To determine the gross structure of Drosophila and human promoters, we determined the abundance of the four mononucleotides (1-mer; Figure 1a) across the 1,500 bp from -1,000 bp to +499 bp for 10,914 Drosophila promoters and compared these to distributions in 15,011 (UCSC) and 12,926 (DBTSS) human promoters (Figure 1b,c). Drosophila promoters are more A and T rich (56%) than human promoters (44%). In addition, Drosophila promoters had a peak for both A and T between -200 bp and the TSS, while the human promoters had a broad peak for both G and C centered at the TSS, suggesting a fundamental difference in global promoter architecture. The two human datasets show the same general distribution patterns, but the DBTSS set has more pronounced peaks and valleys at the TSS.

The CA dinucleotide is often associated with the TSS 15 and is often associated with a unique TSS 16. RNA polymerase is known to prefer an adenine in the +1 position 17. This provides an important quality control metric. A tight cluster of CA sites at the TSS would indicate that enough TSSs have been accurately assigned to permit analysis of other motifs. Figure 1d presents the CA dinucleotide distribution plotted at a single nucleotide resolution, rather than the 20 bp bin shown in Figure 1a-c. The CA distribution in both Drosophila and human promoters showed a spike exactly at the TSS (the A of the CA dinucleotide is at position +1 in the peak). The Drosophila CA spike at the TSS occurs in approximately 20% of all promoters while the spike is less pronounced in the human (UCSC) dataset (approximately 10%) and more pronounced in the human (DBTSS) dataset (approximately 40%). This CA peak is part of the initiator (INR) motif (TCAGTY) that is positioned at the TSS (see below). That CA is often present at the TSS suggests that the TSS has been appropriately assigned in many of the transcripts in both the Drosophila and human promoter dataset. If the CA peak is taken as a relative measure of the quality, or precise alignment, of the datasets, then the two human sets bracket the Drosophila set with respect to the accuracy of the positioning of the TSS.

Having validated the quality of the TSS assignments, we determined the distribution of all 8-mers in the set of Drosophila and human putative promoters to identify potential DNA binding sites for transcription factors that are localized relative to the TSS. A clustering factor (CF), describing the presence of a peak in the distribution of each 8-mer, was calculated three ways, by examining the distribution on both strands (CF), on the positive strand (CF⁺), and on the negative strand (CF^-). For these calculations we divided the 1,500 bp of genomic DNA, from -1,000 bp to +499 bp relative to the TSS, into 75 bins of 20 bp each (see Materials and methods).

When CF values were plotted against the bin with the maximum number of members for the Drosophila and human promoters, respectively (Figure 2a-c), all distributions showed similar patterns, with a grouping of DNA sequences that peak within 100 bp of the TSS. The highest CF values for all plots is 20 to 30, indicating that these 8-mers are approximately 20 to 30 times more abundant at one position relative to the TSS than elsewhere in promoters. In contrast to the similarity in CF values, when the data were plotted for CF⁺, (Figure 2d-f), a profound difference between Drosophila and both human datasets was revealed. Drosophila 8-mers have a maximum CF⁺ value of approximately 50 while the maximum CF⁺ for human sequences is approximately 20. This suggests that Drosophila has more 8-mers that occur preferentially on one strand of DNA, and that the Drosophila strand-dependent 8-mers have a higher degree of localization than their human counterparts. Control data, using 7th-order Markov random datasets, show a complete lack of clustering for any 8-mers for either human or Drosophila (data not shown).

To determine if an 8-mer has a peak in its distribution on only one strand of DNA, we compared the CF⁺ with the CF on the opposite strand (CF^-). In Drosophila, we identified two types of peaking 8-mers; those that peak on both strands and thus have similar CF⁺ and CF^- values (termed non-directional motifs (NDMs)), and 8-mers that peak preferentially on one strand (termed directional motifs (DMs)) and thus have significantly different CF⁺ and CF^- values (Figure 3a). Indeed, many motifs are randomly positioned on one strand and >20-fold enriched at a given position of the opposite strand. These two distinct types of motifs are potentially bound by proteins that have different roles in transcription regulation. The 8-mers with a high CF⁺ but a low CF^- contain directional information and could be binding sites for core promoter selectivity factors. In contrast, in both human promoter sets, we observed a significant number of 8-mers that peak on both strands (Figure 3b,c), and few that preferentially peak on one strand (as shown below, these are predominantly TATA and INR-like sequences). While the human DBTSS dataset contains a greater number of DMs than does the UCSC dataset, both sets are clearly more biased toward NDM than is the Drosophila dataset. These data suggest that there is a significant difference in the sequence organization of promoters between these human and Drosophila datasets.

Are the motifs that peak in humans similar to the motifs that peak in Drosophila? To answer this, we directly compared the CF values for all 8-mers between human and Drosophila (Figure 3d,e). The majority of 8-mers with high CF values are different between the two species. In contrast, 8-mers with the largest CF values are common between the two human datasets (Figure 3f), lending confidence to the idea that the differences between the two species are real.

To determine the statistical significance of the CF⁺ values, we converted the CF⁺ into a probability term using the 8-mer frequencies observed in the 10,914 Drosophila promoter dataset. The probability term, P, represents -log₁₀(1 - p), where p is the area under the normalized curve of the distribution of CF_expt. A high P value indicates that it is very unlikely that the peak for the 8-mer occurs by chance. A plot of the P values versus the most populated bin number (Figure 4a) shows a group of 8-mers near the TSS whose distributions are very unlikely to occur by chance. We analyzed the 298 8-mers that have a P value ≥ 16. All these 8-mers had peaks centered between -100 bp and +40 bp. As illustrated in Figure 4a, P ≥ 16 is a conservative cutoff. We plotted CF⁺ versus CF^- for these 298 sequences to examine their strand specific localization (Figure 4b). DMs (black circles) predominate, but NDMs (red circles) were also identified.

The 298 8-mer sequences were manually grouped into 15 families and a consensus motif was determined for each family (Figure 5). The placement of an 8-mer into a particular motif was guided by: the similarity amongst DNA sequences; the shape of the distribution histogram; the peak position relative to the TSS; and whether the 8-mer was directional or non-directional. The total number of 8-mers in each of the 15 motifs varied dramatically, with over one-third of the 298 8-mers representing variations of the INR motif (TCAGTY) and 8 motifs were represented by 5 or fewer 8-mers. We determined the abundance of the 15 motifs by counting unique promoters that contained a motif in the peak (Figure 4c). A total of 6,067 promoters contain one or more of the 15 motifs. The most abundant motif is the non-directional DRE, found in 15% (1,593) of Drosophila promoters, followed by directional INR, found in 14% (1,501) of promoters. The least abundant motif identified, DMp5, is found in 0.7% (80) of all promoters.

Figure 6 presents the distribution of each of the 15 consensus motifs, showing the number of occurrences on each DNA strand. To gain more insight into how constrained motif position is relative to the TSS, we examined the distribution of the 15 DNA motifs at a single base-pair resolution. The inserts in Figure 6 show the single base-pair distribution plots for the motifs in the region -100 to +100 relative to the TSS. Five of the DMs (Figure 6a-e) are positioned at a single base-pair resolution relative to the TSS while the other five DMs (Figure 6f-j) and the five NDMs (Figure 6k-o) are spread across a broad region of up to 50 bp, though they all clustered near the TSS. We thus classified the DMs as either precise or variably positioned. The DMs are named DMp1 to 5 (for directional motif precise) and DMv1 to 5 (for directional motif variable). The NDMs are named NDM1 to 5. Where a motif has a previous common name we use that name, for example, DMp1 is TATA, DMp2 is INR, DMp4 and DMp5 are DPE-like, NDM1 is GAGA and NDM4 is downstream responsive element (DRE). The single base-pair resolution plots not only reveal the precise versus variable positioning of the motifs, they also reveal the power of the initial analysis based on 20 bp bins. Many of the motifs (DMvs and NDMs) would not have been identified at a single base-pair resolution. Also, the number of promoters identified that contain a specific motif is much greater at a 20 bp resolution than a 1 bp resolution (for example, for INR there are approximately 1,500 versus approximately 400).

To further examine the localization of DNA sequences at a single base-pair resolution, we examined the CF⁺ values of all 6-mers for both Drosophila and human promoters (Figure 7). We chose 6-mers to produce enough occurrences at each base pair position to be able to determine peaks reliably. The Drosophila data (Figure 7a) showed three distinct regions in which individual 6-mers were preferentially localized. Examination of the DNA sequences that cluster around each of these three positions indicated they can be grouped into a single motif that is localized at a specific base-pair position relative to the TSS. The three motifs are TATA, INR and DPE. Where promoters have two of these motifs, they are precisely positioned relative to each other (Figure 7d).

The clustering of 6-mers at a single base-pair resolution in the UCSC human promoters showed generally lower CF⁺ values and only two peaks corresponding to the TATA and INR positions (Figure 7b). While the DBTSS dataset (Figure 7c) showed more pronounced peaks than the UCSC dataset, it still failed to show a clear DPE peak. Examination of the sequences localized under the main human (DBTSS) peaks produced a result similar to that seen form Drosophila. The sequences lying under the TATA peak were exclusively TATA-like sequences. The sequences under the INR peak represented INR variants localized exactly at the TSS and other NDMs, predominantly erythroblast transformation specific (ETS), localized close to the TSS. However, the variety of INR sequences that localized in the human dataset was greater than that seen for the Drosophila data. Attempts to identify distinct human INR motifs six nucleotides or greater were unsuccessful due to the wide degeneracy in sequences that surround the prominent central CA core.

We examined if motifs that peak in Drosophila also peak in human and vice-versa. Of the 15 Drosophila motifs that peaked, four also localized in human promoters (TATA, INR, DPE1 and NDM2; Figure 8a,b,d,l) with INR, DPE1 and NDM2 occurring at much lower frequency in human promoters. While both the human and Drosophila promoters showed a clear overabundance of the CA dimer at the TSS (Figure 1d), we were previously 11 unable to detect an INR signal in human promoters using the degenerate human consensus sequence (YYANWYY). However, mapping the Drosophila INR motif (TCAGTY) to human promoters does produce a weak peak at the TSS in the UCSC dataset and a more pronounced peak in the DBTSS dataset (Figure 8b). Analysis of this peak at a 1 bp resolution (Figure 8x) revealed that both human datasets contain significantly fewer of these precisely positioned elements than does the Drosophila dataset. This result suggests that this TCAGTY motif plays a less significant role in human gene transcription than it does in Drosophila, and agrees with previous findings that the human INR is more degenerate than its Drosophila counterpart. It should be noted that in all cases, the motifs that contained a peak in one human dataset also showed peaks in the other human dataset, although the DBTSS dataset showed more pronounced peaks. This confirms both the qualitative similarity of the two datasets and the suggestion that the DBTSS data contains greater numbers of accurately positioned TSSs. Of the eight motifs previously identified to abundantly peak in humans 11, only TATA also peaked in Drosophila promoters (Figure 9).

In comparing the distributions of the Drosophila and human motifs, it is apparent that some sequences, even when they occur outside of the peak, display different abundances for the two organisms. This is true for DRE (Figure 8n), which peaks in Drosophila but is also a highly abundant motif outside of the peak (total of 7,058 across 1,500 bp of 10,914 promoters). In humans, there is no indication of any clustering, and this element is also very rare (total of 1,015 across 1,500 bp of 15,011 promoters). The reciprocal observation is made for human promoters, where SP1 (Figure 9h) is characterized by a very large peak and is also abundant outside of the peak but is virtually absent from Drosophila core promoters. In contrast, the INR (Figure 8b), which peaks in both organisms, albeit on different scales, shows very similar total abundance in both organisms (a total of 17,377 and 20,320 occurrences across 1,500 bp, in 10,914 and 15,011 promoters, for Drosophila and human, respectively).

NDM5 (CAGCTSWW) is a derivative of the general DNA sequence termed an E-box (CANNTG) that is bound by B-HLH-ZIP transcription factors, including the oncogene Myc|Max. A recent paper 18 has shown that an E-box sequence is located near the TSS of Drosophila genes. The sequence CACGTG is the core of the upstream stimulatory factor (USF) sequence previously identified in humans to peak near the TSS 11. We compared the distribution of these related sequences in Drosophila and human. The USF consensus sequence (TCACGTGR) does not show any clustering in Drosophila (Figure 9b). However, the 6-mer E-box variants CACGTG and CAGCTG have peaks in both human and Drosophila promoters (Figure 10a,b). In Drosophila, the sequence CACGTG peaks downstream of the TSS while in human it peaks upstream of the TSS. The E-box variant CAGCTG peaks in both human and Drosophila just upstream of the TSS. Figures 9c,d highlight two E-box 8-mer variants with dramatically different peaking properties where sequences outside a conserved 6-mer define the peaking properties of the 8-mer. The sequence RCACGTCY peaks only in Drosophila while YCACGTGR peaks only in human, suggesting that distinct B-HLH proteins bind these related sequences.

We examined correlations in the occurrence of the 15 peaking motifs in Drosophila to gain insight into their potential combinatorial or redundant function. Table 1 presents a matrix showing: the number of promoters that contain one motif in a peak that also contain a second motif in a peak (a); the frequency of this co-occurrence (b); and the probability (c). There is a complex pattern of positive and negative correlation for individual motifs, suggesting that combinations of motifs act to regulate core promoter function.

For the precisely positioned directional motifs (DMp1 to 5: TATA, INR, INR1, DPE, and DPE1), promoters that contain INR also preferentially contain either the TATA or DPE sequence. However, TATA and DPE motifs negatively correlate. All five members of the DMp class negatively correlate with some or all of the DMv class. DMp1 to 5 positively correlate with three of the NDMs (NDM1 to 3) but negatively correlate with NDM4 and NDM5.

The five variably positioned directional motifs (DMv1 to 5) have both positive and negative correlations amongst themselves and with the NDMs. The DMv class members positively correlate with NDM4 and NDM5 and negatively correlate with NDM1 to 3, correlations that are exactly the opposite of those observed for the DMp class (see above). On average, members of the NDM class positively correlate with each other. Positive correlations between motifs suggest the possibility of physical interactions between the proteins that bind the co-occurring DNA motifs. Negative correlations, as are observed between the precisely positioned DMs (DMp) and the variably positioned DMs (DMv), suggest that the proteins that bind them have distinct functions.

The non-random distribution of individual motifs and motif combinations at core promoters strongly suggests that the identified motifs are biologically significant and promoters that share the same motif in a peak may also share similar biological functions. To evaluate this possibility, we calculated statistical over- and under-representation of 5,200 Gene Ontology (GO) annotation terms 19 for Drosophila genes whose promoters contained any of the 15 motifs, either within the peak or elsewhere in the promoter region. We found highly significant correlations (p < 10^-4) for each motif only when they occurred in the peak (Figure 11a). With one exception, the simple presence elsewhere within the 1,500 bp promoter region does not correlate with GO terms, demonstrating that the position of a motif in the promoter is critical for predicting biological function, as was observed in human promoters 11. The directional positioned motifs, DMp and DMv, not only co-occur in promoters with either NDM1 to 3 or NDM4 and NDM5, respectively, but also correlate with similar GO terms. This indicates a combinatorial code of motifs at core promoters directing batteries of genes.

Additional insight can be inferred by examining individual GO terms that correlate. For example, Drosophila mitochondrial ribosomal genes contain the E-box (p < 10^-8). In contrast, promoters of human mitochondrial ribosomal genes contain the ETS motif, a motif that peaks in human but not in Drosophila. Thus, even though the mitochondrial ribosomal genes are highly conserved, their regulation is evolving.

If core promoter motifs are used to drive the expression of gene batteries participating in a common biological process, this should be evident in global gene expression profiles. We turned to Drosophila mRNA expression patterns determined by micoarray experiments 2021 to evaluate whether genes that are co-expressed have the same motif in their promoters. Figure 11a shows correlations between all 15 motifs, either in the peak or elsewhere in the promoter region, and gene expression in testis (male germline), ovary (female germline), and soma. The presence of TATA in the peak in the promoter positively correlates with gene expression in somatic tissue but negatively correlates with expression in germline tissue. The presence of positioned DMv3 to 5, and DRE in promoters positively correlates with female germline expression and negatively correlates with male germline expression. If the motif occurs outside the peak, few correlations are observed, supporting the conclusion that motif position is functionally important.

We see more striking correlations between promoter motifs and mRNA expression in the embryonic and adult stages of Drosophila development that express different sets of genes. Figure 11b presents a hierarchal clustering of mRNA expression for 89 samples from a survey of gene expression in embryos and adults for promoters containing any of the 15 motifs (either in or outside the peak). Genes with motifs in the peak show strong mRNA expression differences between embryo and adult samples, suggesting that these motifs help direct the differential utilization of the genome between embryos and adult. Genes with promoters containing DMv1 to 5 and co-occurring NDM4 and NDM5 are preferentially active in the embryo. In contrast, genes with promoters containing the three abundant precisely positioned directional motifs (TATA, INR, and DPE) and the co-occurring NDM1 to 3 are preferentially active in the adult.

Both Drosophila and human promoters have a CA peak exactly at the TSS in a significant number of promoters. About 2,100 Drosophila promoters contain the CA sequence at the TSS but only 400 of these are part of the consensus INR sequence (TCAGTY). We examined the remaining promoter sequences for related INR sequences and identified 4 more motifs, resulting in 1,080 promoters with INR related sequences exactly positioned at the TSS. To evaluate if these INR related sequences correlate with distinct functions or are variants of a single motif, we investigated the correlation of the INR variants with different biological properties by examining GO terms and mRNA expression properties. Figure 12a shows that the variant INR motifs have distinct patterns of enrichment with categories of GO terms. Similarly, the developmental mRNA expression analysis (Figure 12b) indicates that one of the INR motif variants (BCACWS) is preferentially associated with genes with embryonic expression while the other variants are preferentially associated with adult expression genes. While some of the GO categories enriched for specific INR variants (for example, mesoderm development) appear at odds with the adult/embryo expression patterns, the overall impression suggests that these variant INR sequences are functionally distinct and may be recognized by distinct proteins. The discrepancies between the GO term enrichment and adult/embryo expression patterns can be explained if one assumes that the preferential use of INR signals is not absolute. Thus, even though there is a general trend toward preferential use of different elements at different stages in development, certain genes may use the 'adult INRs' during embryogenesis.