We begin by defining the measurement of two kinds of tissue specificity, 'overall' tissue specificity and 'categorical' tissue specificity. (To avoid confusion we will always use the words 'specificity' and 'specific' to refer to the degree of tissue-restricted expression a gene exhibits and never as a synonym for the word 'particular'.) Overall tissue specificity ranks a gene according to the degree to which its expression pattern differs from ubiquitous uniform expression. We use the term 'ubiquitous' expression to mean expression at any level above background in all tissues. Categorical tissue specificity places special emphasis on a particular tissue of interest and ranks a gene according to the degree to which its expression pattern is skewed toward expression in only that particular tissue. In both cases, a gene's specificity to a tissue, cell type or other condition is decreased as the gene is more uniformly expressed in a wider variety of conditions. In addition, the categorical tissue specificity should decrease as the tissue of interest becomes a smaller component of the overall expression pattern of the gene.

Given a static multi-tissue expression profile for a gene, there are at least two dimensions along which we can assess the profile to measure tissue specificity. The first dimension is the number of tissues that express the gene above some background level. It can be argued that this dimension measures tissue restriction, that is, a gene shows restricted expression if it is expressed in only a subset of tissues. The second dimension is the uniformity of expression over all tissues that express the gene. A gene that shows significant non-uniform expression is exhibiting tissue-dependent regulation, in addition to any tissue restriction that may be occurring. We assume that a gene that exhibits no tissue-specific regulation will be expressed at the same level in every tissue. We do not assert that such genes are not regulated, only that they are regulated in a way that is not sensitive to tissue.

The term 'most tissue-specific' will refer to the range of genes that are closer to the extreme of expression in a single tissue than to the extreme of ubiquitous uniform expression. We will refer to genes close to the uniform and ubiquitous end as either 'least tissue-specific' or 'nonspecific' though the latter term may not be strictly true. The range in the middle will be termed 'semi-tissue specific'. The term 'housekeeping' has been applied to genes that are widely expressed and may show little tissue-specific changes in expression level. We can use such genes as an example of genes that will tend to be ubiquitously and uniformly expressed and thus ought to be nonspecific on average. We will use the phrase 'gene sharing' to refer to the situation that occurs when a gene is tissue-specific, and is expressed in a small number of tissues that can be said to share the gene.

We used two gene-expression datasets to evaluate our methods; Affymetrix-based data from the GNF Gene Expression Atlas (GNF-GEA) 22 and the distribution of source tissues for EST libraries in the clusters and assemblies of ESTs in the DoTS mouse and human gene index 33. As described in Materials and methods, the GNF-GEA data were used as provided; EST counts in the DoTS gene index were adjusted with pseudocounts and normalized to account for the different number of ESTs sampled from each tissue across all libraries. Given expression levels of a gene in N tissues, we defined the relative expression of a gene g in a tissue t as p_t|g = w_g,t/∑_{1 ≤ t ≤ N}w_g,t where w_g,t is the expression level of the gene in the tissue. The entropy 34 of a gene's expression distribution is H_g = ∑_{1 ≤ t ≤ N} - p_t|g log₂(p_t|g). H_g has units of bits and ranges from zero for genes expressed in a single tissue to log₂(N) for genes expressed uniformly in all tissues considered. The maximum value of H_g depends on the number of tissues considered so we will report this number when appropriate. Because we use relative expression the entropy of a gene is not sensitive to the absolute expression levels. To measure categorical tissue specificity we define Q_g|t = H_g - log₂(p_t|g). The quantity -log₂(p_t|g) also has units of bits and has a minimum of zero that occurs when a gene is expressed in a single tissue and grows unboundedly as the relative expression level drops to zero. Thus Q_g|t is near its minimum of zero bits when a gene is relatively highly expressed in a small number of tissues including the tissue of interest, and becomes higher as either the number of tissues expressing the gene becomes higher, or as the relative contribution of the tissue to the gene's overall pattern becomes smaller. By itself, the term -log₂(p_t|g) is equivalent to p_t|g. Adding the entropy term serves to favor genes that are not expressed highly in the tissue of interest, but are expressed only in a small number of other tissues. As described earlier, we want to consider such genes as categorically tissue-specific since their expression pattern is very restricted. Figure 1 shows examples of patterns of GNF-GEA expression data for different values of H_g and Q_g|t. The top five genes specific to mouse amygdala, lymph node, and liver as assessed by this data are listed in Table 1. Tables of H_g and Q_g|t values for all genes in all tissues in the GNF-GEA datasets are available in Additional data files 1 and 2.

To compare results from microarray and EST-based expression data we mapped the tissues from the GNF-GEA study to the hierarchical controlled vocabulary of anatomical terms used by DoTS and chose a set of 45 tissue terms grouped into 32 groups shown in Table 2. In both cases, the vast majority of genes are widely expressed as measured by H_g as shown in Figure 2a. Of the 7,714 probe sets in the GNF-GEA data with an average normalized intensity value above 50 arbitrary units (AU), 6,167 (80%) of genes had H_g ≥ 4 bits, which implies expression in at least 16 tissues and typically corresponds to wider, but uneven, expression. Only 87 (2%) of genes had H_g ≤ 1.5 bits, which corresponds to expression in as few as three tissues. Both microarray- and EST-based data yielded similar overall curves. The EST curve peaked at a lower H_g than the microarray curve. This was due to the small numbers of EST sequences in some of the tissues we considered; EST counts for tissues ranged from 1,933 in the adrenal gland to 331,582 in the central nervous system (CNS). Genes that are ubiquitously expressed may not have ESTs from several of the lightly sequenced tissues, making them appear to have more restricted expression, and hence a lower entropy, than they really do. Figure 2b shows the correlation between estimates of H_g derived from microarray and EST data. Visual inspection of the plot reveals that while there are no strong contradictions between the two methods, quantitative agreement is limited. Detailed analysis shows that the standard deviation of the difference of paired H_g values is 0.61 bits. Under the null hypothesis that the estimates from the two data sources are totally uncorrelated the average standard deviation was found to be 0.91 bits. We can reject the null hypothesis (P < 10^-5 as estimated by Monte Carlo methods). The distribution of Q_g|t for selected tissues is shown in Figure 2c. These curves can be used to characterize tissues in terms of the number of tissue-specific genes and the amount of gene sharing; for example, liver has a relatively large number of genes shared with a small number of other tissues. In contrast, there were no genes in this set that are uniquely expressed in the amygdala.

It is important to determine how well the H_g and Q_g|t statistics can be estimated from a dataset to determine the smallest meaningful difference in scores and to guide interpretation of gene rankings. To assess the standard deviations of and H_g and Q_g|t, we sampled from the replicates in the GNF-GEA microarray data to compute a large number of H_g values for each probe set. We found that the standard deviation for H_g was less than 0.2 bits for 97% of genes. Q_g|t was not estimated as well; the standard deviation was 1 bit or less for 95% of gene and tissue pairs. This was probably due to the high standard deviation of the -log₂(p_t|g) term for low expressing gene-tissue pairs. We found much more variation when we measure reproducibility by considering genes that have two or more probe sets (and therefore two or more different transcripts) in the microarray data. In this case, the standard deviation of H_g estimates was as high as 1 bit for 97% of the genes but less than 0.3 bits for about 70-80% of the genes. We chose a minimum of 1 bit for H_g bins and 2 bits for Q bins in the rest of the analyses that require binning. This bin size ensured that most of the genes are in the proper bin and thus the bin could be reliably used to determine associations with the tissue specificity of a class of genes.

A test of the H_g and Q_g|t statistics is to determine values for a set of nonspecific genes such as housekeeping genes. A list of 797 human housekeeping genes 35 was evaluated using these statistics based on the GNF-GEA dataset using RefSeq accession numbers to identify appropriate probe sets. The housekeeping genes had a mean H_g = 4.6 ± 0.27 bits in a set of 27 tissues with a maximum H = lg(27) = 4.75 bits; thus they are nonspecific as expected. Interestingly, a small number of these genes did show some degree of tissue specificity yet were ubiquitously expressed. For example, the median expression of NM_021983 the major histocompatibility complex, class II DR beta 4 gene (32035_at) is approximately 200 AU, but it shows much higher expression in a small set of tissues (spleen, thymus, lung, heart and whole blood), which lowered its entropy. A more extreme case is NM_001502 glycoprotein 2 (zymogen granule membrane protein 2), which is expressed between 250 and 1,000 AU in all tissues except pancreas, where it is expressed at 34,183 AU. This is a ubiquitously expressed gene that entropy categorizes as specific since it showed such extreme tissue-specific induction. The housekeeping genes had a mean Q_g|t = 9.5 ± 0.14 bits in the same set of tissues. The expected Q value for a uniformly and ubiquitously expressed gene is 2 lg(27) = 9.5 bits. Thus, the H_g and Q_g|t statistics successfully captured the expected expression properties of housekeeping genes.

Although the housekeeping genes assessed above have relatively high entropies, they do show some small degree of overall tissue specificity. We therefore sought to determine how many genes show evidence of tissue-dependent regulation. Since random biological and experimental variation introduce fluctuations in the expression levels of genes, we made a probability model of the effect of these fluctuations on the observed entropy. The experimental variability was estimated from the GNF-GEA data using all normal tissues. The random tissue-to-tissue biological variability was modeled by assuming that each gene has an average expression level across all tissues and that the log base 2 of the tissue-dependent fold changes from the average level follow a normal distribution with mean equal to zero and some unknown, but 'small', standard deviation(s). We obtain a conservative estimate of the number of genes showing evidence of tissue-dependent regulation by using s = 0.5, which allows for a relatively large amount of variation; up to 1.4-fold tissue-to-tissue variation around the mean expression level in about 63% of tissues and larger changes in the remaining tissues. As a threshold for selecting genes with tissue-dependent expression, we choose H_g = 4.52 bits which has a p-value of 0.005 under the null hypothesis that all genes are uniform. We then find that 5,837/8,703 (67%) of human genes have entropies less than this and so are probably regulated in a tissue-dependent manner. If we use a more stringent definition of uniform expression that allows half as much variation in tissue-to-tissue expression levels (s = 0.25), then the threshold is H_g = 4.62 bits and we find that 7,584/8,703 (87%) of human genes show evidence of tissue-dependent regulation. Similar results are found in mouse using all 42 distinct tissues, where the corresponding thresholds are H_g = 5.24 bits (s = 0.5) and H_g = 5.35 bits (s = 0.25) and the fractions of genes showing tissue-dependent expression are 5,467/7,913 (69%) and 7,482/7,913 (94%) respectively. Thus we conclude that most genes show evidence of tissue-dependent expression levels.

A test of Q_g|t with respect to specific genes is to evaluate the tissues in which they rank highly (that is, have low Q) for consistency. This was accomplished by clustering tissues with similar tissue-specific genes and inspecting the clusters formed. We used 27 normal human tissues and, separately, 39 tissues from the GNF-GEA data for mouse and selected the genes (N = 3,768 human and N = 1786 mouse) that express at least 200 AU in at least one tissue and have Q_g|t = 7 in at least one tissue. With these genes, we made a consensus hierarchical clustering of the tissues as shown in Figure 3. We found that the tissues in the nervous system, reproductive structures (excluding testis), immune system, and digestive system reliably cluster together in both species. In addition, skeletal muscle and heart clustered in mouse; the human survey did not have skeletal muscle. These results suggest that Q_g|t is correctly identifying tissue-specific genes. Interestingly, testis is an outlier in both trees, indicating that the collection of genes expressed in testis are distinct from any other tissue or organ. Furthermore, H_g and Q_g|t can also be used in conjunction with a tissue hierarchy to answer more complex questions about the tissue distribution of genes such as 'what genes are specific to the brain but are widely expressed throughout the brain?' In Table 3 we list the top five mouse genes expressed specifically but uniformly across three of the highlighted groups in Figure 3b.

It has been proposed that CpG islands are predominantly associated with promoters of housekeeping genes 2. We performed a quantitative test of this hypothesis using the GNF-GEA data and determining the frequency of CpG islands in promoters as a function of H_g. We considered only predicted CpG islands that span the start of transcription (see 3 for a justification of this definition), and genes that expressed at least at the median level of 200 AU (that is, were moderately expressed) in at least one tissue, and were represented by a single probe set on the Affymetrix chip used in the GNF-GEA experiments. Promoter sequences were obtained from DBTSS and were based on the 5' ends of full-length transcripts 17. We found that there is a strong, roughly linear, correlation between a gene's entropy H_g and the probability that the gene will have a predicted start CpG island as shown in Figure 4. Start CpG islands were associated with only nine of the 100 most tissue-specific human genes as compared to 80% of the least tissue-specific genes. Similar numbers were found for mouse (7% start CpG island frequency for the 100 most tissue-specific genes; about 64% for the least tissue-specific genes). A comparison of CpG islands from the most and least tissue-specific genes did not reveal any significant difference in the overall base composition, or ratio of observed to expected CpG dinucleotides. The distribution of the position of the 5' end point of CpG islands was also very similar for the most and least tissue-specific genes though CpG islands tend to start further upstream in the least tissue-specific genes (data not shown).

Another group of genes observed to be associated with CpG islands are those expressed in the early embryo 3 from the fertilized egg to the blastocyst. The question arises as to whether there is an association of genes having start CpG islands and the developmental stage of expression (that is, embryonic versus adult) in addition to the one for tissue specificity. We investigated this possibility in the mouse using DoTS 33 EST and mRNA assemblies by tabulating the number of DoTS genes that contain at least two ESTs from a mouse early embryo library as shown in Table 4. We considered 933 genes with start CpG islands (CGI+) and 1,007 genes without start CpG islands (CGI-) that were expressed in the adult. If there were no developmental bias, this distribution of CpG+ and CpG- genes should be maintained in genes expressed in the embryo. However, only 139 (14%) of the CGI- genes were expressed in the early embryo in contrast to 365 (39%) CGI+ genes (P = 3 × 10^-70 exact binomial). Therefore, a gene expressed in the adult was 2.8 (= 0.39/0.14) times more likely to be expressed in the early embryo if it contained a start CpG island. Furthermore, the most tissue-specific genes expressed in the adult were four times more likely to have been expressed in the early embryo if their promoter contained a start CpG island. These results strongly suggest that CpG islands are promoter features for both embryonic and the least tissue-specific genes.

Analysis of base-composition profiles of promoters provides clues to common features, including motifs associated with promoter categories. We examined the base composition profiles of human promoters of high (0 ≤ H_g ≤ 3.5 bits) and low (4.4 ≤ H_g ≤ 4.71 bits) tissue-specificity genes. We considered CGI+ and CGI- genes separately, as it is clear the presence of a CpG island will strongly influence the base composition and that the fraction of start CpG islands varies with entropy. In addition, the presence of a start CpG island may indicate a different regulation mechanism related to either tissue specificity or embryonic expression (or both). The number of promoters from DBTSS in these four classes that were used in the analysis were: 310 CGI- and 129 CGI+ high specificity; 342 CGI- and 1,501 CGI+ low specificity. Genes that have only non-start CpG islands represented a minor component and were not included in this analysis. We used the full set of normal tissues in the first GNF-GEA microarray study for human and mouse. Base composition profiles with 10 base-pair (bp) windows are shown in Figure 5 for human genes. Each of the features we report were observed in human and mouse (unless noted otherwise) and compare G to C or A to T over spans of at least 10 positional bins; the probability of observing a feature at least this long by chance is less than 0.5¹⁰ which is equivalent to 0.001. Promoters of CGI+ genes (Figure 5a,b) shared features but could also be distinguished on the basis of tissue specificity. A common feature of CGI+ promoters was the increase in C+G content that starts at 1,000 bp upstream of the transcription start site and continues at 200 bp downstream. The C+G bias reached p(C+G) = 0.7 at the start of transcription and continued into the 5' UTR. Nonspecific (Figure 5c) and tissue-specific (Figure 5d) CGI- genes still showed a C+G bias around the start of transcription, but it was much smaller in magnitude at p(C+G) = 0.54. The low specificity CGI+ genes (Figure 5a) showed upstream base composition biases that were not found in any of the other three gene classes. There was a preference for C over G (p(C) > p(G)) in the (-350, -150) region and also a preference for p(A) > p(T) in the -600, -200 region in human (this region is located (-400, -150) in mouse). In tissue-specific CGI+ (Figure 5b) genes the strong C+G bias held but p(C) = p(G), except for the (+50, +100) region where p(C) > p(G). These base-composition differences observed between nonspecific and tissue-specific promoters over regions of hundreds of base-pairs, even in the context of a CpG island, suggest different structural features and regulatory mechanisms for these CGI+ classes.

Most striking were differences between nonspecific and tissue-specific promoters that are independent of the presence of a CpG island. A sharp spike in the proportion of A and T was seen in the (-50,-1) region for all classes but was most pronounced in the tissue-specific promoters (Figure 5b,d). These spikes correspond to the presence of a TATA box and suggest a correlation of this motif with tissue-specific genes (explored more fully later). Conversely, all low-specificity genes (Figure 5a,c) shared a common feature in the (+1, +200) region where p(G) > p(C) and p(T) > p(A) that was not seen in tissue-specific genes (Figure 5b,d). As shown later, this low-specificity feature could be partially explained by the presence of a YY1 motif. These base-composition differences observed between nonspecific and tissue-specific promoters are likely to indicate motifs that distinguish the two classes.

We next examined the distribution of basic core promoter features: the TATA box, the initiator element, and two binding sites for selected ubiquitous transcription factors, Sp1 and YY1, to see if their presence in the proximal promoter was correlated with the tissue specificity of a gene. Two approaches were taken using different datasets and motif-searching methods that gave similar results, providing independent confirmation of results. First, we searched for core motifs using weight matrix hits in promoters of genes selected using H_g calculated from the GNF-GEA data. Second, we searched for core motif consensus sites in promoters of genes selected using Q_g|t calculated from EST data.

We grouped the human genes that expressed at least 200 AU (average value) in the GNF-GEA data by entropy and start CpG island status. The number of genes in each category is shown in Table 5 along with a summary of results. We used alignments of position-specific scoring matrices and scoring thresholds included in the Eukaryotic Promoter Database (EPD) 36 to identify the TATA box and initiator element. Matches to these motifs were preferentially located at the expected positions relative to the transcription start site based on the ratio of the number of observed set to the expected number using a set of random sequences with the same position-dependent base composition as each of the promoters.

We searched for the TATA box in the (-45, -10) region where the average observed/expected ratio for the TATA box was 3.1. As shown in Table 5, the most-specific CGI- genes were six times more likely to have a TATA box than the least-specific CGI+ genes (117/215 (54%) versus 183/2072 (9%), P ≈ 0 exact binomial). Similar numbers are found in mouse (52%/11% = 4.7) This trend also holds within CGI- genes and CGI+ genes. The most specific CGI- genes were three times more likely to have a TATA box than the least specific CGI- genes (117/215 versus 110/607, P ≈ 0 exact binomial). While less common in CGI+ genes, TATA boxes were still almost four times as likely to be found in the most specific CGI+ genes than the least specific CGI+ genes (19/56 versus 183/2,072, P = 2 × 10^-7 exact binomial). Thus TATA boxes are clearly associated with tissue-specific genes and provide a second axis (with CpG islands) for distinguishing between the most and least specific genes.

In contrast, the frequency of occurrences of the initiator element (Pol II binding site) was roughly constant across all tissue-specificity classes for both CGI+ and CGI- genes. We searched for the initiator element in the (-10, +10) region. It occurred in 762 of 1,118 (68%) of CGI- genes and 1,273 of 2,434 (52%) of CGI+ genes. Similarly, it occurred in 149 of 215 (69%) of the most specific genes and 388 of 607 (64%) of CGI+ genes. The observed frequency of TATA+/Inr+ promoters was not significantly different from the expected rate assuming independence of the two individual features (data not shown).

Sp1 3738 is a ubiquitous transcription factor with a G-rich binding site with consensus sequence GGGCGGG that might explain the observed G-richness of the 5' UTR in non-specific genes. We used the GC-box weight matrix and scoring threshold from EPD 36 to identify Sp1 sites. We found that Sp1 sites are preferentially located in the (-150, +1) region in all sets of genes where they occurred on average at twice the expected rate in agreement with previous findings 36. In both human and mouse, Sp1 sites were rarely found in the 5' UTR despite the G-richness of this region; they occurred at the expected rate of between 2 and 5%. Thus Sp1 sites were not the cause of the G-richness in the 5' UTR.

Sp1 sites are associated with CpG islands but are an important component of GGI- promoters as well. Considering just the (-150, +1) region, Sp1 sites occurred in 1,105/2,434 (45%) of human CGI+ gene promoters, and 316/1,118 (28%) of CGI- genes at about 2.5 to 3.0 times the expected frequency in both cases. Frequencies in mouse are 927/2075 (45%) of CGI+ promoters and 464/1652 (28%) CGI- promoters. Sp1 sites were also weakly associated with the least specific genes occurring in 1,105/2,679 (41%) of these genes as compared to 94/271 (32%) in the most tissue-specific genes (P = 0.016). Similar numbers are found in the mouse; 38% of the least specific and 26% of the most specific promoters have Sp1 sites. Thus, although Sp1 shows a preference for the least tissue-specific promoters, it is not a strong predictor of the tissue specificity of a gene.

The transcription factor YY1 5678 is also ubiquitously expressed and is thought to bind close to 39 and downstream of the transcription start site. There is evidence that the function of YY1 depends on its orientation 40. The location and G-richness of the reverse complement consensus sequence (AANATGGCG) make YY1 a candidate for explaining the prominent G > C feature in the (+1, +200) region of low-specificity genes. We consider YY1 because a YY1-like motif was frequently included among the most statistically significant motifs identified by the motif discovery programs AlignACE 41 and MEME 42 in the (+1, +60) region of nonspecific CGI+ promoters (Figure 6a). Our form is most similar to the activating form 43, which may be associated with low-specificity genes. Because of the demonstrated functional sensitivity to the orientation of binding sites we considered each orientation separately. Indeed, as shown in Figure 6b we found each orientation exhibits different position preferences. Sites in the reverse orientation (YY1_r) were preferentially located in the (+1, +25) region but with some elevated levels to +80 bp. Start positions of sites in the forward orientation (YY1_f) showed a very sharp preference for -3 bp, which probably represents a YY1-like initiator sequence reviewed elsewhere 44. Both orientations were found predominantly in the least specific genes (Table 5). YY1_f initiator sites are rare; only 55/2,679 (2%) were found above background in human low-specificity genes. The rate in mouse, 22/2,832 (0.8%) of low-specificity promoters, is even lower. The YY1_r sites are more common and were found above background in 217 (8%) of the 2,679 least specific genes. YY1_r sites were more common in CGI+ genes than in CGI- genes (202/2,072 (10%) versus 15/607 (2%) P = 3.7 × 10^-9 two-population binomial). The corresponding rates in mouse confirm these observations; 178/2,832 (6%) for all low-specificity genes and 152/1,779 (9%) in CGI+ and 26/1,053 (2%) of CGI- low-specificity promoters. These YY1-like sites therefore constitute a feature strongly associated with the least specific genes and may partially explain the observed G > C ratio in the (+1, +200) region.

A second analysis of TATA box and Inr motifs was done to determine if the association of the TATA box with tissue-specific genes is also found in genes ranked by Q and is robust to using EST data as well as promoters that did not specifically rely on full-length cDNA clones. The definition of Q implies that genes with a particular Q-value can have a variety of H_g values and thus it may be more difficult to identify features related to tissue specificity. We tabulated all DoTS genes that contained at least two ESTs from an islet-cell library then ranked the genes by Q_pancreas computed using EST counts. We used Q_pancreas ≤ 7 bits as the criterion for selecting pancreas-specific genes which we grouped into 2-bit Q intervals. For comparison we selected 50 genes with Q_pancreas = 8.5 bits, and 50 genes with 10 ≤ Q_pancreas ≤ 10.6 bits. Genes with high specificity for the pancreas (0 ≤ Q_pancreas ≤ 2 bits, N = 9) preferentially had TATA boxes (8 of 9) with half of these also having an initiation element (4 of 9; Figure 7a). With decreasing specificity, the fraction of genes containing TATA boxes drops with only18 of 81 (2/9) genes with Q > 6 bits having TATA boxes. Thus, the strong correlation of TATA boxes with specific genes found with H_g and microarray data was also seen with Q and EST data for pancreas-expressed genes. Also consistent is the observation that initiator elements were found at similar frequencies (around 60%) across all specificity classes (Figure 7b). Similar patterns were observed in other tissues (data not shown).

The consistency of findings for the TATA box with human islet genes based on Q and ESTs was next tested with orthologous genes in mouse. This test provides a measure for whether the global pattern observed (TATA box with tissue-specific genes) is also found for the same set of genes in another mammal. We also added bins of genes with higher Q-values that represent more widely expressed genes. For each human gene, the orthologous mouse gene was determined (see Materials and methods for details) and analyzed as described above. Overall, 18.8% of the human genes and 22.9% of the mouse genes that were analyzed carry the TATA box motif. Except for the last group (Q >10 bits) the percentage of the genes with TATA box motifs decreases with the increase in the Q-value. This is to be expected since genes with high Q may be specific to other tissues and hence are more likely to have a TATA box. Discrepancies between human and mouse promoters were noted for only about 10% of all human-mouse pairs analyzed and may reflect sequence differences and possible annotation discrepancies for the transcription start site. Nevertheless, there is overall excellent agreement for the presence of TATA motifs in human and mouse genes. Thus, our assessment of preferential presence of transcription regulatory motifs in the human pancreas-expressed genes also applies to their mouse orthologs. We conclude that genes expressed with restricted tissue-distribution may be preferentially regulated via TATA-mediated transcription, and that genes with broader expression profiles are more likely to be regulated by non-TATA mediated mechanisms (such as YY1).

Since the presence or absence of a start CpG island and a TATA box appear to be the primary sequence feature that correlate with tissue specificity, we consider them in more detail. We observe that CpG islands and TATA boxes are not mutually exclusive features of promoters and so we consider all possible combinations of these features.

Figure 8 shows the cumulative fraction of each class of promoter as a function of increasing H_g in human (Figure 8a) and mouse (Figure 8b). The data from human and mouse follow similar trends even though the mouse has a lower proportion of CGI+ genes. Overall, CGI+/TATA- genes are the most common, at 50-60% depending on the species. Interestingly, the CGI-/TATA- class is the second most common overall, comprising 20-30% of genes, depending on the species. Genes in this promoter class are roughly equally common across the entire entropy range and are the most common promoters in the mid-specificity range in both species. The classes CGI-/TATA+ and CGI+/TATA+ are the least common (8 to 12% overall). CGI-/TATA+ genes are concentrated in the most specific genes. CGI+/TATA+ are found relatively uniformly across all but the most specific genes. Although the TATA box and CpG islands are strongly predictive of a gene's entropy, Figure 8 also illustrates the limitations of the promoter classes as an explanation for expression patterns. First, although the CGI-/TATA+ and CGI+/TATA- classes are strongly associated with the most and least tissue-specific genes (respectively), instances of genes in each class cover virtually the entire range of tissue specificities. Second, the CGI-/TATA- class is the second most common, illustrating that any degree of tissue specificity can be obtained without these sequence features.

To try to understand the functional correlates of the four promoter classes, we looked for trends in the cellular localization and biological process of the products of genes from each promoter class. We used the DAVID system 4546, which identifies over-represented Gene Ontology (GO) 47 terms in a set of genes. A summary of the results for human and mouse genes are shown in Table 6. In each case the set of genes in each promoter class were compared to all genes on the corresponding Affymetrix chip.

Products of genes in the CGI-/TATA+ class were often (70/198) located extracellularly. Examples of such genes are the insulin-like growth factor family, serum albumin and chymotrypsin. Some extracellular CGI-/TATA+ genes, such as luteinizing hormone beta (LHB) and bone morphogenetic protein 10 (Bmp10) in the mouse, have a high H_g because they are not induced in the tissues or at the developmental stages surveyed, but otherwise fit the pattern of secreted proteins. Gene products that are secreted from the cell must be produced at high level to be effective. Indeed we found the maximum expression level of TATA+ genes is higher than TATA- genes; 454/745 (61%) of TATA+ genes express at least 1,000 AU in one or more tissues, whereas only 1,321/3,773 (35%) of TATA- genes express that highly (p-value = 0; two-sample binomial population). A second group of CGI-/TATA+ that is common, but with a p-value just over the p-value cutoff are the muscle contraction-related genes, actin, troponin and members of the myosin family. Products of these genes are also required in large amounts to create the contractile apparatus but are only produced in a few cell types. The biological processes that are enriched in the CGI-/TATA+ class differ between human and mouse, but nearly all of them are descendants of the GO term 'response to stimulus' (GO:0050896).

The CGI+/TATA- promoters produce proteins that are typically located in the cell, especially in the cytoplasm and mitochondrion. These locations are consistent with many housekeeping functions. The human results for biological process suggests a large number of housekeeping processes, but these were not confirmed in the mouse using all CGI+/TATA- genes. When we consider just the least specific CGI+/TATA- mouse genes (4.45 ≤ H_g ≤ 5.57 bits), we find cellular locations (including the nucleus) and biological processes that match the human results.

No significant concentrations of cellular locations or biological processes were found among the CGI+/TATA+ genes. A manual examination of genes in both human and mouse identifies a number of heat-shock proteins, histones and ribosomal proteins although these are not statistically significant as a result of the multiple testing correction. Many of these genes fit the expected expression pattern in that they are widely expressed and at high levels.

Interestingly, the products of CGI-/TATA- genes are often located in the plasma membrane (244/499 of human genes with a cellular location) and support signaling and response to the environment. Such products, for example, bradykinin receptor B2, prolactin receptor or protocadherin 9, may be expressed in a tissue-specific pattern, but not at the high levels required for secreted proteins. The exact biological process GO terms that are statistically significant vary between mouse and human, but a common core includes defense response (GO:0006952), immune response (GO:0006955) and response to stimulus (GO:0050896). Thus these genes are similar to CGI-/TATA+ genes in that they are involved in response, but are not (typically) required to be expressed at such high levels.

The GNF-GEA data are processed as described 22. Given a set of N tissues we define p_t|g = w_g,t/∑_{1 ≤ t ≤ N}w_g,t where w_t is the expression level of the gene g in tissue t. DoTS, available through the AllGenes 33 site, contains ESTs and mRNAs assembled into transcripts that are then clustered into genes. We did not consider any transcript that contains only one EST as this may represent a spurious sequence and did not consider any gene with fewer than five ESTs because they provide a poor estimate of H_g. To accommodate the great disparity in sampling depth across tissues we normalized EST counts by tissue. To avoid artificially low entropies for genes that contain relatively few ESTs we used pseudocounts to smooth the data. The expression level of a gene in a tissue is computed as w_g,t = (n_g,t + 1)/(N_t + N_g) where n_g,t is the number of ESTs from libraries for a tissue included in a gene, N_t is the total number of ESTs from a tissue assembled into genes, and N_g is the number of genes. We used different sets of tissues depending on the task. H_g and Q measures in Figure 1 used the full GNF-GEA mouse set with a few modifications; adipose tissue and brown fat were merged, epidermis and snout epidermis were merged, digits and tongue were not considered as they are both a combination of skeletal muscle and epidermis. The expression level for a set of merged tissues is the median of the individual tissue replicate medians. For comparison of microarray and EST data we used a set of 27 tissues that were common to both datasets and