In the absence of predictive animal models of disease, companies are faced with making investments without certainty that they are on the right track, and that the tools will be available to carry a program through development, said Steven Hyman. Moreover, Doug Cole said some of the unique aspects of many central nervous system (CNS) disorders present additional challenges for drug discovery and development. For example, he noted that the inaccessibility of the nervous system, combined with the fact that many of the signs of dysfunction cannot be tied to a single element in the brain, make it difficult to study the basic mechanisms of CNS disorders. In addition, certain aspects of the human nervous system are not represented in virtually any other animals, making modeling difficult, said Cole.
The challenges associated with traditional animal models and the fact that the preclinical pipeline is currently failing to predict what will work in the clinic have contributed to a series of late-stage and very expensive failures, making it imperative to develop novel modeling systems, said Steve Finkbeiner. These systems include novel cellular models derived from induced pluripotent stem (iPS) cells, genetically engineered primate models, and computational models.
STEM CELLS AND ORGANOIDS
Cultures of human neurons provide an alternative approach to disease modeling. According to Lee Rubin, professor of stem cell and re-
generative biology at Harvard University, these models are valuable even for diseases where good animal models are available.1 For example, spinal muscular atrophy (SMA) is an early-onset monogenic disease caused by a mutation in the survival motor neuron 1 (SMN1) gene, which results in the degeneration of spinal motor neurons, muscle atrophy, paralysis, and death. The disease is more heterogeneous than initially recognized, in part because of multiple disease-modifying genes, including a duplicated copy of SMN1 called SMN2. Several SMA mouse models have been generated that recapitulate some aspects of human SMA (Edens et al., 2015). However, why motor neurons were selectively vulnerable in this disease remained unclear, said Rubin.
Rubin and colleagues generated iPS cells from SMA patients across a range of severities. These cell cultures show selective death of motor neurons and respond to drugs that have been clinically effective—but not to drugs that failed in human trials. Because these cells can be produced in very large numbers, they provide a rich source of material to study why motor neurons die. In addition, Rubin said that by studying these iPS cells and comparing those results to approximately 50 million medical records that included hundreds of children with SMA and their parents, researchers were able to predict aspects of the disease that had not yet been recognized by clinicians, especially phenotypes in parental carriers.
Finkbeiner agreed that there are many reasons to get excited about iPS models. Because the models derive from patients themselves, they provide a direct connection between the model and patient. He noted that in an informal survey of pharmaceutical company partners they found about half indicated they would be willing to proceed without in vivo efficacy if they had efficacy in iPS cells, in vivo toxicology was acceptable, and a good target engagement biomarker was available, he said.
iPS cells enable scientists to move beyond pathology and emphasize prediction, said Finkbeiner. For diseases without a clear genetic mutation, such as the predominant forms of Parkinson’s disease (PD) and Alzheimer’s disease (AD), taking cells from a patient may be the first step to creating a model with face validity, he added. This may enable the development of new technologies to be used in drug development for screening, patient stratification, prediction of efficacy, and personalized medicine. For those reasons and more, the Gladstone Institute of Neurological Disease now has more than 250 iPS cell lines for all the major
___________________
1 For additional information on the use of organotypic cultures, or minibrains, for toxicology testing see Chapter 7.
neurodegenerative diseases, including AD, PD, Huntington’s disease, frontotemporal dementia, and motor neuron diseases such as amyotrophic lateral sclerosis, as well as schizophrenia and autism, said Finkbeiner (Haston and Finkbeiner, 2016). They are using a variety of techniques, including transcriptomics, to look at transcriptional changes in patient cells; whole-genome analysis to find variants that correlate with clinical phenotype; and proteomics, epigenomics, and imaging to achieve deep phenotypic characterization of the cells. Rubin noted that this phenotyping approach, particularly for phenotypes that are robust and druggable, could also allow grouping of patients according to their likely response to a drug for drug trials. Refining these in vitro systems so that they are predictive of what goes on in vivo will be key, said Rubin.
Finkbeiner’s lab is developing cellular biosensors to monitor both structural and functional biological processes over time, enabling deep phenotyping of individual cells. For example, he described an array of about 270 multiplexed, optical electrophysiology biosensors to assess synapse structure and function, neurite extension and retraction, mitochondrial structure and trafficking, bioenergetics, proteostasis flux, autophagy, DNA damage and repair, and protein aggregation (Finkbeiner et al., 2015). Finkbeiner and colleagues are also trying to relate this to transcriptomic data and disease-related phenotypes from the same patient, such as survival, calcium signaling, glial phenotypes, electrophysiology changes, and gene profiling. They are also mining imaging data and using machine learning approaches to extract information and develop quantitative, dynamic, multidimensional measures of cellular health or sickness.
However, a number of obstacles remain to be addressed to realize the potential of iPS cells as drug discovery platforms, noted Finkbeiner. A major issue is heterogeneity: even small variations in the recipe used to induce iPS cell differentiation can produce substantial variation in the final product, leading to excess noise in the system. Even when a differentiation protocol is completed, there will still be cells at varying stages of maturity. Some of this variability is by design, said Finkbeiner; for example, to study the neuromuscular junction, one needs both motor neurons and skeletal muscle cells. However, tracking the biology of drug effects in these different cell types can be challenging.
Finkbeiner’s group is addressing some of the issues related to heterogeneity by developing high-throughput, single-cell analysis methods, which he thinks will make the iPS cell platform more useful for target
identification and drug discovery. For example, they have developed a robotic microscope that can be programmed to recognize where a plate is in a three-dimensional (3D) space so that it can register that position and return the plate to the same position at a future time point. This allows a single cell to be followed over time, similarly to how a single patient is followed over time in a drug trial. Finkbeiner said this system is about 100 to 1,000 times more sensitive than conventional endpoint analyses.
Recognizing that simple systems such as this may not work for more complex psychiatric diseases such as schizophrenia and autism, researchers at Harvard University and other institutions have developed techniques to grow neurons in three dimensions, creating two different culture systems called organoids and spheroids. Neuronal organoids are complex, brain-like structures that resemble early stages of the fetal brain, said Rubin. Long-term 3D cultures of human brain organoids develop into radially layered structures with many different cell types, not only different types of neurons, but perhaps even retinal cells and non-neuronal, and other cells. According to Rubin, this allows scientists to examine multiple processes and disorders. His Harvard University colleague Paola Arlotta, for example, is using electrophysiological techniques to study the formation of functional synapses and neural networks.
Compared to organoids, spheroids are structurally simpler, but more consistent and can be produced by the billions, said Rubin. They are made by taking uniform clusters of iPS cells and exposing them to different factors that induce neuronal differentiation. The induction factors chosen determine the types of cells that make up the spheroids. They are smaller and lack the distinct layers of organoids, but contain multiple types of neurons organized as clusters into geographically defined regions. As connections are made within the spheres over time, electrophysiological activity becomes more synchronous, reflecting development of networks.
Comparing spheroids from mutant and control cell lines has enabled scientists to identify and characterize disease-relevant phenotypes. For example, Rubin showed images of cortical spheroids from patients with autism that had a CHD8 mutation compared to isogenic cells without the mutation. The mutated cells showed a significant increase in size compared to the isogenic line.
Rubin’s team, in collaboration with Steven McCarroll, is also using dissociated spheroids plated in two dimensions as screening tools, for example, to screen for compounds that decrease the levels of comple-
ment C4 in cortical neurons (high levels of C4 are a risk factor for schizophrenia). New techniques are being developed to make the screens more viable, allow for comparisons among different kinds of neurons, and facilitate multiplexed labeling to identify many different targets and cell types in one screen. One of the projects includes embedding single spheres within an extracellular matrix to allow for more structured growth for use in high-throughput screens. Rubin said he expects the field to move in the direction of 3D production and 3D imaging.
Organoids and spheroids, while promising, are not likely to model many aspects of complex CNS disorders, said Stuart Hoffman, senior scientific advisor in the Office of Research and Development, Department of Veterans Affairs. For example, he noted that in traumatic brain injury and stroke, it would not be possible to model the systemic reaction to brain damage with organoids.
GENETICALLY ENGINEERED PRIMATE MODELS
Genetically engineered mouse models have transformed the study of disease biology, although their impact in the area of high brain function and related brain disorders has been limited, according to Guoping Feng. One of the major reasons, he said, is the vast difference in brain structure and function between rodents and humans. However, according to Feng, animal models are still critical for neuroscience research and drug development, and better animal models are possible. In particular, he advocated the use of nonhuman primate models given that, compared to rodent models, they are evolutionarily closer to humans and have a prefrontal cortex—the part of the brain that is dysfunctional in many neurological disorders (Kaiser and Feng, 2015). Old World primates such as macaques are phylogenetically closer to humans and thus more genetically similar than are New World monkeys such as marmosets. However, marmosets have the advantage of a short reproductive cycle, reach sexual maturity earlier, and give birth to twins or triplets twice per year, making it much easier to generate a large cohort for studies, said Feng.
At the Massachusetts Institute of Technology and the Broad Institute, Feng and colleagues have set out to establish a genetic engineering platform for brain disorder research and drug discovery, using genetically engineered marmosets and the Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR) technology genome editing system for targeted genetic mutations. The platform and its tools and resources will
be available as an open resource for academic and industry research nationwide, said Feng.
One project under way involves trying to generate an autism model by mutating a gene called SHANK3, which monogenetically causes severe autism spectrum disorders in about 1 percent of all patients with autism (Zhou et al., 2016). Feng’s team has successfully modified the gene in marmoset embryos with about a 50 percent targeting efficiency and is now trying to implant these embryos into live monkeys. In another collaboration with scientists in China, they have generated macaque monkeys with SHANK3 mutations, which should enable them to study how prefrontal-subcortical circuit defects develop over time, said Feng. They are also developing new behavioral testing and monitoring systems that assess vocal and social behaviors. Although animals will never behave like humans, Feng said that if a read-out, whether behavioral or neurophysiological, can be identified in the monkeys that reflects a defect in circuitry analogous to that seen in humans with autism, that readout could be used as a predictor of treatment efficacy.
Currently, Feng and colleagues are generating only monogenic disease models focused on prefrontal cortex developmental processes. However, he predicted that as more information about the genetics of complex disease becomes available and gene editing technology and other genetic tools improve, it will be possible to generate multiple mutations and/or multiple knock-ins in one injection per animal, map disease-relevant cell type–specific molecular changes, and identify cell type–specific targets for correcting circuit dysfunction.
COMPUTATIONAL QUANTITATIVE SYSTEMS PHARMACOLOGY MODELING OF BRAIN CIRCUITS
Computational modeling may be an additional tool to expedite progress in drug development, according to Hugo Geerts, chief scientist at In Silico Biosciences. The approach he advocated—quantitative systems pharmacology (QSP)—is based on lessons learned from other engineering-based industries. These lessons include using collective knowledge for advanced modeling and simulation in silico; embracing complexity in a quantitative way with circuit analysis and mathematical modeling; conducting failure analysis when something goes wrong, such as a failed clinical trial; and using advanced statistical techniques traditionally applied to physics and chemistry to analyze biological processes.
QSP integrates data from various modalities to create a computer model of the brain. These modalities include those discussed earlier: iPS cells, organoids, animal models, and human clinical trials. The current iteration of this disease model, described by Geerts, represents a relatively simple circuit with eight different types of neurons in a “closed cortico-striatal-thalamic-cortical basal ganglia loop of the dorsal motor circuit.” This model has been applied to study a number of pathways involved in schizophrenia, PD, and AD (Roberts et al., 2016). For example, the model enables simulation of the impact of different interventions on the circuit motor symptoms in PD, including the placebo effect, said Geerts.
Several studies have been completed demonstrating the value of this approach, he added. In one study, the model was used to predict the clinical outcome of an experimental antipsychotic drug for patients with schizophrenia, compared to placebo and an active control. The model correctly predicted the results from the actual clinical trial: The drug was less efficacious and had greater adverse effects than the active control (Geerts et al., 2012). If this modeling approach was used earlier, it could have prevented a large investment into the compound, said Geerts. In another study, the QSP model correctly predicted an unexpected worsening in clinical outcome of a Phase I study of a 5-HT4 (serotonin) partial agonist for the treatment of AD (Nicholas et al., 2013).
Geerts and colleagues have also built an in silico model linking amyloid modulation to cognition in AD. This model links biochemical and imaging data on the aggregation of amyloid-β monomers into oligomers and fibrils, and their deposition and removal by microglial cells, with a functional cognitive readout. Together with the Cohen Veterans Bioscience Foundation, they have also initiated a project modeling tau neurobiology, incorporating neuropathological, biomarker, and functional data to identify the best targets to engage, and how to engage them.
Geerts maintained that QSP can help answer critical preclinical and clinical questions in drug development, including those related to target identification, selection of a clinical candidate, demonstration of target engagement, determination of what constitutes a clinically meaningful outcome, dosing parameters, patient selection, and impact of genotypes and co-medications of clinical dose response.
ARE NOVEL MODELING APPROACHES CHANGING THE OUTLOOK FOR CNS DRUG DISCOVERY?
One question raised at the workshop was when, and if, these various modeling platforms will prove as robust as those coming from other fields such as cancer. Until they do, said John Reppas from the Neurotechnology Industry Organization, investment in nervous disorders will continue to lag behind these other areas. Several workshop participants agreed that competing with oncology and other areas will continue to be a challenge, but suggested that as the new models in development become more predictive and provide more solid connections between endophenotypes and patient outcomes, this gap may begin to narrow. Geerts also predicted that pressure from the market to find treatments for brain disorders will push companies in this direction.
The difficulties inherent in CNS drug development, including penetration of drugs into the brain, the long-term nature of many of these diseases, and the great degree of heterogeneity, have added to the reticence of pharmaceutical companies to invest in this area, said Rubin. However, he suggested that the field of neurology may achieve greater status as people recognize the importance of dealing with heterogeneity with more targeted approaches. Hyman suggested that another factor holding back progress in using these new models is slow uptake by academia. Finkbeiner agreed, noting that neuroscience research has been driven by pathology and behavioral dysfunction in mice, rather than by clinical effects. He further suggested that this results from a lack of sufficient interaction between clinical and preclinical scientists.
Frances Jensen suggested that applying these models to the study of orphan diseases, which are relatively pure disorders with more homogeneous endophenotypes, might provide the best opportunity to advance these new models, in part because studies can be much smaller and enroll more homogeneous participants. She also suggested that working with non-animal and animal models in tandem might expedite discovery. Rubin and Feng agreed that the simplest way to make progress is to start with the simplest forms of disease, which would be the monogenic or single mutation driven forms.
This page intentionally left blank.
