A human pluripotent stem cell platform for assessing developmental neural toxicity screening

A lack of affordable and effective testing and screening procedures mean surprisingly little is known about the health hazards of many of the tens of thousands of chemicals in use in the world today. The recent rise in the number of children affected by neurological disorders such as autism has stirred valuable debate about the role chemicals play in our daily life, highlighting the need for improved methods of assessing chemicals for developmental neural toxicity. Current methods of testing chemicals for developmental neural toxicity include animal testing with rats or mice and in vitro testing using cultured primary cells or cell lines. Here, we review the current state of neural toxicity screening, analyze the limitations of these methods and, under the National Institutes of Health's new Microphysiological Systems initiative, describe a human pluripotent stem cell-based platform for developmental neural toxicity screens.

testing between diff erent animal species (typically rat and rabbit) is only about 60% [3], and there is no evidence that the concordance between either species and human toxicity is any better. A good example of this is the chemical thalidomide, which causes birth defects in humans but has little eff ect in rats [9]. In addition, because of species-specifi c diff erences, subtle cognitive changes and human conditions with no known counterparts in animal behavior (for example, autism) are particularly diffi cult to model in animals [10]. Finally, the cost of multigeneration animal studies is considerable (>$1 million/study). Animal testing is projected to account for a stagger ing 70% of the European Union's 2006 Registra tion, Evaluation, Authorization, and Restriction of Chemicals cost of evaluating a new chemical, and to consume an average of 3,200 rats per chemical [4].
Performing chemical screens on in vitro human models is another method of developmental neural toxicity testing. Currently, however, no in vitro human models are in widespread use for assessing developmental neurotoxicity. A few studies have examined the eff ects of toxins on human fetal-derived neurospheres. Human central nervous system neurospheres of fetal origin include stem cells, neurons, and astrocytes that self-assemble into structures which recapitulate some early neural developmental events [11], providing rare access to the developing human neural system. However, these neurospheres lack microglia and vascular cells and therefore do not represent the entire repertoire of cell-cell interactions comprising the brain. Th ey are also hard to obtain in large quantities, making large-scale chemical screens impractical. Furthermore, individual diff erences between donors may introduce variations between screenings, potentially complicating downstream data analysis.
Human embryonic stem (ES) cells and induced pluripotent stem (iPS) cells, which have unlimited proliferative potential, provide a more practical option for building in vitro models [12][13][14]. Th e effi cient derivation of neural cells from human pluripotent stem cells is now possible, paving the way for recent reports of developmental neural toxicity testing on human ES cell-derived neurons [15][16][17][18]. However, the number of toxins examined in these studies has been extremely small (typically one to four) and the types of cells included in the screen have been limited [17,18]. Th e value of these model systems to actually predict neural toxicity in wider chemical screens remains to be examined.

An innovative platform for developmental neural toxicity screen
Under the NIH's Microphysiological Systems initiative, we are building a platform for developmental neural toxicity testing on a physiologically relevant human model.
A combination of stem cell biology, tissue engineering and bioinformatics, this platform is capable of meeting the needs of large-scale chemical toxicity screens ( Figure 1).
A central characteristic of our platform is that a remark able degree of self-assembly, diff erentiation, and maturation occurs if appropriately specifi ed precursor cells are brought together in the right environment. Th e most impressive experimental example of self-assembly is the formation of teratomas by ES cells and iPS cells. If allowed suffi cient time to develop, these teratomas form well-developed, highly stratifi ed neural structures that closely recapitulate early brain development [19], and form other advanced structures requiring complex inductive interactions between germ layers, including gut, teeth, and hair [12]. Th e diff erentiation and selfassembly of polarized cortical tissues, including ventricular, early and late cortical plate, and Cajal-Retzius zones, has also been demonstrated previously from the in vitro culture of ES cell-derived aggregates, a signifi cant fi nding for our platform [20].
To create these physiologically relevant three-dimensional structures, we embed human ES/iPS cell-derived endothelial cells, pericytes, and primitive macrophages (microglial precursors) into tunable poly(ethylene glycol) hydrogels displaying specifi c peptide motifs that promote capillary network formation (Figures 1a and 2). We then overlay this mesenchymal layer with neural and glial precursors to mimic in vivo cephalic mesenchyme-neural epithelial interactions, both to promote the formation of the polarized layers of the cerebral cortex and to allow the formation of endothelial networks with blood-brain barrier properties (Figure 1a). Poly(ethylene glycol) hydrogels have found widespread use in three-dimensional cell culture due to their ease of processing and biocompatibility [21]. Cells can be readily encapsulated within these gels using photo-polymerization [22][23][24] (Figure 2a), providing a simple mechanism to generate hydrogel arrays with well-defi ned local spots directly amenable to high-throughput screens. A key prerequisite of this approach is a consistent and scalable source of correctly specifi ed precursors. Recent progress in deriving neural, glial, and vascular precursors from human pluripotent stem (ES and iPS) cells [25][26][27][28][29] provides this scalable source for the fi rst time.
Th e nature of developmental timing poses perhaps the most signifi cant challenge to this project. Th e timing of human ES/iPS cell diff erentiation largely recapitulates the timing of normal human fetal development. Although it is possible to set up cultures of diff erentiating human cells that interact for 9 months or longer, such culture systems would not be practical for high-throughput drug screens. Our strategy, then, is to prediff erentiate the early precursors of the components of the cerebral cortex in large, defi ned batches, cryopreserve them, and later combine them into three-dimensional hydrogel assemblies, allowing those assemblies to interact and further mature for a more limited time during drug exposure.

Platform readout
Optimized screen readouts and machine learning tools are also important components of this platform. Our basic premise is that toxicants change the basic physiology of cells, and that these changes directly or indirectly cause changes in gene expression profi les that can, in turn, be used to classify specifi c classes of toxicants. Th e three-dimensional neural/glial/vascular assemblies, which initially consist of seven cell types, are complex; with further maturation, additional cell types or subtypes may also emerge. Because the specifi c targets of a novel toxic agent will be unknown and could be as diverse as the ensemble of molecules contained in these assemblies, readouts must refl ect this complexity and have a substantial dynamic range, as minor alterations in the assemblies' cellular subcompartments may refl ect relevant in vivo toxicity.
Importantly, the recent dramatic fall in the cost of high-throughput sequencing makes a readout based on expression profi ling very desirable for drug screening purposes, as RNA-Seq allows the monitoring of complex samples with an excellent dynamic range and at a reasonable price per sample. Th is desirability of using gene expression profi ling to better understand toxic responses is shown in recent reports using ES cell-based and other cellular models [30][31][32][33].
Th e platform's fi nal component, a machine learning algorithm, uses the gene expression profi les of cells exposed to known developmental neural toxins to predict The neural vascular assembly from (a) will be exposed to a training drug set. The gene expression profi les from the training set will be used to establish a drug toxicity prediction model using a machine learning algorithm. (c) The model established in (b) can be used to predict the toxicity of an unknown chemical.
the neural toxicity of chemicals without previous toxicity information (Figure 1b,c).

Conclusion
In short, we combine the developmental potential of human pluripotent stem cells, the modular nature of the tunable hydrogels, and the discriminatory power of machine learning tools to create a highly sensitive model suitable for large-scale predictive developmental toxicology screens (Figure 1). Th is platform, and other NIH Microphysiological Systems, should provide a better understanding of chemical impacts on human health.
Abbreviations ES, embryonic stem; iPS, induced pluripotent stem; NIH, National Institutes of Health.

Competing interests
JAT is a founder, stockowner, consultant and board member of Cellular Dynamics International (CDI), and serves as scientifi c advisor to and has fi nancial interests in Tactics II Stem Cell Ventures. The remaining authors declare that they have no competing interests.