Differences in human and chimpanzee gene expression patterns define an evolving network of transcription factors in brain

Abstract

Humans differ from other primates by marked differences in cognitive abilities and a significantly larger brain. These differences correlate with metabolic changes, as evidenced by the relative up-regulation of energy-related genes and metabolites in human brain. While the mechanisms underlying these evolutionary changes have not been elucidated, altered activities of key transcription factors (TFs) could play a pivotal role. To assess this possibility, we analyzed microarray data from five tissues from humans and chimpanzees. We identified 90 TF genes with significantly different expression levels in human and chimpanzee brain among which the rapidly evolving KRAB-zinc finger genes are markedly over-represented. The differentially expressed TFs cluster within a robust regulatory network consisting of two distinct but interlinked modules, one strongly associated with energy metabolism functions, and the other with transcription, vesicular transport, and ubiquitination. Our results suggest that concerted changes in a relatively small number of interacting TFs may coordinate major gene expression differences in human and chimpanzee brain.

Humans differ from chimpanzees in a number of important anatomical and physiological respects, most strikingly in our enhanced cognitive abilities and a substantial increase in the relative size of the human brain (1). Although the human brain is relatively energy-efficient per cell compared with brains of other species, this increased capacity imposes a significant metabolic and oxidative burden (2, 3). Several studies have noted the up-regulation of genes and metabolites involved in oxidative metabolism and mitochondrial function in human brains compared with chimpanzee brains (2, 4, 5). These data, together with evidence of positive selection acting on the promoters of genes involved in energy metabolism during human evolution, indicate that increased energy production has been essential to the evolution of the human brain (6). The relative up-regulation of human genes in other functional categories, including neuroprotection and synaptic transport, has also been documented (7). However, the molecular mechanisms underlying these well-documented species differences have not been elucidated.

Although some differences in human–chimpanzee gene expression may be due to cis-regulatory element divergence, transcription factors (TFs) represent another potential source of expression variability. Whereas most cis-element mutations would be expected to have limited, localized effects, alterations in TF sequence and/or expression could alter the expression of hundreds of target genes in a coordinated fashion (8, 9). Because of these predicted consequences, it is often assumed that TFs are evolutionarily stable, and indeed, TFs as a class are structurally well conserved (8). However, two recent studies have identified TF genes as enriched among genes with expression patterns that are under directional selection in humans (4, 10). These studies raise the intriguing hypothesis that differences in the expression and networking of specific TFs could be driving major changes between primate species. However, previous analyses of human and chimpanzee TF gene expression did not include a comparison of gene expression in brain.

Other large microarray studies have included human and chimpanzee brain comparisons (5, 7, 11), but the assessment of TF gene expression is complicated by several issues. In particular, many TF loci are members of extended gene families, yet most microarray platforms are not designed to uniquely detect the specific family members. The problem is particularly acute for the largest family of TFs in mammals, the KRAB zinc finger (KRAB-ZNF) genes. About one-third of these genes are primate-specific, including many recent duplicates (12). In striking contrast to other TFs, KRAB-ZNFs have on average accumulated more amino acid differences between humans and chimpanzees than other genes, indicating that they may have contributed disproportionately to the phenotypic differences between these species (13, 14).

To enable an accurate comparison of TF gene expression and network structure in human and chimpanzee brain, we devised a strategy to reliably distinguish expression levels of individual gene family members. Our analysis of an established dataset (11) uncovered 90 TF genes that are differentially expressed and revealed that they are organized in a coexpression network comprised of two modules with distinct functions. Both modules are enriched for primate-specific KRAB-ZNF genes, which despite their recent advent are robustly embedded in the chimpanzee and human brain networks. Our results implicate a network of TFs with differential expression in human and chimpanzee brain involved in regulation of energy metabolism, vesicle transport, and related functions required to build and maintain the larger and more complex human brain.

Results

TF Genes with Differential Expression in Human Compared to Chimpanzee Brain.

We reexamined the expression of TF genes (15) in a published Affymetrix microarray data set that contrasts five human and chimpanzee tissues: heart, kidney, liver, testis, and brain—specifically prefrontal cortex (PFC). Although all cortex regions display very similar expression differences between humans and chimpanzees (16), the PFC is a good study object because of its marked differences in structure and function between the two species. For convenience, we will refer to these samples simply as “brain” in the following discussion.

Expression data were analyzed from all five tissues in each of six individual humans and in five individual chimpanzees (11). To improve the reliability of the comparisons, we masked all probes that do not match both genomes perfectly (10). To detect gene family members uniquely, we also removed probes with more than one exact match in either genome. This approach yielded modified probe sets for many genes, including 1,463 of 1,715 probe sets detecting TF loci, and allowed unique expression measurements for 1,063 TFs, including 331 KRAB-ZNF genes (http://znf.igb.uiuc.edu/human). Especially for KRAB-ZNFs this is a clear improvement over an earlier study that used smaller, unmasked Affymetrix expression arrays and reliably detected the expression of only 211 KRAB-ZNF genes, almost none of which were primate-specific (12).

To identify the genes with the largest differences in robust multiarray average (RMA) expression level between species in each tissue, we initially used a two-sample t test with a P value cut-off of P < 0.01. For this cut-off we calculated a false discovery rate of <2% for all tissues (Table S1). Implementation of the unique-sequence mask clearly affected the discernment of differentially expressed genes (Table S2). Twenty TF genes that would otherwise have been selected as differentially expressed in brain did not display significant species differences after masking; likewise, 16 TFs were added to the differentially expressed gene set.

As previously noted (11), after our masking, most expression differences were seen in testis, whereas the four somatic tissues showed approximately the same number of differentially expressed genes (Fig. 1). We were particularly interested in expression differences in brain, which was not investigated in previous studies of human and chimpanzee TFs (4). In contrast to these studies, we did not find TFs as a whole to be over-represented in the differentially expressed genes from any examined tissue (Table S1). This difference could have several explanations, including differences in array platform and gene representation. However, when we analyzed KRAB-ZNF genes separately we observed a significant [permutation test (PT), P = 0.029] enrichment of genes of this family among differentially expressed genes in brain (Table S1). When we excluded KRAB-ZNFs from our analysis, TFs as a group were significantly under-represented among the genes with changed brain expression in human compared to chimpanzee (Table S1). We confirmed this finding through a series of tests, including calculating the expected number of differences among genes with similar expression intensities like KRAB-ZNFs, analysis on the probe set (rather than gene) level, and comparing gene expression variation within species to the interspecies difference (SI Text). On the other hand, KRAB-ZNFs are significantly depleted (PT, P = 0.999) from the set of differentially expressed genes in testis. The only other large TF families, encoding Homeobox and the basic helix–loop–helix (bHLH) proteins, do not show such enrichment: Whereas 23.2% of brain-expressed KRAB-ZNFs are differentially expressed between species, only 8.2% of the Homeobox and 5.5% of the bHLH genes have changed in brain expression. Our analysis therefore revealed a clear contrast between KRAB-ZNFs and other TFs as well as between brain and other tissues (Fig. 1).

Fig. 1.

Percentage of differentially expressed genes among human and chimpanzee tissues. The proportion of all genes (white), all transcription factors (light gray), all KRAB-ZNF (KZNF) genes (gray), conserved KRAB-ZNF genes (dark gray), or primate-specific KRAB-ZNF genes (black) that are differentially expressed between species (t test, P < 0.01) is shown separately for each tissue. Asterisks mark values that represent significant enrichment. Numbers of genes per category are between 7 and 6,720.

To focus on the differentially expressed TFs most likely affecting brain functions, we applied three additional filters. In the remainder of this paper, we will only refer to genes as “changed” between humans and chimpanzees if they (i) are significantly different in expression after correcting the P value obtained from the two-sample t test within each tissue for multiple testing (P < 0.05; Benjamini–Hochberg correction), (ii) have a difference of at least 1.2-fold, and (iii) have a difference of 20 units of expression values. The latter criterion ensures that the gene is expressed at a modest level in at least one species. We found 90 TFs, including 33 KRAB-ZNFs, that met these more stringent requirements for differential expression between human and chimpanzee brains (Table S2). Despite the fact that most TFs (79 of 90) are expressed in all five tissues, about one-quarter of them (18 of 79) have changed specifically in brain. The proportion of KRAB-ZNFs that have changed only in brain is higher (9 of 29). Interestingly, recent primate-specific KRAB-ZNFs [defined as genes within human segmental duplications of at least 90% overall sequence identity (Segmental Duplication Database, http://humanparalogy.gs.washington.edu, accessed March 30, 2007) (17) that recognize no ortholog in the mouse or dog genomes (12)], are enriched among brain-changed genes (PT, P = 0.027) (Fig. 1). Importantly, primate-specific KRAB-ZNFs are not overrepresented among the genes displaying human-chimpanzee expression differences in any other tissue (Table S3). The other TFs that display changed expression levels in human and chimpanzee brain correspond to a wide variety of different protein families (Table S2). While these include a few well-studied TFs most have no known regulatory functions or established roles in the brain.

Brain-Changed TF Genes Form a Bimodular Coexpression Network.

Structure of the human brain TF network.

The diversity of TF types raised the question of whether these proteins are acting independently or coordinately in the brain. To address this question and to gain clues to possible TF gene functions, we identified potentially concordant changes between TFs and other genes using expression correlation-based methods.

As a first step, we computed expression correlations for the brain-changed TFs (Fig. 2). To provide maximum information for these correlations, we used all 30 human samples (six individuals per five tissues). This step permitted us to compute gene expression correlation values with higher statistical power, but it restricts our analysis to genes that are expressed in all five tissues, including 79 of the 90 brain-changed TF genes. KRAB-ZNFs are thought to function primarily as repressors, but many TFs can act as activators or repressors depending on context (e.g., refs. 18 and 19). To allow inclusion of genes that are negatively regulated by differentially expressed TFs, we identified both positively and negatively correlated genes for each TF from the human tissue dataset.

Fig. 2.

Strategy to obtain TF-associated gene sets. Summary of the strategy used to identify positively and negatively associated gene sets for the network analysis. Details are described in the text.

To further enrich for genes that are potentially directly affected by changes in TF expression, we next applied an interspecies filter (Fig. 2). Specifically, from each set of positively correlated genes, we retained only those genes with a human–chimpanzee brain expression difference in the same direction as the TF. For example, for each TF that is relatively up-regulated in human brain compared with chimpanzee brain, we retained only those genes that are also more highly expressed in human brain in the interspecies comparison. Similarly, we filtered the negatively correlated gene sets by retaining only those genes that changed in the opposite direction as the TF in brain expression. Thus, for example, for each human up-regulated TF, we retained only genes that displayed lower expression in human compared to chimpanzee brain. In the following discussion we will refer to the set of genes retained by this method as the TF-associated gene sets. In many of the TF-associated gene sets we found significant enrichment of Gene Ontology (GO) groups indicating related functions, and in gene sets associated with two TFs with known DNA binding sites we found significant enrichment of those motifs in predicted promoter regions (SI Text). These data support the idea that the associated gene lists are indeed functionally related to the TFs.

We noted considerable overlap between the gene sets associated with certain TFs and the GO categories enriched among those gene sets (Table S4), indicating that groups of TFs might be acting cooperatively. To test this hypothesis, we calculated the weighted topological overlap (wTO) matrix [ω_ij] (20–22) for the brain-changed human TFs. The wTO method allows us to visualize the interactions between TFs in a large network by not just displaying possible direct TF correlations, but also by taking the overlap between TF-associated gene sets into account. In other words, instead of drawing all links between all correlated genes in the network, the overall commonality in the associated gene sets for two TFs is represented. Detrimental effects resulting from incorrectly assigned associated genes (false positives) are therefore strongly suppressed, and we expect the resulting network to provide more robust information about the connections between TFs than a simple gene correlation network.

The resulting human brain wTO network (Fig. 3 and Fig. S1A) shows a striking biclustering of human up-regulated and human down-regulated TFs into separate modules with high connectivity within each module but few links between them. We hereafter refer to these modules as Module 1 (dominated by human brain up-regulated TF genes) and Module 2 (dominated by down-regulated TFs). Testing multiple wTO cutoffs in the interval [0.2, 0.5] we demonstrated that this structure is robust with respect to cutoff choice. Furthermore, we implemented several tests to ensure that the observed network structure is significantly different from random expectation and compared our wTO calculation with an alternative approach (22) (SI Text and Fig. S1D).

Fig. 3.

Weighted topological overlap network between brain-changed transcription factors. Using as an input the correlations of the full set of TF-associated genes in human samples, the wTO network was calculated between the 79 ubiquitously expressed brain-changed transcription factors (∣ω_ij∣ > 0.3). TFs up-regulated in human brain compared with chimpanzee brain are shown in red, and down-regulated TFs are shown in green. Positive and negative links between TFs are shown in red and green, respectively. Numbers label the four TFs with the highest BC scores.

The topology of the two modules of the observed human wTO network (Fig. 3) is noticeable different; in particular, the average density (23) of Module 1, d = 0.65, is significantly larger than that of Module 2, d = 0.43 (Mann–Whitney test, P = 2.11 × 10⁻⁸). The two modules are interconnected by a relatively small number of links: Only 54 intermodule links are connecting significantly fewer TFs than expected by chance (28 up-regulated TFs; PT, P = 0.047; 17 down-regulated TFs; PT, P < 10⁻⁶). To determine the TFs important for connecting the two modules in an unsupervised way, we calculated the betweenness-centrality (BC) score for all TFs (24). Simply stated, the BC score of a gene corresponds to the number of shortest paths between all pairs of genes that pass through that gene and is hence a measure for the centrality of a gene in a network. Genes with high BC scores are sitting on a large number of shortest paths, thus acting as bridge nodes in a modular network. We find that ZNF542, NFYA, TAF11, and ZNF717 are the TFs with the largest BC scores (numbered 1–4 in Fig. 3, respectively; see Table S5 for a complete list).

Comparing network structure between species.

To investigate species-specific differences in network architecture, we used the same approach to calculate a wTO network representing the connections between the changed TFs in all five chimpanzee tissues (a total of 25 samples). The TFs in the chimpanzee brain network are the same as in the human brain network; however, because expression patterns for genes have changed between species, the TF-associated gene sets might be different. Because there were six humans but only five chimpanzees in the original study, we produced 50 human wTO networks based on random sampling of five out of six human individuals per tissue to facilitate a direct comparison between the networks for the two species. These samples were then used to generate a consensus human network consisting of all links found in most of the resulting networks. Note that the structure of the human network is robust with regard to this resampling (Fig. S1C). We find that the overall topology of the chimpanzee network is very similar to the human consensus network (Fig. S1B). However, human TFs are more interconnected with each other (PT, P < 0.02) and have mainly gained links between Module 1 TFs (PT, P < 0.02), whereas chimpanzee TFs have significantly more links between Module 2 TFs (PT, P < 0.02). The links connecting the two modules are enriched for potential species-specific links (χ² test, P = 2.2 × 10⁻¹⁶).

Next we identified TFs that are differently integrated into the network in the two species. Therefore, we identified all species-specific links between pairs of TFs (Table S6). We defined human-specific links as those that are present in all 50 resampled networks but not in the chimpanzee network. Likewise, chimpanzee-specific links are those present in the chimpanzee network but in none of the resampled ones. Note, that these definitions probably lead to an underestimation of species-specific links. Human TFs have on average four links that are not present in the chimpanzee network, whereas chimpanzee TFs have on average two potential chimpanzee-specific links (Fig. 4). TCEAL1, TBX19, ZFHX1B, and ZNF295 (human) and TFDP2 (chimpanzee) stand out as having gained a significant (>2 SD) number of species-specific links. These species-specific links can shift the position of a TF in the network. For instance TFDP2 has 14 chimpanzee-specific links, all connecting this gene more tightly to Module 2, whereas it is linked more closely to Module 1 in humans (Fig. S1, compare A with B).

Fig. 4.

Difference in number of species-specific links among TFs in the human and chimpanzee wTO network. For each TF, the difference in the number of human- vs. chimpanzee-specific links is plotted. For definition of species-specific links, see the text. Genes with the largest numbers of human-specific links are positioned at top. The mean difference is 1.3 and 2 SDs are 10.3.

Functions associated with the TF modules.

To identify potential functions for the two modules, we determined which genes were most frequently associated with TFs in each module and tested whether these genes were enriched for certain GO categories using the program FUNC (25). We found that each module is characterized by distinct enrichment of GO categories (Fig. 5), and enriched GO categories were almost identical for the corresponding chimpanzee modules (Fig. S2). In particular, Module 1 TF-associated genes are highly enriched for GO categories involved in transcription, ubiquitination, and vesicular transport; among Module 2 TF-associated genes, GO categories corresponding to translation, mitochondrial function and energy metabolism are most highly over-represented. Interestingly, for the human modules, we found significant enrichment for GO categories only for genes that are positively correlated with TFs in Module 1 and only for genes negatively correlated with TFs in Module 2. Because most Module 1 TFs are up-regulated and most TFs of Module 2 are down-regulated in human compared to chimpanzee, the enriched GO functions are all predicted to be relatively increased in human brain.

Fig. 5.

GO categories over-represented among genes associated with the TFs of each module. The program FUNC (25) was used to test for GO categories enriched among associated gene sets of TFs of each module. The sizes of the pie segments are proportional to the number of genes annotated in a given enriched (P < 0.05) GO category. The legend is sorted by size of the pie segments, going from the biggest to the smallest segment in Module 1 and then from the smallest to the biggest segment in Module 2.

Discussion

In this study we examined TFs with differential expression between humans and chimpanzees and the potential impact of these TFs on global gene expression differences between both species. To visualize potential interactions and combinatorial effects of these TFs, we represent the similarity of their associated gene sets in a wTO-based network. Several recent studies of gene coexpression networks have used a wTO measure, based on pairwise correlations of all expressed genes as input for clustering algorithms, to identify gene modules of interest (26, 27). In contrast, we focused on TFs that are differentially expressed in human and chimpanzee PFC together with their correlated genes. After calculating expression correlation between the TFs and other genes within each species, we further enriched for potential direct targets of the TFs using interspecies comparison as a filter. Finally, we implemented a wTO measure that accounts for both positive and negative correlations. This approach allowed us to incorporate genes affected by TFs that are predicted to function as negative regulators, like the KRAB-ZNFs (28).

The resulting wTO network is TF-focused, but it captures the TF's patterns of overlapping influence on a wider collection of brain-changed genes. We should note that the TF-associated genes used to construct the network are ubiquitously expressed, because the initial, intraspecies correlations were calculated using data from all five tissues. Nevertheless, we expect the brain network to have many unique features because ≈23% of the differentially expressed TFs in it have not changed significantly in the other tested tissues. Furthermore, only 25% of TF-associated genes in brain are associated with the same TF in heart and testis, and <7% of these associations are maintained in kidney and liver (Table S7). Hence, the wTO network of other tissues would be constructed of different gene sets and links. This dataset included samples of the PFC as the representative brain tissue, and it is possible that the brain TF network we have identified is specifically active in PFC rather than being common to all regions of the brain. Previous studies have shown that, although gene expression is very similar among different cortical regions (16), cortex displays more human–chimpanzee coexpression network differences than other regions of the brain (26). These findings are consistent with the known evolutionary differences and make the PFC a good choice for our comparative study.

Our analysis has identified a cadre of brain-expressed TF genes that have changed expression patterns in a concerted fashion in recent primate history. The TFs cluster into a tight bimodular network suggesting that, rather than functioning independently, this diverse set of proteins acts coordinately to regulate specific processes differently in human brain compared with chimpanzee brain. This type of concerted TF change is the signature of dramatic shift within larger biological pathways, clues to which may be found in the distinct sets of functions associated with each of the two network modules.

Genes associated with TFs in Module 1 are associated most strongly within GO categories related to transcription, derived from the TFs themselves along with cofactors and chromatin proteins (Fig. 5 and Fig. S2). A second set of functions related to vesicle-mediated transport reflects the prominence of genes encoding small GTPase proteins, which play critical roles in neurite outgrowth, axonal transport, and synaptic transmission (29, 30). These functions are predicted to be relatively up-regulated in human, consistent with added requirements of building and maintaining the larger human cortex. Module 1 TF-associated genes are also annotated in ubiquitination and the unfolded protein response, processes that serve neuroprotective functions and are critical to the maintenance of neuronal health (31). Most Module 1 TFs are of unknown function, but some of them have been implicated in processes related to our GO-based functional predictions. For example, OPTN, has been implicated in regulation of vesicular trafficking and is thought to play a neuroprotective role (32). Other Module 1 TFs function in neuroprotection as well as neurite outgrowth and cortical neurogenesis; most notably in this latter category, ZFHX1B is associated with Mowat–Wilson syndrome, a complex human disorder associated with cortical hypoplasia and microcephaly (33). The tight linkages between these known proteins and other Module 1 TFs indicate roles for the uncharacterized TF genes in forebrain development, neurite outgrowth, and neuroprotection.

Module 2 is associated overwhelmingly with pathways linked to energy metabolism (Fig. 5). The overall up-regulation of energy metabolism genes and metabolites in human brain has also been noted in previous studies (2, 4, 5, 7). While higher metabolic rates may provide additional fuel for the larger human brain, they also yield increased levels of reactive oxygen species (ROS), which can lead to neuronal death (34). It might therefore be predicted that peroxisomal functions, which play a critical role in ROS metabolism and the management of oxidative stress (35), would be salient in Module 2. Our results indicate that ROS-producing and -metabolizing functions are indeed coordinately regulated and point to a module of TFs that may be involved in this important functional coupling. A few well-studied Module 2 TFs provide consistent functional links. For instance, homeobox protein PKNOX1 is involved in the regulation of genes associated with mitochondrial membrane permeability and organelle homeostasis (36), and AHR is involved in the regulation of mitochondrial membrane potential and production of ROS (37).

It is difficult to infer TF hierarchies from the data presented here. However, certain TFs stand out for their position in the network structure. For example, ZNF717, a KRAB-ZNF of unknown function, serves as a bridge node between the two portions of the network, tied by positive links to Module 2 and negative links to Module 1 (Fig. 3, labeled as number 4). It is tempting to speculate that these negatively linked TFs constitute direct repressive targets of ZNF717, a hypothesis that can now be tested experimentally. Like several other KRAB-ZNFs in this network, ZNF717 is a primate-specific duplicate, with an ortholog in chimpanzee and orangutan but not in rhesus macaque (as determined by reciprocal BLAST and investigation of synthetic genomic region); its predicted role as a connector between the two modules would therefore represent a very recent evolutionary change. In addition, some TFs stand out for striking species-specific differences in the number and identity of their network links. Given the dramatic differences in human and chimpanzee brain size, it is intriguing to note that ZFHX1B, mentioned above for its association with human microcephaly, is one of the genes with the highest differences in connectivity between species (Fig. 4).

The primate brain network we have identified contains deeply conserved genes, but our analysis indicates that recently evolved KRAB-ZNF genes play a disproportionate role. Furthermore, the links connecting the two modules are enriched for species-specific links, suggesting that the cross-talk between the two modules is still evolving. This argues that the observed human brain network is a very recent evolutionary invention. To test this hypothesis, to investigate how the human network formed from a putative ancestral network and how newly evolved genes have been incorporated, it will be important to analyze expression data from additional primate species, for instance orangutan and rhesus macaque, in future studies.

Although several brain-changed KRAB-ZNFs have been implicated in brain development or human neurological disorders (38–40), most of these proteins are presently of unknown functions. Data presented here provide clues linking them to potential target genes and regulatory partners in well-studied functional pathways. These data also suggest concrete approaches to test these hypotheses in future studies. For example, siRNA knockdown and ChIP experiments with ZNF717 and other “bridge” genes can test the relationship between these TFs, their associated gene sets, and other TFs in the network. However, any gene in this brain-changed TF network can play an important functional role. Given the fact that the functions of most of these brain-expressed TFs are unknown, new data regarding their expression patterns and interactions in specific cell types, their target genes, and their DNA binding activities will further our understanding of the differences between primate species and will provide new clues to the basis for phenotypic differences between the human and chimpanzee brain.

Materials and Methods

Microarray Data Preprocessing and Analysis.

We downloaded CEL files for HG U133 Plus 2.0 arrays from a previous study; tissue samples were taken from apparently healthy male individuals of various ages from each species, all of which had died a sudden death (11). Note, that the donors of the tissues are not the same individuals across the five tissues. We aligned probe sequences against the hg18 and panTro2 genomes and removed all probes with DNA sequence differences between species similarly as previously described (11). Next, probes with multiple matches in the hg18 or panTro2 genome were removed. Probe sets with less than four remaining probes were discarded from further analysis, leaving 47,506 probe sets with on average 8.08 (median 8) probes per set. The numbers of probes per probe set remaining for all probe sets, for TFs, or for KRAB-ZNF genes were not significantly different from each other (Kruskal–Wallis test, P > 0.22; all pairwise comparisons). RMA (41) expression values and MAS5 detection P values were calculated with the Bioconductor “affy” package (42). Probe sets were considered to be “expressed” at a detection value of P < 0.05. For genes with multiple expressed probe sets, a mean expression intensity was calculated.

FDR was calculated for unadjusted P < 0.01 (two-sample t test) using a PT (based on random assignment of species labels 1,000 times) and comparing the median number of significant genes in the permuted datasets with the observed number. To test for over-representation of TFs and KRAB-ZNFs among differentially expressed genes, we used PTs (based on drawing random genes 1,000 times). P values were calculated based on the number of times that the number of significant genes in a permuted dataset was equal to or larger than the observed number.

TF-Associated Gene Sets and wTO Calculation for TF Network.

We defined genes as associated with a TF if they were significantly positively or negatively (P < 0.05) correlated with the TF in a Spearman rank correlation test performed across all five tissues and all six human individuals (30 samples) and were differentially expressed between human and chimpanzee brains in the predicted direction (see Structure of the human brain TF network). We then calculated the weighted topological overlap (20–22) matrix starting from the adjacency matrix A = [a_ij], with a_ij = Corr(i,j) ∈ [−1,1] and a_ii = 0, representing the 79 differentially expressed TFs and the 1,273 members of their associated gene sets and containing a total of 29,361 significant (P < 0.05) pairwise correlation scores. Note that the wTO also incorporates the correlations of two TFs' associated gene sets. Further note that previously published expressions for the wTO (e.g., ref. 20) are only appropriate for positive adjacency matrices. Here we have extended the range of validity of the wTO to also include the case a_ij ∈ [−1,1] when a_ij ≤ 0⇒a_iua_uj ≤ 0 for all u (and conversely a_ij ≥ 0⇒a_iua_uj ≥ 0 for all u):

where the weighted connectivity of a node i is k_i = Σ_j |a_ij|. Thus, when generating the wTO network by calculating ω_ij between the 79 TFs, we take all members of TF associated gene sets into account when assigning a wTO score to each TF pair. We used a cutoff of ∣ω_ij∣ > 0.3 for our network analysis (see SI Text for a discussion of this cut-off choice).

Enrichment of GO Groups.

To relate GO categories to a module in the wTO network, we counted how often a gene was a member of a gene set associated with a TF of Module 1 or Module 2. Next, FUNC (25) was performed with the following settings: Wilcoxon test, cut-off of five genes per group; GO Annotation from April 2008; and GO categories with P < 0.05 before and after refinement are reported.