Utilizing two different methods, FASTA and TRIBE-MCL, we identified 6,536 and 7,027 non-redundant human duplicate gene pairs, respectively (see Materials and methods for details). These pairs represented 3,313 and 3,555 gene families, respectively. After filtering out duplicate gene pairs with synonymous rate (K_S) >2 and/or lacking a start codon, we obtained 2,790 and 2,750 duplicate gene pairs using the former and the latter methods, respectively. A total of 1,600 duplicate gene pairs overlapped between these two data sets (Additional data file 2). All subsequent analyses were carried out for duplicate genes identified with each of the two methods. Because the results were similar, we present the results only for duplicate genes identified with the FASTA method (2,790 gene pairs in group A), as this method is stricter for clustering proteins into families compared with the TRIBE-MCL method 2930.

From human U133A and GNF1H oligonucleotide arrays 28, we defined 14,505 genes that mapped to probes with a one-to-one correspondence (see Materials and methods), thus minimizing cross-hybridization. Among these genes, we were able to detect 2,924 non-redundant duplicate gene pairs belonging to 1,792 multiple gene families. After filtering out duplicate gene pairs with K_S >2 and/or lacking a start codon, we obtained 1,015 duplicate gene pairs (group B, representing a subset of group A). In the remainder of the manuscript, we consider duplicate genes of group B when gene expression is investigated and duplicate genes of group A otherwise.

Initially, we analyzed the divergence in the position of TSSs between copies in each duplicate gene pair. Using tag clusters, which were built by grouping overlapping tags (namely, 5'-end-sequences) with the same strand, from large-scale tag clustering of the cap analysis of gene expression (CAGE) 20 and the paired-end ditags (PETs) 31, putative TSSs of each gene were identified (see Materials and methods). From 2,790 duplicate gene pairs in group A, we excluded duplicate gene pairs that were duplicated by retrotransposition or for which at least one copy lacked a TSS(s) identified by either CAGE or PETs. As a result, 1,124 duplicate gene pairs were retained. To evaluate sharing of TSSs between duplicate genes, we compared the sequences of genomic regions surrounding putative TSSs (as identified by CAGE or PETs) between the two copies for each of these 1,124 duplicate gene pairs. We considered 110 bp (-20 bp to +90 bp) surrounding each TSS (later called the 'TSS region'), because there was a clear peak in the average sequence similarity between TSSs of duplicate genes in this region (Additional data file 3) and because several studies indicated that a region of this size surrounding TSSs was well conserved between human and mouse orthologs 3233. Sequence similarity between all possible combinations of TSS regions from each duplicate gene pair was considered. If at least one pair of TSS regions had an identity greater than 60%, it was defined as a TSS(s) shared between the two duplicate copies. As a result, 13.6% (153 out of 1,124) of duplicate gene pairs had shared TSSs.

We observed that the relative frequency of gene pairs with shared TSSs decreases with increasing K_S, a proxy of time since duplication (Figure 1). The L-shaped distribution observed in Figure 1 implies a rapid turnover of TSSs after gene duplication. Already at K_S = 0.1, corresponding to only about 33 million years ago since duplication 34, a mere 64% of duplicate genes share TSSs. Considering an instantaneous K_S rate according to 35 did not alter our results (Additional data file 4).

Interestingly, the turnover of TSSs between human duplicate genes was much more rapid than between human-mouse orthologs. Indeed, for 1,610 human-mouse orthologs considered (see Materials and methods), the mean K_S was 0.61 (with a 95% confidence interval of 0.60-0.63), while the proportion of orthologs with shared TSSs was 0.71, several fold higher than the proportion of human duplicate genes with similar K_S (Figure 1).

To estimate the relationship between TSS usage patterns (for example, shared TSSs versus non-shared TSSs) and gene duplication mechanisms, the duplicate genes were divided into three classes: retrotransposed duplicate genes, tandem, and nontandem duplications (see Materials and methods for details). The relative frequencies of gene pairs with shared TSSs in each class were calculated (thus, we analyzed 1,124 non-retransposed genes as above plus 220 retrotransposed genes). Duplicate gene copies in which one of the pair has one exon and the duplicate copy has multiple exons were called retrotransposed duplicate gene copies. We found that among paralogs with shared TSSs, the majority of pairs represented tandem duplicates (Additional data file 1).

Interestingly, about 30% (67 out of 220) of retrotransposed duplicate gene pairs retained the same TSSs (Additional data file 1). To evaluate whether the retrotransposed gene pairs with shared TSSs tend to undergo stronger purifying selection than those without shared TSSs, the median nonsynonymous-to-synonymous rate ratios (K_A/K_S) were compared between these two groups of genes; however, no significant difference was detected (0.475 versus 0.499; P > 0.1, Mann-Whitney U test).

Next, to test whether the turnover of TSSs may contribute to the expression divergence in duplicate genes, the Pearson correlation coefficient of expression values (R_expression; calculated for 61 non-redundant tissues) between the two copies in each pair was computed and compared among group B duplicate gene pairs with shared TSSs versus those without shared TSSs (a total of 581 group B pairs with available TSS data were included in the analysis). Duplicate genes with shared TSSs had significantly higher R_expression values than those without shared TSSs (0.437 versus 0.080; P < 0.01, Mann-Whitney U test). It is conceivable that the significant difference in R_expression values is due to different synonymous rates in genes with shared TSSs versus those without shared TSSs. Indeed, we observed that all duplicate genes (belonging to group B) with shared TSSs had K_S <0.4, while more than 97% of gene pairs without shared TSSs had K_S ≥ 0.4. However, if only genes with K_S <0.4 were considered, the gene pairs with shared TSSs still had higher (but not significantly so) R_expression values than those without shared TSSs (0.437 versus 0.140; P > 0.05, Mann-Whitney U test).

The 60% identity threshold among the TSS regions that was tentatively inferred from substitution rates between human and mouse ortholog core promoters 36 may be inadequate for estimating the sharing of TSSs among human paralogous genes. Thus, we reclassified the sharing of TSSs between copies of duplicate genes using several identity thresholds (40%, 50%, 70%, and 80%). Although the numbers of duplicate genes with shared TSSs in each bin varied with the threshold, the frequency of gene pairs with shared TSSs decreased over divergent time independent of the threshold used (Additional data file 5), consistent with the pattern observed with the 60% identity threshold (Figure 1). Moreover, regardless of the identity threshold, the R_expression values were significantly higher in duplicate genes with shared TSSs versus those without shared TSSs (data not shown).

By reconstructing the full-length coding sequences via concatenating exons from multiple splicing variants for each gene separately, each pair of duplicate genes was classified into one of two structural categories: completely similar and incompletely similar. If the proportion of aligned sequences was greater than 0.9, duplicate gene pairs were categorized as completely similar and as incompletely similar otherwise. For some analyses, incompletely similar duplicate gene copies were classified in one of the three non-overlapping groups: 5' similar, 3' similar, and neither 5' nor 3' similar. If alignments between the two copies started at the start codons of both copies, then such duplicates were classified as 5' similar. Alternatively, if the alignments ended at the stop codons of both copies, we classified the duplicate genes as 3' similar. The remaining duplicate gene pairs were labeled as neither 5' nor 3' similar.

After excluding genes that lacked start/stop codons or consensus splice sites, 2,591 duplicate gene pairs were retained (from 2,790 pairs of group A; for group B, 889 duplicate gene pairs were retained). We found that 55% (1,429 out of 2,591) of duplicate gene pairs had incompletely similar structures. As expected from the divergence of the coding sequence over time, the proportion of duplicate gene pairs with completely similar structures decreased gradually with divergence between the two duplicate copies, approximated by K_S (Figure 2). Considering an instantaneous K_S rate according to 35 did not alter our results (Additional data file 6). Interestingly, even at the smallest duplicate gene divergence (K_S <0.1), the proportion of genes with completely similar structures was only 80% (Figure 2). Although this finding might be affected by misannotations, our results suggest that some duplicate genes might have acquired structural differences during duplication.

To analyze whether the incompletely similar structures of duplicate genes can lead to expression divergence, we compared the relationship between R_expression and K_S for duplicate genes with completely versus incompletely similar structures. Before addressing this issue, retrotransposed duplicate genes (a total of 108 out of 889 genes retained in group B) were excluded because, as retrotransposition does not include a promoter, it can lead to expression divergence regardless of structural heterogeneity in coding sequence between duplicates. We found that: the correlation coefficient between R_expression and K_S for duplicate gene pairs with completely similar structures was significantly lower than that for pairs with incompletely similar structures (R = -0.315 versus R = -0.001; Fisher's z test, z = -4.028, P < 0.001; Kolmogorov-Smirnov test for normality, P < 0.010; Figure 3 and Table 1); and duplicate genes with completely similar structures had significantly higher y-intercepts of regression lines than duplicate genes with incompletely similar structures (0.407 versus 0.134; z = 2.672, P < 0.01). These observations suggest that, immediately after duplication, the expression pattern is more similar for duplicate gene pairs retaining the same versus acquiring different coding sequence structures, and that divergence of gene expression is more dependent on evolutionary time for duplicate gene pairs with completely versus incompletely similar structures. To estimate the importance of sharing of 5' regions of coding sequences between duplicate gene copies, which can be an indirect indicator of common transcription regulation mechanisms, we separately considered duplicate gene pairs completely similar at the 5' end only (a total of 24 gene pairs from group B that were otherwise genes with incompletely similar structures) and calculated the correlation coefficient between their R_expression and K_S. The correlation was negative, but not significant (Table 1). When duplicate gene pairs having completely similar and 5' similar structures were considered together, the correlation coefficient between R_expression and K_S was somewhat lower than that for duplicate gene pairs with completely similar structures (Table 1), although the difference was not significant (z = -0.093, P > 0.1). We observed that there was no correlation between R_expression and K_S for duplicate genes with 3' similar structure and with neither 5' nor 3' similar structure (Table 1). These results suggest that maintenance of the entire coding region (and not just of its 5' or 3' portion) is important for determining gene expression profile after duplication.

To estimate differences in selective pressure among duplicate genes in different structural categories, their K_A/K_S ratios were compared (Table 1). We observed that K_A/K_S was significantly lower for duplicate genes with completely similar structures than for those with incompletely similar structures (P < 0.001, Mann-Whitney U test; Table 1), suggesting that the former genes are subject to stronger purifying selection than the latter genes.

Next, we evaluated the relative contribution of cis-regulatory divergence to differences in expression between copies of duplicate genes in each pair. The 2-kb (from -1.5 kb to +0.5 kb) genomic regions surrounding TSSs were used as putative cis-regulatory sequences and their divergence was estimated with REALIGNER 37. For genes with multiple TSSs, a TSS supported by the highest number of CAGE/PET tags was selected. This analysis was limited to group B duplicate genes with completely similar structures (a total of 158 duplicate gene pairs). We found a significant positive correlation (R = 0.242, P < 0.01) between the proportion of aligned sequences in the cis-regulatory region (P_cis) and R_expression. This implies that the divergence of cis-regulatory regions leads to expression divergence in duplicate genes. After duplicate genes created by retrotransposition (a total of 23 gene pairs) were excluded, the correlation coefficient was even higher (R = 0.252, P < 0.01). Through comparison between K_S (which may serve as a neutral proxy, although see 38) on the one hand and the proportion (corrected for multiple hits using HKY85 model) of aligned sequences in the cis-regulatory region on the other hand in each non-retrotransposed duplicate gene pair, we estimated whether the cis-regulatory regions evolved neutrally. We found that for 107 out of 135 duplicate gene pairs compared, K_S was significantly higher (P < 0.001, Wilcoxon signed-rank test) than the proportion of aligned sequences in the cis-regulatory region, suggesting that purifying selection acts at cis-regulatory regions.

To investigate whether copy-specific TEs influence divergence in duplicate gene expression, we identified such TEs (TEs that integrated in the cis-regulatory region of only one duplicate gene copy of a pair after duplication) in the same 2-kb regions surrounding TSSs of the above 158 duplicate genes pairs (excluding 23 retrotransposed duplicate pairs; see Materials and methods). However, no significant correlation was found between the proportion of copy-specific TEs and either P_cis for duplicate genes or R_expression (data not shown). This suggests that the effect of copy-specific TEs on divergence in duplicate gene expression may be at best minor, although this issue requires additional studies.

Because several factors studied above might be interrelated, we conducted multiple regression analysis to estimate the relative contribution of each factor to explaining the total variability in R_expression. A total of four continuous predictors (K_A, K_S, the K_A/K_S ratio, and divergence of cis-regulatory sequences (labeled 'Cis') and three categorical predictors (shared versus not shared TSSs (labeled 'TSS'); completely versus incompletely similar gene structure (labeled 'Structure'); and tandem versus non-tandem gene organization (labeled 'Tandem')) as well as all possible pairwise interaction terms were used to build a regression model. After pruning nonsignificant terms, the final multiple regression model explained approximately 10% of the variation in R_expression and consisted of eight predictors (Table 2). Five of these predictors remained significant after applying Bonferroni correction for multiple tests (Table 2). These predictors included: Tandem, TSS, and interaction terms between Structure and Tandem, between TSS and Tandem, and between K_A/K_S ratio and Cis (Table 2). Our computation of the relative contribution of the variability explained (RCVE) for significant predictors (see Materials and methods for details) indicated that each of them makes a sizeable input into the model.

To cluster genes into families, we downloaded 48,218 protein sequences of consensus coding sequences, known and novel genes from Ensembl (release 38 of NCBI build 36) and independently used the FASTA 63 and TRIBE-MCL 64 methods to define duplicate gene families. Briefly, for the FASTA method, each protein sequence was used as a query to search against all other protein sequences using FASTA 65 with E < 10. Two protein sequences formed a link if: the aligned region was >80% of the longer protein; and the identity between two proteins was ≥ 30% for alignments longer than 150 amino acids or ≥ (0.01n + 4.8L^{-0.32 [1+exp(-L/1000)]}) otherwise, where L is the alignable length between two proteins and n = 6. The formula above was derived from empirical data, which suggested that a higher sequence identity was required for shorter proteins 66. These gene pairs were grouped into gene families according to the single linkage clustering algorithm. For gene families derived by TRIBE-MCL, we downloaded the gene annotations through BioMart in the Ensembl database, and considered gene families with at least two members.

To identify independent pairs of duplicate genes within each gene family, we sorted gene pairs in ascending order of K_S and selected the pair with the lowest K_S. After excluding genes that had been picked, we chose the next gene pair with the lowest K_S. These steps were repeated for each gene family. All genes encoding proteins were realigned using CLUSTALW 67, and the yn00 module 68 of PAML 69 was used to calculate K_S. We counted duplicate gene pairs in intervals of size K_S = 0.01 to derive the instantaneous rate of K_S according to 35.

Duplicate gene copies in which one of the pair has one exon and the duplicate copy has multiple exons were called retrotransposed duplicate gene copies. In addition, duplicate gene pairs were classified as tandem duplicates if there were no genes separating them.

Expression data for 61 non-redundant and nonpathogenic human tissues in U133A and GNF1H Affymetrix arrays were obtained from 28. To validate mapping between probe sets and genes, we aligned the transcripts of consensus coding sequences, known genes, and novel genes downloaded from Ensembl (release 38 of NCBI build 36) with the exemplar and consensus sequences for each array using BLAST 70 with E < 10^-20. According to the criteria described in 7172, the acceptable alignments were selected if: the identity was 100% and the length was greater than 49 bp; or the identity was higher than 94% and the length was at least either 99 bp or 90% of the length of the query. We considered three scenarios for mapping relationships: a single probe set hitting one gene (9,508 probe sets); multiple probe sets hitting one gene (13,186 probe sets and 4,997 genes); and a single probe set hitting multiple genes (4,493 probe sets and 6,764 genes). All genes following the first two scenarios were utilized in the present study. For each gene following the second scenario, the probe set with the highest expression value (defined by average difference) was selected. All genes following the third scenario were removed from the analysis due to potential cross-hybridization. Following 28, genes with average difference >200 in a particular tissue were considered to be expressed in this tissue.

The putative TSSs were identified using the method described in the ENCODE pilot project 73. Briefly, we utilized tag clusters from two sets of 5'-end-tag-capture technologies: CAGE 20 and PETs 31. If two tag clusters were located on the same strand and within 60 bp (which was derived from analyzing the distribution of distances between tag clusters in 73) of each other, they were considered as one tag cluster. To map tag clusters to genes, the following two criteria were considered. First, the strand of a tag cluster was required to be identical to the strand of a gene. Second, a tag cluster was required to be located in the 5' upstream region from the most upstream start codon of a gene. Because we constructed artificial coding regions of genes by including all their exons, our analysis is not affected by alternative start codons. To confirm the reliability of the tag data, RefSeq 74, H-Invitational 75 and human ESTs 76 RNA data from the UCSC Genome Browser 77 were utilized. We excluded tag clusters with a single tag as well as those whose coordinates did not overlap with the genomic coordinates of the 5' end of cDNAs or ESTs. To define a representative tag site (to be used as a putative TSS) for each tag cluster, we selected the tag site that was supported by the highest number of 5' start sites. Otherwise, if several sites in a tag cluster had the same number of 5' start sites, the central coordinate of this tag cluster was defined as the representative tag site.

To evaluate conservation of TSSs between human-mouse orthologous genes, we obtained two distinct classes of orthologous genes from 23. Briefly, 'conserved promoter regions' means that upstream sequences of TSSs between human and mouse orthologous genes were aligned; otherwise, 'non-conserved promoter regions' means there were no significant alignments. We excluded orthologous genes that were classified into both classes because alternatively spliced variants of each gene had different conservation patterns of promoter regions. As a result, 1,610 orthologous gene pairs that were classified into just one class in a mutually exclusive manner were retained. We downloaded human and mouse protein sequences from Ensembl (release 38 of NCBI build 36). All genes were aligned using CLUSTALW 67, and the yn00 module 68 of PAML 69 was used to calculate K_S between orthologous genes.

Structural categorization of duplicate genes was performed using reconstructed full-length coding sequences. We downloaded annotated human genome data from Ensembl (release 38 of NCBI build 36). Alternatively spliced variants lacking start or stop codons or lacking canonical exon boundaries (5'-GT...AG-3', 5'-GC...AG-3', or 5'-AT...AC-3') were excluded. For each gene with several alternatively spliced variants, all exons were aligned against each other, and, if some exons overlapped, they were merged in a single exon. Next, exons were sorted by their genomic coordinates and were reassembled to form reconstructed full-length coding sequences.

The reconstructed full-length coding sequences were aligned using AVID 78 with default parameters. Each pair of duplicate genes was classified into one of the four structural categories: completely similar, 5' similar, 3' similar, and neither 5' nor 3' similar. If the proportion of aligned sequences was greater than 0.9, duplicate gene pairs were categorized as completely similar. The other duplicate gene pairs were exclusively classified in just one category of 5' similar, 3' similar, or neither 5' nor 3' similar. If alignments between the two copies started at the start codons of both copies, then such duplicates were classified as 5' similar. Alternatively, if the alignments ended at the stop codons of both copies, we classified the duplicate genes into 3' similar. Finally, the remaining duplicate gene pairs were labeled as neither 5' nor 3' similar.

To detect homologous sequences in cis-regulatory regions, we used a modified version of REALIGNER 37. Using BL2SEQ (part of the Blast suite 70) with mismatch penalty equal to -2 and word size equal to 7, we constructed alignments of 2-kb (-1.5 kb to +0.5 kb) genomic regions surrounding putative TSSs between copies in each duplicate gene pair. We selected alignments satisfying three criteria: hit length >7 bp; identity >70%; and identical hit strand. If two local alignments overlapped, an alignment with the higher bit score was retained. If the bit scores of the two overlapping alignments were identical, a longer alignment or the one closest to TSS was retained. If the two local alignments were not syntenic (the order of blocks in each alignment was inconsistent), an alignment with the lower bit score was removed. Finally, all local alignments ordered by their genomic coordinates were used as a conserved cis-regulatory region for a duplicate gene pair.

TEs within cis-regulatory regions were classified into two sets: with the insertion occurring in the ancestral sequence before duplication of a genomic region; with the insertion in only one duplicate copy after the duplication event. We used the Repeatmasker 79 tables at the UCSC Genome Browser 77 to map the coordinates of TEs into cis-regulatory regions.

Linear multiple regression analysis was performed in the R statistical package. The original model included all seven predictors and their interaction terms, but was pruned to include only significant predictors (and significant interaction terms). RCVE 8081 was utilized to assess the contribution of each predictor to explaining the total variability:

R C V E = R f u l l 2 − R r e d u c e d 2 R f u l l 2 MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciGacaGaaeqabaqabeGadaaakeaacaWGsbGaam4qaiaadAfacaWGfbGaeyypa0tcfa4aaSaaaeaacaWGsbWaa0baaeaacaWGMbGaamyDaiaadYgacaWGSbaabaGaaGOmaaaacqGHsislcaWGsbWaa0baaeaacaWGYbGaamyzaiaadsgacaWG1bGaam4yaiaadwgacaWGKbaabaGaaGOmaaaaaeaacaWGsbWaa0baaeaacaWGMbGaamyDaiaadYgacaWGSbaabaGaaGOmaaaaaaaaaa@4946@

where Rfull2 MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8viVeY=Nipec8Eeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaamOuamaaDaaaleaacaWGMbGaamyDaiaadYgacaWGSbaabaGaaGOmaaaaaaa@352A@ and Rreduced2 MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8viVeY=Nipec8Eeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaamOuamaaDaaaleaacaWGYbGaamyzaiaadsgacaWG1bGaam4yaiaadwgacaWGKbaabaGaaGOmaaaaaaa@37E2@ are the R² for the full model and the model except for the predictor of interest, respectively. In addition, variance inflation factors 82 were calculated for each predictor to diagnose multicollinearity. All predictors and their interaction terms included in the final model had variance inflation factors below 2 (data not shown), suggesting that multicollinearity was not adversely affecting the model.