Molecular networks and the evolution of human cognitive specializations

Abstract

Inroads into elucidating the origins of human cognitive specializations have taken many forms, including genetic, genomic, anatomical, and behavioral assays that typically compare humans to non-human primates. While the integration of all of these approaches is essential for ultimately understanding human cognition, here, we review the usefulness of coexpression network analysis for specifically addressing this question. An increasing number of studies have incorporated coexpression networks into brain expression studies comparing species, disease versus control tissue, brain regions, or developmental time periods. A clearer picture has emerged of the key genes driving brain evolution, as well as the developmental and regional contributions of gene expression patterns important for normal brain development and those misregulated in cognitive diseases.

Introduction

The evolution of human cognitive specializations has been the subject of human ruminations for perhaps as long as the evolution of self-reflection itself. Here, we consider some of the outstanding questions in the field of human cognitive evolution that can be addressed through the comprehension of evolved molecular networks in the brain. An important unanswered question is how genes and, perhaps more importantly, how gene networks have evolved to impart cognitive specializations in humans. What are the key players in human gene networks that are important for specific specializations, such as language?

To begin, one must address whether genetic and/or genomic changes are important for cognition. Most cognitive diseases (e.g. autism and schizophrenia) have a strong genetic component, and there is significant evidence that the evolution of the human genome has been permissive for both cognitive evolution as well as increased risk for developing cognitive diseases [1-5]. Are these genes disrupted in cognitive diseases under selective pressure, and are they important from an evolutionary standpoint?

There are precedents for changes in single genes at the DNA level affecting cognitive specialization. For example, it has been suggested that two human-specific amino acid changes in the transcription factor FOXP2 are under positive evolutionary selection due to the role of FOXP2 in human speech and language [6]. When the human-specific modifications of FOXP2 were knocked-into the endogenous mouse Foxp2, these animals displayed a number of behavioral and pathological changes including alterations in ultrasonic vocalizations [7]. These two amino acids are also sufficient to direct an altered transcriptional program [7,8]. Furthermore, beyond evolution of protein sequence, changes in the regulation of gene expression will likely have profound consequences relevant to cognitive function [9]. For example, recent analysis of differential DNA promoter methylation between humans and chimpanzees has implicated epigenetic control of genes associated with disorders highly prevalent in humans such as autism, neural-tube defects, and alcohol dependency [10].

We propose that studying the emergence of altered gene networks may provide greater insight into cognitive evolution than considering the modifications to only a handful of genes. Scale-free networks exist across multiple domains, ranging from bacteria to the internet [11,12]. The prevalence of such networks across nature reinforces the idea of using gene networks as an improved model for understanding relationships among gene expression changes. In contrast to differential expression approaches, which focus on how single genes change between conditions, a network approach allows for the rapid prioritization of the most interconnected genes, or hub genes, from complex datasets, such as those often generated across tissues and species (Figure 1). In addition, the importance of these hub genes has been validated through the assessment of network structure upon removal of expression of a hub gene [13]. In other words, if one attempts to build a network using expression data from a knockout animal for one of the hub genes, the network essentially falls apart [14]. While the majority of this review will focus on one particular technique, Weighted Gene Coexpression Network Analysis (WGCNA, see Box 1), for building genomic networks, there are many other methods for achieving a similar goal, such as the transcription-factor-focused weighted topological overlap (TF-wTO), Detecting Association With Networks (DAWN), Pearson, Spearman, or Kendall correlations, and Rank Theil-Sen [15-21] (see Box 2).

Figure 1. Understanding the interaction of individual genes at a network level in the evolution of cognitive specializations.

Here, we illustrate that compared to a single gene approach (e.g. FOXP2 regulation of specific cognitive-related genes such as CNTNAP2, contactin associated protein-like 2), network analysis can uncover more subtle and additional candidate genes which could exert important influence on the higher-order function in question (e.g. language). Thus, the incorporation of a coexpression network approach in evolutionary comparisons provides increasingly more information than single gene approaches. The network image is modified from Konopka et al., 2009.

BOX 1. Weighted Gene Coexpression Network Analysis (WGCNA).

WGCNA takes an unbiased approach for ascertaining the relationships among all genes queried across all samples in a dataset [62,63]. Pairwise correlations between all gene expression profiles in a dataset are calculated and input into an adjacency matrix. For human brain RNA-sequencing datasets, this can translate into a square matrix of order up to approximately twenty thousand, depending on the spatio-temporal pattern of brain gene expression assessed. A significant computing resource is needed to analyze such matrices if all expressed genes are used in building the matrix, so often the analysis is restricted to some fraction of the original gene set (typically the six thousand most highly correlated genes.) Once the matrix is built, it is raised to a power to approximate scale free topology and construct the network. Genes with highly correlated expression profiles are then grouped together in clusters called modules, which may correspond to biological pathways or similarly associated groups of genes. Gene relationships within a given module can then be assessed using a number of visualization tools such as Visant [64] or Cytoscape [65]. The graphical representation of a module aids in rapidly identifying hub genes and other biologically meaningful patterns of coexpression within a given module. Overall, WGCNA provides a means for prioritizing specific genes from large expression datasets, particularly those with biologically salient relationships that might otherwise be missed using differential expression approaches. One of the limitations of WGCNA is that its power depends upon the number of samples as well as the complexity of the dataset. Therefore, a minimum of roughly 8-12 samples is required for WGNCA, although more samples are always desirable. In terms of complexity, WGCNA is most powerful when samples contain varying attributes such as different types of cells, tissues, genotypes, or developmental time periods. Finally, WGCNA is limited to expression information as input data, although additional information (e.g. genetic, protein-protein interactions, etc.) can be superimposed on the resultant modules after completion. A new algorithm termed DAWN has directly integrated gene coexpression built using a WGCNA framework with genetic risk factors for autism [17], opening up the possibility of direct integration of WGCNA with a multitude of factors that can influence gene expression (e.g. epigenetic data, transcription factor binding, etc.).

BOX 2. Other methods for building brain coexpression networks.

Several other coexpression approaches have been utilized to understand brain transcriptomes. The TF-wTO method is similar to WGCNA in calculating topological overlap where the links are weighted based on the commonality of the correlated genes between nodes. A major distinction of TF-wTO is that it has been modified from the original definition of topological overlap [66] to specifically examine transcription factor networks by taking both positive and negative correlations into account [15,16]. Thus, in a TF-wTO network the nodes will only represent transcription factors and the links between nodes will be weighted based on the common set of correlated genes between these factors. The TF-wTO approach is a good method for focusing on major regulatory elements in the brain using genome-wide expression data.

Pearson product-moment or Spearman rank-order correlations can provide straightforward relationships between genes in a dataset, however both methods are prone to identifying false positive relationships. Pearson, which WGCNA uses as input, has been used extensively in brain and comparative studies [18-20], and is most appropriate when the correlation of variables under comparison (e.g. brain regions) can be assumed to be linear. Spearman, which is essentially a ranked Pearson correlation and is used by TF-wTO as input, is more appropriate for datasets when such a linear relationship cannot be assumed (e.g. in a developmental dataset).

Comparative networks across species

Although hundreds of differentially expressed genes between human and other species have been identified by directly comparing gene expression in the brain [22-25] (for more details see the review by Somel in this issue), translating these findings into meaningful functional distinctions between species has been difficult. Comparing genomic data in humans with those of our closest genetic relatives on a network level, in contrast, may provide a more straightforward approach to identifying networks important for human-specific cognitive specializations. WGCNA offers an unbiased view of relationships within gene networks and has been used to directly test this idea.

The first application of WGCNA to address human cognitive evolution compared gene coexpression networks in human and chimpanzee brains [26]. By using data from multiple brain regions, modules of coexpressed genes corresponding to specific regions could be characterized and compared between species to identify important evolutionary hub genes. Interestingly, modules specific to the cerebral cortex were significantly less conserved among species than noncortical modules [26]. A more recent study took the same approach to identify human-specific coexpression networks by comparing human and chimpanzee brain expression data, but also including data from rhesus macaque as an outgroup [27]. This analysis again revealed an enrichment of genes differentially coexpressed in the human neocortex, whereas gene expression in non-neocortical regions was more highly conserved across primate species. Furthermore, many of these human-specific modules were enriched for genes implicated in cognitive diseases [27]. This divergence in the neocortex suggests that neocortical evolution has permitted greater flexibility among gene coexpression, and perhaps some of this divergence may underlie cognitive enhancements as well as disease-susceptibility.

Although directly studying human samples yields results clearly relevant to human health and evolution, rodent model systems currently offer the most powerful genetically manipulable mammalian model for studying human cognition and disease. WGCNA can also help to validate mouse models of human disease and cognition. An example demonstrating the power of combining network-level analysis of human data with a mouse genetic model is a study of the role of progranulin mutations in the development of frontotemporal dementia [28]. WGCNA was used to assay transcriptomic changes with alteration in progranulin expression in a human neural progenitor model, and uncovered a role for Wnt signaling in disease pathogenesis. This finding was confirmed using WGCNA from postmortem brain tissue of patients. Upon switching to a mouse model, it was then demonstrated that progranulin knockout mice also exhibit alteration of Wnt signaling genes, such as upregulation of Fzd2 [28]. WGCNA has also been used to identify differentially coexpressed genes in mice following a stress paradigm. A number of orthologous human genes were then confirmed as dysregulated in human post-mortem brain tissue from depressed patients, highlighting the prioritization power of WGCNA [29].

Another study compared the brain transcriptome of human and mouse and identified highly conserved coexpression modules between mouse and human brain for common features such as genes involved in cell-type classification. However, human-specific modules were also identified, including one enriched with Alzheimer's-disease-associated genes, supporting the idea of certain cognitive diseases being human-specific phenomena [30]. Additional network analysis of Alzheimer's disease-relevant data identified network changes and hub genes specific to Alzheimer's disease compared to normal aging brain [31]. By shedding light onto the convergent and divergent pathways in mouse models and humans, network analysis provides a systematic framework for reflecting on the relevance of mouse models to human cognition and disease.

Network approach to human cognition and disease

Given that human-specific evolution of cognitive specialization may underlie our unique susceptibility to neuropsychiatric disease [1-5], understanding the transcriptome-level etiology for these conditions could be invaluable. Autism spectrum disorder (ASD) is a highly heritable and prevalent neurodevelopmental disease which shows remarkable genetic heterogeneity [32]. Gene network-level analysis is thus well suited for studying how numerous genetic risk factors can converge on common molecular pathways in the brain. Previous studies have supported the use of in vitro human neural progenitor cultures as a model system for human neuropsychiatric disease, such as ASD [33]. As this model system is genetically tractable, it permits the study of ASD transcription networks using a human genetic background. The whole-genome transcriptome was analyzed upon differentiation of primary neural progenitor cultures into a post-mitotic neuronal state, and through the use of WGCNA, the unbiased network structure revealed a significant enrichment for ASD-associated genes in specific modules [33]. Moreover, this approach was able to predict ASD genes: the second most connected hub gene in one of the ASD-enriched modules, DNER (delta and notch-like epidermal growth factor repeat containing), was subsequently shown to have genetic association with ASD [34].

Furthermore, using WGCNA, the transcriptome organization of normal and autistic brain tissue has been investigated. Interestingly, this approach revealed that gene expression signatures that differentiate neocortical regions are altered in ASD brain, resulting in a more homogenous pattern of gene expression in the ASD brain and suggesting that potential alterations in cortical patterning in ASD can be reflected at the gene expression level [35]. More recent studies have attempted to address how disparate genetic causes of ASD lead to commonly altered biological functions. ASD-associated genes identified using distinct methods (e.g. rare genetic variants, differentially expressed genes, etc.) were mapped onto coexpression networks built using longitudinal expression data from human cortical development [36]. The ASD genes were enriched in modules that distinguished both the etiological classes of the genes together with altered biological or cellular functions in development. For example, rare variants were enriched in prenatal modules containing genes involved in transcriptional regulation, whereas differentially expressed genes were enriched in postnatal modules containing genes involved in synapse formation. ASD genes were also found to have specific superficial cortical layer expression that may underlie some of the cortical pathology associated with the disease [36]. A second study took a somewhat similar approach, but instead built the coexpression networks by seeding the dataset with nine high-confidence ASD genes identified from de novo loss of function mutations in ASD patients [18]. This complementary approach yielded ASD-enriched modules during midfetal cortical development containing genes important for deep cortical layer neurons [18]. Together, these three studies are beginning to tease apart the relationship among the hundreds of genes associated with ASD and put these relationships into a biological context that is consistent with ASD phenotypes.

Additional methods that incorporate both genetic and gene expression data similar to WGCNA approaches have also been shown to provide valuable coexpression analysis relevant to ASD. For example, the DAWN algorithm builds upon the study discussed above [18] to predict putative ASD genes and ASD-relevant coexpression modules. DAWN can be further applied to new datasets from other cognitive disorders, thereby complementing the WGCNA approach [17]. Finally, genetic information alone (e.g. copy number variation (CNV) and exome sequencing data) can be mined to build networks with functional connectivity through the incorporation of methods such as DAPPLE [37-39], which incorporates protein-protein interaction information, and NETBAG [39,40], which prioritizes genetic information. Interestingly, comparisons of CNV evolution among great apes and humans has led to the first genetic identification of a developmental disorder in a chimpanzee that resembles a human neurodevelopmental syndrome [41]. Thus, further study of CNVs and networks built from such genetic information in humans and closely related species will shed more light onto this important evolutionary force and other cognitive disorders.

Schizophrenia (SCZ) is another highly heritable psychiatric disorder where WGCNA has proven insightful. In one study, two WGCNA modules were shown to be significantly correlated to schizophrenic patients versus controls [42]. One of these modules was associated with gene expression changes in the cerebral cortex and was enriched for genes involved in neuron development and differentiation; the second module was associated with both the cerebral cortex and cerebellum and was enriched in genes involved in responses to stress and neuronal protection [42]. Another study, also using WGCNA in schizophrenic patients and controls, found oligodendrocyte- and neuron-specific modules were enriched for SCZ-associated genetic variants [43]. Furthermore, coexpression network analysis between brain regions uncovered an attenuation in gene expression patterns that normally distinguish neocortical regions of controls [43], a finding similar to patients with ASD as described above. A larger gene expression study relied on peripheral blood samples of schizophrenic patients, both treated and untreated, and healthy controls [44]. Two modules were identified in both treated and antipsychotic-free patients, one of which was significantly enriched for brain-expressed and SCZ-associated genes [44].

Gene coexpression networks throughout human brain development

Prenatal gene expression networks lay the groundwork for how the human brain develops and responds to environmental cues. Such networks ultimately drive cognitive specializations and, as already discussed above, are at risk in cognitive disorders such as ASD [18,36]. Many studies have sought to provide comprehensive data on the developing brain transcriptome. WGCNA of mid-fetal human brain found region-specific co-expression networks as well as hub genes with unknown functions [45]. By increasing the number of brain regions examined and extending the time periods investigated from prenatal through adulthood, WGCNA can uncover distinct regional and temporal coexpression modules [46]. Interestingly, the majority of the variance in the data derives from the prenatal samples, while the postnatal transcriptome exhibits more homogeneity [46]. This spatiotemporal atlas of human gene coexpression has been compared to gene expression in the developing rhesus macaque brain. Approximately 60% of hub genes identified to have distinct prenatal developmental expression patterns in human demonstrate different expression patterns in the developing rhesus macaque brain [47]. In addition, this study comprehensively examined whether the human brain exhibits lateralization in gene expression, a potential mechanism for driving specializations such as language, however, there were no data to support the idea of lateralization of gene expression [47]. More recently, WGCNA has also been used to delineate germinal layer specific modules derived from human prenatal expression data [48]. Transcriptional differences among layers in the human cortex were identified that may provide insight to human-specific features of cortical development that ultimately underlie human cognitive specializations [48]. For example, superficial cortical layers have increased cortico-cortical connectivity relative to deeper layers. Thus genes with differential coexpression between upper and deeper cortical layers might be at risk for altered intra-cortical connectivity. Such changes in cortical connectivity have been linked to human cognitive diseases such as autism [49].

Future directions

How can we use molecular network information to gain further insight into the evolution of human cognitive specializations and their associated diseases? WGCNA and other coexpression analyses such as TF-wTO, DAWN, and Pearson correlations allow for rapid identification of the most potentially fruitful targets for investigation (Table 1). Recent advances in induced pluripotent stem cells [50-53] and human neural progenitors [8,33,54] provide genetically tractable human model systems in which one can manipulate these most interconnected genes to explore their role in neuropsychiatric disease.

Table 1. Examples of hub genes associated with human disease.

WGCNA and other coexpression methods allow rapid identification of the most interconnected hub genes as genes of interest in association with human cognitive disorders.

Gene Name	Description	Associated Human Cognitive Disorder
A2BP1 (MIM 605104) [35]	RNA binding protein, fox-1 homolog (C. elegans) 1	ASD
ABCF1 (MIM 603429) [44]	ATP-binding cassette, sub-family F (GCN20), member	SCZ
CLOCK (MIM 601851) [27]	clock circadian regulator	delayed sleep phase disorder
CUL3 (MIM 603136) [17, 18]	cullin 3	ASD
DLX1 (MIM 600029) [48]	distal-less homeobox 1	ASD
DNER (MIM 607299) [33]	delta/notch-like EGF repeat containing	ASD
GSK3B (MIM 605004) [30]	glycogen synthase kinase 3 beta	Alzheimer's disease
GTF2I (MIM 601679) [15]	general transcription factor II-I	ASD
KCNK12 (MIM 607366) [47]	potassium channel, subfamily K, member 12	epilepsy
LGI2 (MIM 608301) [27]	leucine-rich repeat LGI family, member 2	epilepsy
MOG (MIM 159465) [43]	myelin oligodendrocyte glycoprotein	SCZ
NEFL (MIM 162280) [26]	neurofilament, light polypeptide	Charcot-Marie-Tooth disease
NOTCH2 (MIM 600275) [42]	notch 2	SCZ
NRXN1 (MIM 600565) [36]	neurexin 1	ASD
SYNJ1 (MIM 604297) [31]	synaptojanin 1	Alzheimer's disease and bipolar disorder
TBR1 (MIM 604616) [17, 18]	T-box, brain, 1	ASD

Modeling human genetic changes that have occurred since the divergence of modern humans from the common ancestor of other hominin species (e.g. Neandertals and Denisovans) in cells and animal models will provide particular insight into the most recent changes on the human lineage. It will be provocative to make a multitude of these changes in the same cells or animals and assess alterations in resultant gene networks. The use of new genome editing techniques, such as the CRISPR/Cas9 system [55,56], will provide a rapid means for achieving this goal. Making humanized mice ([57] and Enard review in this issue) will allow us to examine the effect of human-specific changes in protein sequence or expression. Investigating other functional evolutionary changes, such as gene duplications specific to the human lineage (e.g. SRGAP2 [58,59]), will also prove insightful.

A major challenge for using either human cell lines or humanized mice is that neither can really be expected to recapitulate the complex evolved network in the intact human brain or human cognition. Even newly available transgenic monkeys will still likely fall short of such modeling due to the myriad of differences between a human and monkey brain. One approach that is being expanded upon is the correlation of human brain imaging data with genetic information. In humans, functional brain imaging has been correlated with genetic data, and structural brain imaging has been correlated with post-mortem gene expression data [48,60,61]. The next logical step would be to integrate functional brain imaging with post-mortem gene expression and to do this with both control and patient populations. Similar approaches could also be attempted in non-human primates for a comparative study. Such studies would provide a window into the relationship among genes and the working human brain. Individual genes identified could be further examined in detail using cell and animal models.

In summary, while it is impossible to “prove” which specific genetic and genomic changes have permitted specific human specializations, developing methods for assessing human genomic modifications are a critical step towards gaining a basic understanding. Such comparative genomic approaches that yield novel human expression networks coupled with functional assessment in model systems are a powerful means for achieving this goal.