions with promoter predictions.

The next milestone in large-scale promoter analysis will be an in-depth in silico analysis of functional structures of promoters. Promoter function is defined by the specific arrangement of transcription factor binding sites. Promoters often contain subregions called transcriptional modules that are responsible for a specific transcriptional response of a promoter or a promoter group (Kel et al. 1999; Klingenhoff et al. 1999; Werner 1999). We previously showed that specific promoter modeling can yield functional insights into promoter organization in several cases (Frech et al. 1996, 1997, 1998), using a library of currently more than 100 computer models of transcriptional modules (GEMS Launcher). The module analysis will serve as the information base for ongoing research.