Neuroscience in the era of functional genomics and systems biology

Public data sharing and resource generation

Omics data provide a platform for small-scale and large-scale in silico discovery and hypothesis generation that often goes beyond the scope of the original experiment. This notion supports the proposition that data from microarray studies have most value in the public domain, where they can be mined or combined with other studies²¹. However, proteomic and transcriptomic data are inherently less generic than sequence data, so obstacles such as variable sample preparation, experimental annotation and platform differences prevent a single fully unified data resource. The development of a framework, MIAME (minimal information about a microarray experiment), allowed the creation of centralized repositories for these data created on multiple platforms in laboratories around the world (Box 2). A growing number of studies that have capitalized on such databases, including those cited above and others^22–27, have demonstrated that the widespread availability of these proteomic and transcriptomic data are of great benefit. One caveat with any resource based on compilation from multiple sources is that the data are only as good as the experimental and sample phenotype annotation. As microarray-based hybridization platforms for expression analysis give way to sequencing-based approaches, these data will become more generic, diminishing, but not eliminating, cross-platform compatibility issues.

In parallel with the growth of generic data repositories, standardized resources built by coupling high-throughput analyses with bioinformatics tools provide a clear demonstration of the value of large-scale omics data. The development of such resources requires significant foresight, planning and investment, but their worth is unquestionable if they are properly managed. Additionally, just as for DNA sequences and the genome databases, errors exist. Quality control must be balanced with throughput. Ideally, such resources would provide access for users to post corrective data or suggest annotations that would enhance the accuracy of the resource in real time.

Use of these resources to reanalyse public data is often a necessary feature of the well-rounded systems biology approach, raising the issue of the primacy of data analysis versus data generation. There is a clear tension between biological advances obtained in this way and the generation of new data, the latter usually carrying more weight in terms of novelty and degree of difficulty. Often, data analysis is not considered to be as much of an advance as the generation of new data, no matter how narrow the new data may be. However, a new perspective must be adopted to take into account studies that analyse data generated by others or in the public domain, to lead to new levels of understanding, or provide database tools around these data, even if no ‘new biology’ is done. The notion of what constitutes an experiment must expand to include analysis only, in addition to standard bench work. Similarly, we need to ensure that the next generation of neuroscientists has the quantitative skills to allow them to take full advantage of these opportunities. Here we highlight a few examples of advances based either on the generation of new omics data resources or on the analysis of data from such resources to provide fresh biological insights.

Integrating genetic and phenotypic data

Another approach to adding systems-level structure to transcriptome data is the analysis of these data in concert with genetic and phenotypic data to integrate across all three levels of observation. Thus, the advent of expression quantitative trait locus (eQTL) analysis is a major advance in integrating large-scale genomic or genetic data sets to understand a model system or cohort of patients at a systems level. In eQTL mapping, gene expression data are used as a phenotype on which to base quantitative genetic association mapping (Fig. 1), the rationale being that gene expression is a more proximal, intermediate quantitative phenotype to the underlying genetic risk than heterogeneous behavioural or anatomical phenotypes. Because large samples are necessary to provide statistical power for whole genome-wide association, peripheral tissues such as blood cells were used in the initial pioneering studies in humans⁴⁶. Recent studies in animal models demonstrate its utility in brain samples^47,48. A very promising avenue is the combination of gene network analyses of expression for eQTL analysis as first demonstrated in ref. 49 and, more recently, ref. 50, both of which studied phenotypes related to metabolism. Another intriguing recent paper merges a different form of gene network analysis with structural chromosomal variation identified in patients with autism, suggesting a role for genes related to glycobiology in this disorder⁵¹.

One comprehensive, user-friendly interface for eQTL analysis in mouse is the database and toolkit provided by WebQTL¹⁷ (http://www.genenetwork.org). WebQTL allows integration of genetic polymorphism data from several strains of mice and rats, gene expression data from microarrays, and neurobiological traits based on neuroanatomy, pharmacology and behaviour. These tools have been implemented to provide proof-of-principle evidence of the utility of this approach; behavioural, microarray and genotyping data from a recombinant inbred mouse strain have been recombined⁴⁷ to uncover significant QTLs correlating with gene expression and neuronal phenotypes, including, most interestingly, synapse function. Numerous QTLs were mapped to the majority of specific transcripts, which included a number of loci that appear to be ‘master’ transcriptome regulators, because they account for the expression of hundreds of genes. This study also uncovered QTLs that have polymorphisms associated with phenotypes of a specific behaviour. These links indicate that QTL-related transcript regulation can have wide-ranging effects on numerous functions, from the synapse to behaviour. Most of the transcriptional data in WebQTL is from whole brain, and the resource would benefit greatly from transcriptome analysis at higher spatial resolution, because significant variation occurs regionally⁴⁸. However, this work connects genetic polymorphisms to RNA levels and to variation in both anatomy and behaviour, demonstrating the power of systems-level approaches to uncover complex neurobiological interactions.

QTL analysis has been extensively employed in rodents, but so far only two studies have combined gene expression data from human brain with genetic polymorphism data^52,53. In the first study, nearly 200 pathologically proven normal cerebral cortical samples were merged with genome-wide SNP data⁵³. Although this is a small number from the standpoint of genetic association in complex disease, the difficulty and cost associated with profiling human tissue makes this an important demonstration of the approach. Furthermore, this study identified a large number of potential cis-QTLs that, once replicated, can serve as a basis for understanding the direct functional consequences of human genetic variation on brain gene expression in disease. Supporting this notion, these data were used in a recent comparison with eQTLs identified in a parallel study of the brains of sufferers of Alzheimer’s disease, identifying several novel and known candidates for this complex brain disease⁵². The next step for these, and similar, studies is replication of the disease-association or eQTL-association findings in an independent cohort.

Accelerating discovery through next-generation sequencing

The advent of next-generation sequencing has raised multiple possibilities for the study of all layers of regulation leading to gene expression, changing the way we think about designing functional genomic experiments. Early studies clearly indicated that sequence data surpasses microarrays for studying gene expression, in terms of both depth of coverage and analysis of splicing, among other factors^54,55. This technology can be used to quantitatively query not only mRNA expression but also RNA splicing, microRNA expression, epigenetic modifications, DNA binding, copy number variations and genetic deletions, insertions and mutations. Bridging these different outputs to generate a complete picture, from DNA to modified and bound DNA to precursor mRNA to mature mRNA, challenges even the most seasoned systems biologist. In addition, combining these data with proteomic data sets will completely revolutionize our understanding of the central dogma of molecular biology in action in any given cell. Another feature of next-generation sequencing that surpasses microarrays is the ability to assay species for which arrays do not exist, and eliminate the bias inherent in cross-species comparisons on microarrays⁵⁶. The major challenges in the use of next-generation sequencing revolve around data storage and handling, but they are not unlike those encountered in the early years of microarray technology (Box 3), despite the amount of data involved being several orders of magnitude greater than that gained from microarrays. The relative platform independence of next-generation sequencing, leading to its more generic nature, will lower barriers for data sharing, meta-analysis and integration.

Moving from lists to networks

Even the most elegant of the multilevel functional genomic approaches essentially involve the analysis of overlapping lists of microarray or other data. In most transcriptomic or proteomic studies, data are organized in order from most differentially expressed to least. This is useful because these very large data sets need to be put into a form that allows them to be analysed and understood. However, such simple levels of data organization cannot and do not represent even a small fraction of the potential information inherent in the data. Furthermore, the use of standard pathway tools limits the analysis to known relationships. Fortunately, we are just beginning to appreciate that the data itself has an underlying structure, and acknowledgement of this network structure as a general biological principle is opening new avenues of complementary investigation^57,58.

The demonstration that transcriptome data can be organized into networks based on expression correlation, the application of graph theory, and robust statistical methods to develop weighted gene co-expression network analysis (WGCNA; http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork) raised expression profiling to the level of systems biology by elucidating the relationships among all of the elements being studied^27,59 (Fig. 2). Several studies have now demonstrated that such networks derived from human or animal brain represent a reproducible and robust structure (for example refs 22–24), and that network position has significant functional implications. Similarly to earlier work in simpler organisms, this work has shown that proteomic and transcriptomic data derived from complex tissue such as brain show a high level of correspondence^24,25. Such structure can be used to inform a new level of neuroscientific investigation that is not possible using standard analysis of differential expression^22–25.

For example, one of the first such studies²³ showed that gene networks could be used to provide a unifying method of identifying transcriptional targets of human brain evolution in the context of the neutral model of evolution for the transcriptome⁶⁰. More recently, it has been demonstrated that the human cerebral cortex transcriptome is organized into a robust network and shown how its modular nature could be used to drive functional understanding and discovery in many directions, including the identification of markers for human adult neural stem cells²⁴. One of the more remarkable observations here, especially with relevance to the discussion of neuronal heterogeneity and the need for individual cell study, is that from whole tissue WGCNA can recover modules that represent the transcriptional programs of the major cell classes in human brain, an example of in silico tissue dissection²⁴. Importantly, these analyses relied on several public data sources, including the ABA and a resource provided by the transcriptional analysis of purified neurons, astrocytes and oligodendrocytes⁷.

These same network methods have recently been applied to provide a platform-independent comparison of pathways altered in normal ageing and dementia²². Data from two different microarray studies were combined and reanalysed to identify overlapping network modules corresponding to the synapse and mitochondria, illustrating common mechanisms shared by normal ageing and Alzheimer’s disease. Similarly, reanalysis²⁵ of single-cell expression data¹⁰ has confirmed the initial findings and led to new biological insights, including a major transcriptional distinction between two classes of mitochondria: those located in the synapse and neuronal processes, and those located in the cell body. Both of these studies show how network methods can elucidate organelle-specific or cellular-component-specific expression profiles without the need for their purification (another form of in silico dissection), and further demonstrate the value of data placed in the public domain for subsequent reanalysis or use by others.

Proteomic networks can be constructed either through the investigation of actual protein–protein interactions or by the correlation of protein levels across samples or observations. The coupling¹⁵ of a high-throughput proteomic screen for protein interactions with bioinformatic analyses has been used to uncover an interaction network specific to ataxia. A screen for protein partners of ataxia-associated proteins was performed using the yeast two-hybrid system and a human-brain complementary DNA library, ultimately validating interactions in silico and in vivo. These interactions were expanded by culling known protein–protein interactions from public databases and the literature to create a final interaction network containing nearly 7,000 protein–protein pairs among almost 4,000 proteins. This network is an important resource for researchers studying neurodegeneration; it has already either provided confirmation of specific interactions or the impetus for their study^61,62.

A more recent example of the combination of network methods and proteomic analysis identified a large number of postsynaptic protein complexes that show significant overlap with putative schizophrenia-susceptibility genes³⁷. Five interconnected protein clusters were found by combining in vivo tandem affinity purification of PSD-95 (also known as DLG4) and mass spectrometry with network analysis. Almost half of the proteins in the clusters were related to at least one neurological disease, and, remarkably, 70% of the constituents of one of the clusters were specifically linked to schizophrenia. These and other early protein network analyses in brain should fuel future investigations and database development to catalogue and organize the entire neuronal proteome.

Conclusions and future directions

Many of the systems-level approaches that have been described here involve omics-level analysis of large data sets that span one or two dimensions of biological experimentation (Fig. 3). The most ambitious systems approach would see the integration of enormous data sources across multiple levels of genotypic, genomic, proteomic, epigenetic and phenotypic data (Fig. 3). Great efforts and powerful tools will be needed to achieve this in neuroscience, including the standardization of complex measurements of animal and human nervous system phenotypes and the development of accompanying ontologies⁶³. Most forms of molecular data can be made into relatively generic forms, but translating complex phenotypes, from neuronal morphologies to the cognitive and behavioural profiles of neuropsychiatric disease entities, will require far more groundwork. Furthermore, although there is a compelling rationale to study nervous system phenotypes as they evolve over time, in many cases current funding and review processes are barriers to the collection of longitudinal data.

However, even one data dimension, such as knowledge of transcriptome organization by means of network analysis, can promote large conceptual leaps by providing a new view of gene function, independent of the proteome or genome. For example, we have observed groups of genes annotated by gene ontology as mitochondrial, ribosomal and proteasomal within single co-expression modules in several data sets. From a proteomics standpoint these organelles are distinct, but the high co-expression of genes within them suggests that they are part of a highly coordinated and interconnected system that spans cellular compartments as they are typically defined. Here transcriptional network analysis provides a new, systems-level view of gene function within cellular pathways that was not readily apparent from proteomic or genomic data alone.

Another frontier in systems-level analysis in neuroscience is highlighted by the necessary scale and complexity of endeavours analysing human phenotype data from a genetic perspective. Recent work combines text mining^64,65 with genetic analysis and modelling to integrate across disease phenotypes^66,67. In a remarkable study⁶⁷, the co-morbidity among 161 different medical conditions abstracted from 1.5 million medical records was analysed in the context of a basic probabilistic model of inheritance. This work not only detected known genetic relationships between diseases, but also showcased the power of such methods to detect unexpected genetic relationships between human disorders considered to be distinct. For example, it predicted significant causal overlap between disorders such as bipolar disorder, autism and schizophrenia, as well as a shared genetic connection among autism and autoimmune and infectious disorders. In the future, such work should be greatly facilitated by detailed and standardized clinical phenotype ontologies, as well as integration with other levels of quantifiable phenotypic data, from molecular biomarkers such as gene expression and epigenetic profiles to neuroimaging data.

The application of systems-level analyses to neuroimaging data alone, based on a theoretical framework similar to WGCNA, has also begun to reveal remarkable insights into human brain networks⁶⁸. This work shows a robust relationship between specific aspects of brain functional networks and structural connectivity⁶⁹. Analogously to gene-based networks, particular functional modules and key hub regions can be identified on the basis of maps of functional or structural connectivity. At this point, the possibility of connecting such networks to the panoply of potential genetic, genomic or environmental factors that regulate them may seem distant. However, there is likely to be a reasonable systems biology solution. Standardized network analyses based on graph theory, such as WGCNA, could be used to integrate anatomically based brain networks with gene expression or proteomic networks from profiling of the same brain regions from post-mortem samples (Fig. 3b). This type of analysis would reach a level of integration of ‘neuro-omic’ and neuroimaging data that goes far beyond the current state of genotype–phenotype correlations. Such studies would be greatly accelerated by the requirement that data from all published neuroimaging studies be made publicly available in a usable, normalizable form, similar to omics data.

In many ways, the omics revolution provides an example of the enormous value of public data sharing in making efficient use of our relatively limited resources to fuel scientific and biomedical advances. The value of these data highlights the urgent need for common language and measurements in addition to incentives and portals for data sharing from even the smallest-scale studies. In parallel, it also speaks to the need for large-scale collaborative endeavours that collect data in a way that facilitates such sharing and permits integration across multiple disciplines. This is not to suggest that omics should or could necessarily replace ology in any way. However, omics approaches offer the possibility of a new foundation for ologies to build on, by allowing for the testing of many hypotheses in parallel and providing a systems-level context for data interpretation that is necessary to further our understanding of brain function.