Understanding the etiology and pathophysiology of complex human neuropsychiatric disorders will require modeling disease in systems ranging from humans to stem cells. While this will enable insights appropriate to each model system, the development of novel therapeutics ultimately requires that these models converge at tractable interfaces. In the previous discussion, we laid out a roadmap for developing technologies to achieve this convergence. Here we describe these model systems in greater detail, in an effort to highlight the appropriateness of each to answering a particular question. Rather than stratifying models into “top-down” and “bottom-up” heuristics, we highlight opportunities to interface between systems, as well as the iterative nature of this process.
Role of Human Genetics for Developing Models
Genetic technology has exploded over the last several years, and soon we will have the ability to obtain full genetic information for each person. Gene manipulation in model systems can provide a platform for investigations of the consequences of genetic mutations on neurochemistry, plasticity, circuits, and behavior in a manner that cannot be done in humans. Advances in human genetics have facilitated tremendous progress toward identifying pathogenic or risk genes for several neuropsychiatric disorders, and this information has been applied to model systems. The genetic variations in autism spectrum disorders (ASD), for example, are quite diverse, encompassing rare and common alleles, sequence and structural variants, and de novo as well as transmitted mutations. Moreover, in addition to gene discovery in what is often referred to as “common idiopathic” forms of these disorders, there are also multiple examples of established “syndromes” caused by monogenic mutations, including fragile X and Rett syndromes, highly penetrant copy number variants, including 22q11 deletion syndrome (also known as velocardiofacial syndrome), and rare recessive syndromes. It is important to note that the clinical presentation of these genetic disorders can vary widely and that distinctions between syndromic and idiopathic categories are largely historical. It also remains to be determined whether the underlying pathophysiology of the relevant psychiatric manifestations differs as a result of divergent transmission modes.
The characterization of differences in penetrance, effect size, and likelihood that a given risk gene may be implicated in more than one disease state will inform both the selection of genes for modeling in nonhuman systems and the strategies that will need to be applied. For example, mutations carrying relatively large effects and mapping to coding segments of the genome may be suitable to be immediately modeled in an animal system, whereas noncoding single nucleotide polymorphisms found in association with a condition may require additional fine mapping and systems biological analysis before pursuing in vivo model experiments.
Given the finding of a very high level of locus heterogeneity for both schizophrenia and ASD, efforts are underway to use a variety of approaches to organize disparate genes into more biologically meaningful groups. These include protein-protein interaction analysis and gene ontology approaches that seek to identify molecular pathways of interest. In addition, several recent studies have leveraged genome-wide expression data from typically developing brain (in humans and other species) to gain traction on the question of when and where specific risk genes might show pathophysiological convergence. Studies in both autism and schizophrenia have implicated mid-fetal cortical development (Gulsuner et al. 2013; Parikshak et al. 2013; Willsey et al. 2013). One recent study mapped an ASD-associated co-expression network based on only high-confidence ASD genes to layer 5/6 projection neurons in mid-fetal development (Willsey et al. 2013). All of these “systems biological” approaches are aimed at prioritizing key parameters in model systems studies, including which cell types, signal pathways, and circuits may be of particular interest.
Etiological and symptom heterogeneity has led some to propose that it might be more useful to consider risk genes as distinct clinical entities (i.e., the existence of “autisms” rather than a single “autism”). Others have suggested an analogy to Alzheimer disease, which is considered a single clinical entity based on pathology, despite the fact that only 10% of the cases have been linked to a specific genetic cause. The remaining 90% likely result from a set of changes that take place during aging which, presumably, compromises the ability of neurons to process proteins correctly. Thus, it is taken for granted that a single therapeutic, when it is finally found, will be effective on all patients with Alzheimer disease. This makes a huge difference in the search for therapeutics and explains why knowing the number of classes (in the treatment sense) of diseases is so essential. In the long run, the complexity of psychiatric disorders, both in terms of their waxing and waning symptoms and genetic etiology, suggests that a spectrum of phenotypic and etiologic factors will need to be considered when defining individual disease classes for the development of tractable models and treatment studies.
The distinction between defining a “disease,” based on symptoms or pathology, or “diseases,” based on genes, could influence selection of models, since these definitions may lead to different predictions about the level (i.e., biochemical, synaptic, circuit, behavior) of pathogenic convergence across etiologies. For example, at one extreme, it could be argued that because the “diseases” are primarily defined by the genetic etiology, the key property required of a model is construct validity (e.g., genetic homology) and the ability to therapeutically target convergent biochemical pathogenic mechanisms. Accordingly, behavioral readouts would be secondary, and a demonstration of face validity for human symptoms need not be prioritized. This view is particularly amenable to using iPS or IN cell-based models to subclassify complex disorders. At the other extreme, one could argue that because the “disease” is primarily defined by the symptoms and pathology, models which achieve the best face validity (e.g., behavioral changes that recapitulate human symptoms) will provide the greatest opportunity for understanding pathogenesis. Accordingly, therapeutic strategies would focus on correcting conserved circuit and behavioral abnormalities, independent of the heterogeneity of genetic disruptions that produced them. This view is particularly useful when the etiology of the disorder is unknown and may require development of model organisms (e.g., marmosets, discussed below) amenable to interrogation of more complex behaviors.
Between these two extremes lies the view that since some of the identified risk genes encode synaptic proteins (e.g., receptor, downstream signaling, and adhesion molecules) with known functions that are largely conserved across species, understanding pathogenesis at the level of synapses will yield the greatest short-term opportunity for therapeutic intervention. The term “synapsopathy” was first coined to describe this view as it applies to autism (Bear et al. 2008; Dölen and Bear 2009) and to distinguish this disorder from those brain diseases where the primary pathology is localized to a specific brain region (e.g., substantia nigra in PD). Despite enthusiasm for the idea, an inherent challenge is that the brain is remarkably heterogeneous in its cellular composition. The wide variety of cell types might thus be the basis for genetic “pathoclisis,” a selective vulnerability of subsets of neurons to the effects of mutation. Indeed, recent studies of Huntington disease and autism raise the possibility that genetic pathoclisis will be an important pathogenic mechanism, with profound implications for developing therapeutics (Dölen and Bear 2009; Paul et al. 2014). Moreover, synapsopathy and pathoclisis need not be mutually exclusive; together they might account for symptom heterogeneity despite overlapping clinical presentation. Understanding the relative contribution of each of these mechanisms to the pathogenesis of disease will likely inform decisions about the suitability of modeling synaptic disruptions at the cellular or circuit level, so as to guide future development of tractable interfaces. For example, synapsopathic features might be particularly amenable to ERPs, MMN, gamma oscillations, and ERG measurements (discussed above), whereas TMS, tCS, and brain imaging (see previous discussion) of specific circuits might be more appropriate for interrogating disease symptoms that result from pathoclisis. Of course, a risk of focusing solely on genes, the functions of which are currently somewhat familiar, is that many of the genome-wide significant loci associated with disorders, such as schizophrenia, are noncoding (Schizophrenia Working Group of the Psychiatric Genomics Consortium 2014). Thus we may end up ignoring important pathogenic mechanisms which, in the long run, need to be understood.
Model Systems: Mice
To date, the modeling of identified risk genes has primarily been conducted in mice using constitutive or conditional transgenic systems. This model organism offers the opportunity to interrogate pathogenesis at the level of biochemical, synaptic, circuit, and behavioral mechanisms. Moreover, the availability of a number of other molecular genetic tools—such as BAC-Cre-recombinase driver and BAC-EGFP reporter lines, viral-mediated gene transfer (Callaway 2008; Grimm et al. 2008; Luo et al. 2008; Marie and Malenka 2006; Neve et al. 2005; Salinas et al. 2010), RNA interference (Morris and Mattick 2014), optogenetics (Boyden et al. 2005; Lima and Miesenböck 2005; Lin et al. 2009), genetically encoded voltage (Cao et al. 2013; Dimitrov et al. 2007; Siegel and Isacoff 1997) and calcium sensors (Chen et al. 2014a; Wang et al. 2004) and pharmacosynthetics (Dong et al. 2010; Lee et al. 2013)—has made it possible to address how complex, goal-directed behaviors occur from organized networks of neurons. Many of the plasmids for these molecular manipulations are available from Addgene,3 a nonprofit, public repository. An in-depth review of the molecular toolkit is beyond the scope of this discussion. Here we will present an overview, with a focus on resources, advantages, and limitations, followed by a handful of examples that have demonstrated the power of these approaches to help understand the pathophysiology of brain disease.
The Gene Expression of the Nervous System Atlas project,4 in collaboration with the Intramural Program of the National Institute of Mental Health, offers transgenic BAC-EGFP reporter and BAC-Cre recombinase driver lines which allow for cell-specific gene manipulations in the mouse CNS. The aim of this project is to provide the scientific community with reporter and Cre driver lines that will target selected neuronal or glial populations in the brain and spinal cord. An important caveat is that cell type specificity must be confirmed for each line, since BAC transgenics are not generated by targeted mutation; furthermore, cell type specificity in adult brain need not be reflective of expression patterns during development. In addition to this resource, a vast array of constitutive and conditional knockout and knockin mice are available through commercial repositories, such as Jackson Labs.5 When transgenic mice, particularly conditional knockouts, for the gene of interest are not available, RNA interference is a viable alternative at most synapses. Because various forms of RNA interference are encoded by relatively short sequences, it is feasible, and indeed often necessary, to control for off-target effects using molecular replacement strategies (Alvarez et al. 2006; Jurado et al. 2014).
The introduction of exogenous DNA into neurons by viral transfection has become a standard technique in molecular neurobiology, particularly as our understanding of viral tropism, life cycle, transport, and toxicity has led to the development of increasingly sophisticated recombinant strategies. An illustrated comparison of the viruses commonly used for neuroscience () highlights the relevant features of each (for in-depth reviewa, see Callaway 2008; Grimm et al. 2008; Luo et al. 2008; Marie and Malenka 2006; Neve et al. 2005; Salinas et al. 2010). This approach can be a powerful adjunct or alternative to the BAC driver lines, particularly when synapse-specific or developmentally restricted expression is required. Nevertheless, in vivo viral infection requires labor-intensive stereotaxic injection, except in cases where the vector can cross the blood-brain barrier (e.g., AAV-9; Foust et al. 2009). Finally, this approach is also a promising method under development for therapeutic gene delivery in human patients, with the caveat that viral immunogenicity is often species specific, so not all vectors used in mice are appropriate for humans (Mingozzi and High 2011).
Comparison of commonly used viruses in neuroscience.
The ability to optically stimulate molecularly specified neurons (termed “optogenetics”) has transformed neuroscience (Boyden et al. 2005; Lima and Miesenböck 2005; Lin et al. 2009). Stimulation of neurons with metal or glass electrodes, while still a mainstay, can only resolve individual input pathways when these are arranged in such a way that they can be physically segregated (e.g., Schafer collateral versus perforant path inputs to the CA1 region of the hippocampus). However, this anatomical arrangement is exceptional: in the vast majority of brain regions, inputs are intermingled, and thus stimulation of known inputs to a specific cell type is frequently impossible. Furthermore, it is increasingly apparent that many of the assumptions concerning input and output homogeneity, even in well-circumscribed pathways, do not hold (e.g., co-release of transmitters, novel parallel pathways) (Graves et al. 2012; Tritsch et al. 2012; Varga et al. 2009). In addition, optogenetics enables convergence of cellular and behavioral studies since molecularly isolated inputs can be stimulated to both record electrical responses in specific neurons as well as to interrogate the behavioral consequences of evoked responses. This is particularly important in brain regions where the “receptive fields” of the neurons in question are internal states and are not reliably evoked by direct manipulation of the sensory or motor environment. Despite the remarkable advances enabled by the implementation of optogenetics, its current use is restricted by the speed with which the optogenetic proteins can be delivered to membranes. It often takes many weeks or months for adequate expression of optogenetic proteins in axon terminals, thus limiting experiments to late stages in development. Future iterations will likely improve subcellular targeting (e.g., dendritic versus axonal membranes), channel properties (for better temporal fidelity at higher stimulation frequencies), opsin properties (e.g., for better resolution of distinct activation wavelengths), and toxicity due to overexpression.
Genetically encoded voltage (Cao et al. 2013; Dimitrov et al. 2007; Siegel and Isacoff 1997) and calcium (Chen et al. 2014a; Wang et al. 2004) sensors represent a parallel set of emerging technologies, which will do for the recording electrode what optogenetics has done for the stimulating electrode: allow sub-(action potential)-threshold recordings in molecularly specified neurons, both ex vivo and in vivo. Currently, voltage sensors are limited to in vitro approaches in mammalian systems, and the use of genetically encoded calcium indicators in vivo is restricted to microscopically accessible brain regions (e.g., the somatosensory cortex). However, ongoing development of brighter fluorophores and endoscopic techniques may likely overcome these hurdles in the near future (Deisseroth and Schnitzer 2013; St-Pierre et al. 2014).
Finally, the G protein-coupled receptor superfamily represents the canonical targets of more than 30% of clinically available pharmacotherapies (across all indications), largely because these molecules are readily druggable, act as modulators of nearly every known physiological process, and have been implicated in the etiology or pathogenesis of numerous disorders. Despite this profile, interrogating the neuronal function of these receptors is complicated by the fact that a single endogenous ligand typically binds multiple receptors, can modulate different signaling pathways through a single G protein-coupled receptor, and are expressed across mixed populations of cells within a given brain region. The advent of second-generation molecularly encoded, pharmacosynthetics, namely the designer receptors exclusively activated by designer drugs, has made deconstructing this functional complexity a tenable goal (Dong et al. 2010; Lee et al. 2013).
While many of these techniques can be used independently in other model organisms, the power of modeling in mice is the opportunity to use a combinatorial approach. For example, AAV (adeno-associated virus)-mediated expression of channelrhodopsin-2 in the striatum of BAC-transgenic Cre driver lines under control of regulatory elements for the dopamine D1 or D2 receptor has enabled direct activation of basal ganglia circuitry implicated in PD. These studies have validated the long-standing hypothesis that two parallel pathways exert bidirectional control over motor behavior and, furthermore, shown that in a mouse model of PD, direct pathway activation rescues movement phenotypes (Kravitz et al. 2010). Another approach has combined viral-mediated expression of pharmacosynthetics with Cre-mediated targeting of neurons which receive inputs from agouti-related protein-expressing neurons in the arcuate nucleus, to interrogate G protein-mediated feeding behaviors relevant to the pathogenesis of insatiability in Prader-Willi syndrome (Atasoy et al. 2012). Others have capitalized on the ability of rabies virus and AAV to infect selectively pre- and postsynaptic neurons, respectively, combining this technology with conditional knockout and knockin reporter mouse lines. Such approaches, for example, allowed the characterization of oxytocin and serotonin receptor-containing inputs to the nucleus accumbens and implicated a novel circuit in the pathogenesis of social deficits seen in autism (Dölen et al. 2013).
Despite the remarkable technological opportunities available in mice, determining the suitability of this organism for modeling disease requires the consideration of genetic similarities and differences between mouse and humans. Sequencing of the mouse and human genome has revealed that in protein-coding regions, there is 97% alignment between orthologs, and in 1:1 orthologs 85% DNA sequence identity. These genes are the most likely to have maintained ancestral function in both species and are therefore most appropriately targeted as disease models (Church et al. 2009). Despite this remarkable homology in protein-coding regions, only 33% of whole genomes align, with 60% DNA sequence identity in aligned regions. This discrepancy is due, in part, to structural variants and segmental duplications (i.e., evolutionarily young and rapidly evolving parts of the genome) in gene regulatory regions, as well as to transcribed microRNA and noncoding RNA sequences (Church et al. 2009). Modeling disease that affects these nonhomologous genomic regions will require alternate systems. For example, patient-derived induced pluripotent stem cells may be appropriate for modeling schizophrenia, since most identified schizophrenia risk variants are regulatory (e.g., cis trans transcription regulators) affecting the time, place, and rate of gene expression. Currently, the success of such an approach will depend on the degree to which the pathogenesis of disease is cell autonomous, although chimeric and organoid systems may be able to overcome this limitation in the future (Anderson and Vanderhaeghen 2014).
In addition to these genetic differences, anatomical differences (e.g., a small medial prefrontal cortex) as well as a limited repertoire of behavioral assays for interrogating complex cognitive function (e.g., episodic learning, language) limit the use of mice for modeling all features of complex neuropsychiatric disease. New genome-editing technologies, such as clustered, regularly interspaced, short palindromic repeats, and associated cas genes, have made it possible to make precise genetic manipulations in many other organisms, including primates (Ran et al. 2013). Considerations such as size, cost, generation time, and ability to breed in captivity will importantly influence the selection of organism for genetic modeling.
The common marmoset, Callithrix jacchus, a New World monkey, has significant advantages from a genetic perspective. It is small in size (~400 g), reaches sexual maturity at 12 months, and breeds rapidly in captivity, typically producing two pairs of fraternal twins per year. The neuroanatomy of common marmoset is well described. Like macaques, but unlike rodents, marmosets have a well-developed prefrontal cortex, a critical region for many cognitive functions that are impaired in human psychiatric disorders. Furthermore, marmosets are very social and communicative and can perform some higher cognitive tasks developed for macaque monkeys and humans. The marmoset genome has recently been sequenced, laying the necessary groundwork for genetic manipulations. Moreover, the optogenetic tools described above are also being developed for use in primates (Diester et al. 2011; Galvan et al. 2012).
A number of other organisms offer potential as model systems for understanding the pathogenesis of complex neuropsychiatric disease. For example, SAGE labs6 has developed several rat models of PD, Alzheimer disease, schizophrenia, and autism, as well Cre lines for specific expression of floxed constructs in dopaminergic neurons. The larger size of rats makes them more amenable to in vivo recording as well as an extensive set of well-characterized behavioral assays. Other species of potential interest for future genetic model development are the Etruscan shrew (active touch; Brecht et al. 2011), the prairie vole (pair bonding; Barrett et al. 2013), and the scrub jay (episodic memory; Raby et al. 2007).
