Novel Technologies and an Overall Strategy to Allow Hazard Assessment and Risk Prediction of Chemicals, Cosmetics, and Drugs with Animal-free Methods

Several alternative methods to replace animal experiments have been accepted by legal bodies. An even larger number of tests are under development or already in use for non-regulatory applications or for the generation of information stored in proprietary knowledge bases. The next step for the use of the different in vitro methods is their combination into integrated testing strategies (ITS) to get closer to the overall goal of predictive " in vitro-based risk evaluation processes. " We introduce here a conceptual framework as the basis for future ITS and their use for risk evaluation without animal experiments. The framework allows incorporation of both individual tests and already integrated approaches. Illustrative examples for elements to be incorporated are drawn from the session " Innovative technologies " at the 8 th World Congress on Alternatives and Animal Use in the Life Sciences, held in Montreal, 2011. For instance, LUHMES cells (conditionally immortalized human neurons) were presented as an example for a 2D cell system. The novel 3D platform developed by InSphero was chosen as an example for the design and use of scaffold-free, organotypic microtissues. The identification of critical pathways of toxicity (PoT) may be facilitated by approaches exemplified by the MatTek 3D model for human epithelial tissues with engineered toxicological reporter functions. The important role of in silico methods and of modeling based on various pre-existing data is demonstrated by Altamira's comprehensive approach to predicting a molecule's potential for skin irritancy. A final example demonstrates how natural variation in human genetics may be overcome using data analytic (pattern recognition) techniques borrowed from computer science and statistics. The overall hazard and risk assessment strategy integrating these different examples has been compiled in a graphical work flow. tice and solid toxicological experience with new concepts and innovative technologies (van Thriel et al., 2012). Traditional approaches are rooted mainly in the fields of biology , medicine, and veterinary sciences (Leist et al., 2012). New technologies are emerging from many additional fields (Leist et al., 2008a,b). These include computer information sciences, statistics, mechanical engineering, molecular biology , drug discovery, and many others. Many technologies, approaches, and assays have been described individually. Recently , more integrated strategies for using combinations of in vitro data and in silico models to predict drug and chemical


Introduction
Current toxicological risk assessment to ensure human health or the safety of the environment is still based predominantly on animal studies.Besides ethical considerations, there are also scientific reasons to stop the use of animals and to switch to scientifically validated modern test systems, which may provide a deeper insight into the mechanisms of toxicity.
As science and technology continuously move ahead, and new innovations appear on the market daily, we suggest taking advantage of this situation by combining good laboratory prac-Erschienen in : Alternatives to animal experimentation : ALTEX ;29 (2012), 4. -S. 373-388 Konstanzer Online-Publikations-System (KOPS) URL: http://nbn-resolving.de/urn:nbn:de:bsz:352-271940 safety have emerged as well.One of the most comprehensive strategies to make toxicology largely independent of animal studies is "Toxicity testing in the 21st century: a vision and a strategy," as put forward by the National Research Council (NRC, 2007).This has triggered the design of new concepts and the setup of research programs dedicated to the implementation of this vision (Hengstler et al., 2012).Some of these programs, such as ToxCast, focus mainly on high-throughput and use mostly technology and methods suitable for this purpose as building blocks (Kavlock et al., 2012;Rotroff et al., 2012).For example, a high-throughput risk assessment (HTRA) framework was published last year (Judson et al., 2011).A different approach makes use of relatively complex models that have been established primarily with the purpose of replacing defined animal experiments.
The approach taken here is to suggest a comprehensive theoretical strategy to undertake risk assessment on the basis of available alternative test systems of differing complexity and with varying throughput.This strategy may be used as a framework to look for, identify, and employ new technologies.It may also be used to incorporate such diverse technologies as high-throughput screening of simple biochemical endpoints and the evaluation of complex functional endpoints in 3D engineered tissue.The different available and newly emerging methods would then support individual steps outlined by the strategy and form the basis for translating the theoretical framework into practice.Such a risk assessment scheme is outlined here, and a heterogeneous set of new technologies was chosen as an example of how to apply the new approaches within this overall framework.

In vitro based risk evaluation approach
The development of suitable in vitro methods for risk assessment is an ongoing effort, and many promising methods are available already.We suggest a general scheme, which can then be filled in with suitable methods according to the individual requirements (Fig. 1).The basic idea of the scheme was also discussed in the context of biomarker usage (Blaauboer et al., 2012).
Potential starting steps for compound evaluation are based on literature review, in silico as well as purely physicochemical and biochemical methods to define the compound in question, its intracellular distribution, and its potential metabolites.These steps would be followed by the application of a battery of methods to define the biokinetic behavior of the compound and metabolites in in vitro systems and to obtain essential data for physiology-based pharmacokinetic modeling (PBPK).Some of this information would be used immediately for subsequent steps, while other information (exposure information) would be used later for the final risk assessment.
The core of the strategy is the concept that in silico evaluation and thorough literature research go hand in hand with an in vitro-based hazard assessment.These are the cornerstones of any type of "in vitro risk assessment." In principle, any point of the circle may be used as starting point, as reconsideration and re-iteration is possible and necessary to ensure that all data and aspects are taken into account.If new compounds are to be evaluated, it might make sense to start with an exposure assessment in order to avoid testing of concentrations that are unlikely to occur in real life."Exposure Assessment is the process of estimating or measuring the magnitude, frequency and duration of exposure to an agent, along with the number and characteristics of the population exposed.Ideally, it describes the sources, pathways, routes, and the uncertainties in the assessment" (IPCS, 2004).Risk assessment would, for instance, be handled differently if the compound under investigation is used only as intermediate in a closed chemical production process or if it is part of a cosmetic product, or if it is an environmental chemical known to enter the food chain.Data on exposure may be scarce, of low reliability, or not available at all, especially for environmental agents.Under such circumstances, the scheme suggests moving to subsequent steps and evaluating compound hazard first.For the risk assessment, exposure would then need to be estimated roughly, classified on the basis of use categories and on the knowledge of environmental fate (e.g., transport models, persistence, bioaccumulation, and transformation).In the absence of all information, risk classification would need to be heavily based on hazard information and then adapted as more exposure information becomes available.
A thorough literature review 1 is usually the basis of a successful evaluation.Defining a search strategy is necessary to ensure coverage of the topic, as the "state-of-the-art" should be the starting point.In this sense, the term "literature" is meant to cover all kinds of databases that are accessible.The literature review would initially focus on the compound, but in further iterations it also would cover the methods used for hazard assessment, as well as other technologies used for evaluation of the compound.In addition, experiments that did not work and information on failed tests or compounds might be very helpful.Unfortunately, this information is seldom freely available.
Due to progress in science and technology, in silico modeling becomes more and more a prerequisite in the risk assessment field, and it was specified by Hartung and Hoffmann in 2009 as "anything that we can do with a computer in toxicology" (Hartung and Hoffmann, 2009).These days, it is difficult to define "in silico toxicology" exactly, as in silico components are present in practically every area of risk assessment (Raunio, 2011), and non-testing data can be generated by several approaches, including: grouping approaches, which consist of read-across and chemical category formation, structure-activity relationship (SAR) and quantitative SAR (QSAR) 2 .A structural physicochemical reactivity characterization of a compound is currently done routinely (Valerio, 2009(Valerio, , 2011)).
Another very important point is the determination of the biokinetic behavior of the compound in all test systems used.This comprises information on the real free concentration in 1 http //www ctu edu vn/guidelines/scientific/scientific/2 1howliteraturesearch htm 2 http //echa europa eu/web/guest/guidance documents/guidance on information requirements and chemical safety assessment tal system, similar information is necessary on the different metabolites.Such background information on the chemical facilitates identification of tests and conditions to set-up an in vitro effects battery.
The in vitro effects battery could include simple model organisms such as drosophila or zebra fish (Sipes et al., 2011b;Padilla et al., 2012), and it would include assays for functional changes, to provide broad omics information (e.g., metabolomics, transcriptomics), and to define biochemical targets of the chemicals.In addition to the mostly correlative methods, the cell culture medium, and ideally also in different cellular compartments.The latter often will require information on transport processes of the compound across different membranes.To perform an in vitro-in vivo extrapolation (IVIVE), the real concentration of a compound has to be determined (Coecke et al., 2012).Some compounds may bind to plastic or to proteins or they may evaporate; others react quickly or become metabolized.Therefore, the freely available concentration is not necessarily identical to the nominal concentration.In case of metabolism of the compound within the experimen-

Fig. 1: In vitro-based risk evaluation approach
The overa strategy re es on two essent a steps: F rst, mportant background nformat on on the compound and ts in vitro hazard s assessed (c rc e of ye ow boxes).Then, an in vitro-in vivo extrapo at on (IVIVE) s used to pred ct re evant human doses.These are then re ated to d fferent exposure scenar os, tak ng a so d fferent human subpopu at ons nto account to arr ve fina y at a r sk eva uat on (upper part of the scheme).The part of the strategy focus ng on n v tro hazard assessment wou d start w th the dent ficat on of appropr ate tests and cond t ons, based on the nformat on ga ned from the n t a steps.An in vitro test battery w th var ous 2D and 3D mode s (see sect ons 3-5) wou d be used to obta n deta ed concentrat on-response nformat on to arr ve at a po nt of departure (POD) for IVIVE.In most cases, steps of the cyc e may need to be rev s ted n an terat ve process of opt m zat on.A steps wou d y e d nformat on to databases, s mu at ons, and new software so ut ons, and vice versa.Bes des these genera nteract ons, the spec fic hazard eva uat on of a compound wou d nvo ve a process of mode ng and systems b o ogy approach to use a the data generated for the dent ficat on/ mapp ng of key pathways of tox c ty (PoT) and for defin t on of thresho ds of the r act vat on.A comb nat on of such steps s demonstrated n sect on 6.Once a defined POD (= cr t ca concentrat on of a compound to tr gger hazardous effects) has been der ved, IVIVE w he p to define the hazardous dose n humans.Th s may be affected, e.g., by genet c var ab ty (as out ned n sect on 7).Comb nat on of human hazard nformat on and exposure data w fina y support the r sk eva uat on process.
encephalic neurons from fetal material from the university of Lund."This cell system allows measurement of functional endpoints and demonstrates the usefulness of image-based high-content screening in human cells.The cell system can be used for disease modeling (Lotharius et al., 2005) and also for examining toxicity to developed neurons (Schildknecht et al., 2009;Pöltl et al., 2012).In addition, it allows developmental neurotoxicity to be addressed (Stiegler et al., 2011).LUHMES are conditionally-immortalized neuronal precursor cells with dopaminergic neuronal features.Differentiation is triggered by a tetracycline-mediated inactivation of the v-myc transgene in the cells, and it results in uniformly post-mitotic neurons forming a neurite network (Fig. 2) (Scholz et al., 2011).Neurite outgrowth can be measured in an imaging-based procedure.Live imaging allows the simultaneous evaluation of cell viability and neurite outgrowth within one culture dish.Some compounds can slow the extension of neuronal processes at lower concentrations than those causing cell death.Extensionpromoting compounds have been identified as well.To evaluate the specificity of the assay, the actions of unspecific cytotoxicants have been tested.Thus described test system may be useful for high-throughput screens to identify neurotoxic agents and for closer characterization concerning mode of action (MoA), compound interactions, or the reversibility of their effects (Stiegler et al., 2011).
As another example for toxicological use, the LUHMES cell system was used to examine the effects of methamphetamine or 1-methyl-4-phenylpyridinium (MPP + ) and the parental compound 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP).As expected, cells were sensitive to MPP+, while no reaction occurred with MPTP.The high homogeneity and purity of the cultures allowed the detection of metabolic changes during the degeneration, which is not possible in mixed primary human cell populations (Hansson et al., 2000).Cellular ATP in LUHMES cells dropped in two phases, cellular glutathione (GSH) decreased continuously, paralleled by an increase in lipid peroxidation.These events were accompanied by a timedependent degeneration of neurites.Blocking the dopamine transporter completely abrogated MPP + toxicity (Schildknecht et al., 2009).By applying different inhibitors, the underlying mechanisms and pathways have been identified.ATP-depletion, as the initial mitochondrial effect of MPP + , requires further downstream processes to result in neuronal death.These processes may form self-enhancing signaling loops that aggravate an initial energetic impairment and eventually determine cell fate (Pöltl et al., 2012;Schildknecht et al., 2011).
Further investigations are facilitated by genetic engineering of the cells.These manipulations can change the susceptibility to chemicals, test mechanisms of toxicity, or provide additional information in the form of reporter assays.Such approaches are a counterpart to the use of transgenic animals, and LUHMES have been used for many such approaches.
The intention of modeling the real world situation as closely as possible makes it necessary to place cells in contact with other cell types.Neuronal networks and tissues naturally consist of different cell types.Therefore, to predict reactions in humans these conditions must be reproduced experimentally, and this testing step also would use approaches to define key toxicity events and/or biomarkers more causally, such as reporter assays and inhibition of suspected pathways.
Such battery approaches are well known from the field of genotoxicity (Kirkland et al., 2011).They are under active development for the area of skin sensitization, and it is expected that they will be broadly applied in the future in many other domains, such as developmental neurotoxicity (Kadereit et al., 2012).The battery should be validated carefully, including appropriate controls.The tests would generate in vitro concentration response curves.This should give a first idea of the biological profile of a compound.These data and their use in an appropriate integrated model should yield a point of departure (POD).The POD is defined as the concentration that results in a significant hazardrelated change in the in vitro system, which is considered predictive for the in vivo situation.Different systems with different endpoints, considered also at different incubation times, will result in a large set of data on the compound's effects at different concentrations.This will be particularly important, as test systems diversify.In addition to the "classical tests," mimicking certain complex modes of action (e.g., inflammatory activation of cells, or disturbed differentiation), new test systems will either test only the effects of compounds on defined biochemical targets (e.g., receptor activation or enzyme inhibition), or they will obtain broad information by different omics technologies (metabolomics, transcriptomics, proteomics).Moreover, some of the test systems will also evaluate compound effects in 3D systems and between different cell types, e.g., between microglia and astrocytes (Kuegler et al., 2012;Schildknecht et al., 2011).Some of the effects measured may not be related to toxicity but may be epiphenomena or cellular counter-regulations.Sometimes, the effects observed also may depend on the metabolic state of the cells.Modeling such different situations can be useful to predict toxicity under greatly varying conditions in humans (Latta et al., 2000;Falsig et al., 2004).In most cases, extensive modeling will be required to process the data and to determine the relevant pathways of toxicity (PoT), as well as the threshold concentrations that trigger them.These will then be used as POD for IVIVE (Blaauboer, 2010).When, finally, the variability and sensitivity of human subpopulations (Bolt et al., 2003;Pohl and Scinicariello, 2011;Mirlohi et al., 2011) are taken into account, this overall procedure may lead to a risk evaluation without the use of animals.

Use of engineered human cells for complex 2D models of toxicity
To start the risk assessment process, it is necessary to combine different biological methods to finally reach a prediction about the hazard of a substance.Relatively simple 2D systems may be a good starting point.Preferably, human cells should be used, but these often require technologically challenging procedures.Pluripotent stem cells may be a good general source of different cell types (Balmer et al., 2012).
Here, the LUHMES cell system is presented as alternative example.The acronym of these cells stands for "human mes- (D) GFP-over-express ng "mature• day 5 LUHMES ce w th a typ ca neur te ength of 800 Jlm and an end ng w thout growth cone (no further neur te growth).Insert shows neur te end ng at h gher magn ficat on.Adapted from (Scho z et a ., 2011).
possibly further refined by in silico modeling.One first step are co-culmres , as described already by Schildknecht et al. (2012).
As astrocytes serve numerous functions, such as nutrient supply of neurons, regulation of cerebral blood flow, orchestration of neuronal growth and differentiation, maintenance of extracellular glutamate levels, and ion and liquid balance (Kettenmann and Ransom, 2005), they seem to be the right partners to support neurons in their biological function .
TI1e future of this cell culture system lies in the combination of cell culture techniques and innovative analytical approaches.Metabolomics and transcriptomics are especially promising tools for providing a richer set of data.Combination with other cells and 2D or 3D stmcturing of the cultures appear very attractive.The lack of functional N-methyl-D-aspartate receptors is, however, a disadvantage compared to several primary neuron cultures (Volbracht et al., 2006).Thus, it is  important to know both the advantages and the limitations of LUHMES cells.
4 Automation-compatible organotypic microtissues for drug and substance testing Physiological tissue-like culture systems seem to improve the relevance of in vitro testing of substances tremendously.Although the advantages of organotypic 3D cell culture models to increase the performance of in vitro compound assessment have been known for years (Pampaloni et al., 2007;Justice et al., 2009), complex production and elevated readout processes impeded the industrial implementation.Microtissue models can be derived either from cell lines, primary cells, or stem cells.They are spherical in shape and devoid of artificial biomaterials such as hydrogels or scaffold materials.TI1e cells produce their own, cell type-specific extracellular matrix environment, mimicking native tissues in vitro.Often, they display some morphological tissue-like features, but they do not necessarily reflect the overall histology as seen in vivo.Nevertheless, biochemical functions of cells in 3D microtissues can be closer to the in vivo situation than to the one observed in 2D in 378 vitro cultures.As an example of a 3D liver microtissue model, the InSphero system was chosen, as this technology platfonn allows relatively high throughput and adaptation to many different cell types and requirements.
A novel automation-<:ompatible 96-well platform based on hanging drops to produce organotypic microtissues has been developed (Kelm and Fussenegger, 2004), and spherical microtissues can be produced in a scalable production process.This is achieved by a well design where a microfiuidic channel co1111ects an inlet funnel at the top and an outlet fu1111el at the bottom of the plate, allowing for hanging drop formation without turning the plate upside down.At the well outlet a hanging drop is formed by a combination of capillary and surface-tension forces.In this the cells assemble into a tissue-like arrangement (Fig. 3).Microtissues can be initiated with low cell numbers resulting in excellent size unifonnity with variations below 10% in diameter (Drewitz et al., 2011).The size is defined by the initial cell number.In contrast to tumorigenic and ES cells, primary cells are contact inhibited and do not proliferate (Drewitz et al., 2011).
Complementing the manufacturing platfonn for reproducible production of single and multi-cell type microtissues, a spheroid-specific culture plate allows more convenient long-tenn cultures, medium exchanges, optical analysis, and biochemical assays.The lack of artificial biomaterials ensures compatibility with most biochemical assays used to assess cell viability and toxicity.The whole system can be combined with robotic liquid-handling devices equipped with a 96-multichamlel pipette head, with sinlilar volumetric precision as in standard multiwell plates (Drewitz et al., 2011).Future developments may include the combination of periportal and perivenous types of hepatocytes and closer modeling of the liver lobule structure, while at present the focus of the platform is on robustness, throughput and improved overall biochemical function .
An example application is the use of liver microtissues from primary rat hepatocytes or cryopreserved human hepatocytes, both in co-culture with non-parenchymal cells (Godoy, in preparation; Messner et al., submitted) for toxicity testing.h1 standard sandwich culture systems, liver-specific functionality decreases rapidly with prolonged culture time.The 3D enviromnent already improves the functionality, whereas heterotypic cell populations incorporating NPCs further increase liver-specific functions as shown for urea secretion over time (Fig. 4).Actually, levels of secreted albumin are close to values detected for human native liver (Meng, 2010;Uygun et al., 2010) .fucorporation of liver-derived macrophages enables the detection of indirect liver toxic effects such as inftanunationmediated toxicity (Messner et al., submitted).
Microtissues are a versatile 3D cell culmre concept to create a wide variety of different tissue models, which are designed for high-tlu•oughput data generation and easy implementation in current drug development processes to foster drug de-risking.
The key advantages of microtissue models are: -Long tem1 maintenance of tissue structure and functionality -Multi-cell type models -No scaffold requirements (no impact of batch to batch variations from biomatelials and hydrogels) -Production of endogenous cell type-specific extracellular matrix -Direct cell-cell and cell-ECM interactions -Standard 96-well fonnat enables automation and highthroughput compatibility -San1e tissue format for efficacy and safety assessment 5 Use of genetically engineered 3D models for identification of toxicity pathways 3D organotypic in vitro human epithelial models are important advances over traditional monolayer cell culture models.For toxicology applications, these models provide a number of important features (Kandarova et al., 2009;Kaluzhny et al., 2011).
Specific advantages of 3D organotypic human epithelial models include: -Normal (non-inunortal) human cells -Organotypic structure -Barrier function -Real-life exposure options Rat hepatocytes were grown n 20 sandw ch cu tures.or as 30 m crot ssues.The m crot ssues ether conta ned non-parenchyma ver ce s (NPC) as add t ana ce popu at ons or they were formed from hepatocyte monocu tures.The ce s were kept n cu ture for 10 days, and the urea product on was measured every second day, start ng from day 4.The data were norma zed to the number of ce s n cu ture.and are g ven as product on per day.
-Xenobiotic metabolizing capabilities -Co-culture models for epithelial-stromal interactions 3D human skin (e.g., EpiDennTM) and corneal tissue (e.g., EpiOcularTM) models have become increasingly important as replacements for traditional animal-based toxicology testing in the area of cosmetics, personal care products, household cleaning products, and in the chemical and phannaceutical industries.Two EpiDem1TM -based test methods for assessing skin irlitation and dem1al corrosion potential are now fom1ally validated as alternative metl10ds in tl1e European Union (EU).
Several other EpiDenuTM and EpiOcularnt test methods continue tllrough the validation process as defined by regulators in the EU (ECVAM) and the United States (ICCVAM).Nonnal human 3D (NHu-3D) epithelial models with the added feature of engineered toxicological reporter functions are currently being developed (Hayden et al., 2011).These models represent a further advance that will allow development of mechanistic toxicity screening assays.Initial experiments have produced promising results.
Early passage nonual human epidennal keratinocytes, dermal fibroblasts, tracheobronchial epithelial cells, and fibroblasts were transduced witl1 lentiviral vectors containing NFKB reporters linked to either GFP or luciferase.Stably transduced

\ / B
Reporter Cells Schemat c representat on of the process for creal on of organotyp c reporter modes from norma human ce s.Norma ce s are so ated from human exp ants (e.g., sk nor ung) and kept n cu ture (shown n red).They are transduced w th ent v ruses conta n ng reporter constructs and an ant bot c res stance.The successfu y transduced ce s (shown n green) are se ected w th puromyc n, expanded and cryopreserved for ong-term storage.After recovery, cryopreserved ce s are ut zed for product on of organotyp c reporter modes.cells were selected by puromycin resistance, expanded over several passages and cryopreserved to produce large pools of reporter-expressing cells (Fig. 5).Reporter-expressing cells were then utilized to produce NHu-3D skin and airway epithelial models.
Important considerations for the development of organotypic reporter models are: -Stable integration of reporter into cells -Attainment of normal organotypic development -Attaitunent of adequate reporter activity -Generation of adequate number of cells to support commercial production -Simple, robust assays for HTP screening Organotypic stmcture and barrier properties of models produced from reporter-expressing cells were fmmd to be similar to models produced from untransduced cells, as detemlined by llistological and barrier assessment.NFKB reporters linked to either GFP or luciferase were found to be activated approximately 5-fold above background by treatment of the organotypic models with twnor necrosis factor (TNF)a.GFP was detected in formalin fixed paraffin sections by epifiuorescence microscopy, and luciferase activity in tissue extracts was quan-380 tified with a microplate luminometer (Fig. 6).Production of models containing other reporters of toxicological significance (e.g., for DNA damage, oxidative stress, heavy metal stress, .ER stress, etc.) by the same process will provide a suite of human epithelial reporter models that can be utilized to provide mechanistic toxicity screening assays.

Prediction of skin irritancy using mechanism-based integration of in vitro and in silico methods
In the field of skin irritation, numerous test models have been developed over the years, some of \Vhlch have been validated and accepted by OECD as stand-alone alternative methods to the Draize rabbit skin test (OECD, 2010).In silico methods also are available, including mle-based expert systems and QSAR prediction of binary classification of skin irritancy.For example, ToxTree implemented a decision tree for skin irritation mles (HUizebos et al., 2005) that requires users to provide physicochemical properties as exclusion mles prior to applying inclusion mles based on stmctural alerts.After st mu at on w th 50 ng/m TN Fa for 48 h, t ssue morpho ogy appeared norma (C).V sua zat on of the reporter s gna by fluorescence m croscopy showed the act vat on of the NFKB reporter n TN Fa treated t ssue (D).(E) For quant tat ve assessment of Ep Derm-FT"' NFKB reporter act v ty, t ssue was generated w th fibrob asts (FB) conta n ng reporter constructs, or w th kerat nocytes (KC) conta n ng reporter constructs or w th both ce types conta n ng reporter constructs.The t ssues were then st mu ated ether w th TN Fa or w th the to -ke receptor agon st po y nos n c ac d (Po y(I:C)).
After 48 h, the t ssue was harvested, homogen zed and used for determ nat on of uc ferase ( uc) act v ty.
Here we present an automated workflow algorithm developed and implemented to predict a molecule's potential for skin irritancy (including severity grading).In vivo Draize legacy data (OECD, 2002) have been used as learning and training data for the model.

MoA QSAR modeling approach
A theoretical framework to link initial interactions of a chemical with its target site to toxicologically relevant endpoints on the level of the organism or the population are the adverse outcome pathways (AOP), defined initially in the field of ecotoxicology (Ankley et al., 2010).The concept has been taken up by several authorities dealing more broadly with environmental and chemical risk assessment, and it is further elaborated and applied to case srudies, for instance by the OECD.An AOP is identified by a causal linkage between a molecular initiating event (MIE) and in vivo endpoint of regulatory value for risk assessment.Levels in between may include cellular changes and responses of tissues or organ systems.Before AOPs were defined, the related concept of mode-of-action (MoA) was in use for a long time.It also describes the chain of events, explaining how a chemical leads to functional effects on a higher level of complexity (cellular or organism level).In many cases, the term MoA is used to describe the initial parts of an AOP at a level of high bioche1nical resolution.
A MoA QSAR modeling approach may be useful to reflect mechanisms when constmcting training sets and selecting model descriptors.The MoA QSAR approach includes the preparation of mechanistic training sets by identifying MoA categories that are linked to the phenotypic effects.MoA categories are defined based on chemical groups that initiate the molecular interactions in putative toxicity pathways and result in biological events thatlead to phenotypic effects.TI1e chemical groups that initiate this chain of events and therefore link s ght (45).moderate (68).and strong (8).
biological event pathways to phenotypic effects are defined as chemical MoA categories.These are identified from individual, highly resolved MoA pathways by aggregating several pathways into one category.The purpose of this "pooling" step was to build large groups of chemicals containing both positives and negatives in the training set.These chemical groups include surfactants , acids and bases (alkali and bleaches), organic solvents, and reactive groups.From a L'Oreal intemal source and the ECETOC report (ECETOC, 1995), a collection of 269 che1nical structures were selected as the global dataset.This global set was then partitioned into individual MoA groups: alcohols (98 stmctures), alkenes (68), amines (50), carboxylic acids/esters (93), reactive groups (135), and surfactants (84) (Fig. 7).
Here the surfactant group is selected as an example and treated in detail.Surfact.ants can act.in a number of ways , including the triggering of infiaimnation and/or cell death (De Jongh et al., 2006), altering signaling pathways (Tonna and Beme, 2009), or enhancing the loss of barrier function by dismpting lipid bilayer structures (Weiss et al., 2004).While the fonner are widely studied by in vitro assays and gene/mRNA expressions, the membrane damaging can be investigated by applying colloid chemistry and in physico assays (Yang et al., 2001).Surfactant molecules can dismpt lipid bilayers in two ways, and the purpose of these in physico assays is to develop descriptors to be used in our MoA modeling process.

Modeling procedure
Molecular descriptors were used to relate aspects of chemical structures to biological responses.Structural features were represented as fingerprints from a fragment library of generic features, as well as stmctural alerts for skin irritation.Physicochemical properties included calculated LogP (lipophilicity), water solubility, polarizability, polar surface area, and dipole moment.Further parameters that were used include molecular shape descriptors (e.g., diameter, moment of inertia) and surfactant-specific parameters, such as HLB (hydrophilic-lipophilic balance) and molecular packing factors (Israelachvili, 1994).Based on the structural features and properties, in the first part of the approach, a binary classification of non-irritants vs .irritants was perfonned for each MoA group as well as the global dataset.This was done using a combination of partial least squares (PLS) and logistic regression.The probabilities of being an irritant were then combined from the different models into one overall outcome using a quantitative weight of evidence (WoE) approach.Each model is assigned a weight to optimize the prediction of the non-irritants during the training stage.

L W;P;
where Wi is the weight of each model and Pi is the probability of being an irritant for a given chemical.The sensitivity was 95% and the specificity was 84%.The app cab ty doma n of the mode s eva uated based on both structura features and phys cochem ca propert es used n the mode s.MoA: mode of act on; WOE: we ght of ev dence.
Next, irritant chemicals are further ranked for severity effects using ordinal partial least squares regression.The applicability domain of the model was evaluated based on both structural features and physicochemical properties used in the models (Fig. 8) .
As part of L'Oreal's integrated testing strategy, a workflow system (Fig. 9) was designed for easy and transparent access to the decision tree of structural alerts and MoA QSAR models for in vivo skin irritation.
7 Human variability: Solutions through the application of in silico pattern recognition and knowledge discovery methods to human data Research tmdertaken with human subjects and samples suffers from large inter-individual variations, and therefore encourages the view that the science is potentially of poorer quality since predictions emanating from such human research will be less reliable and repeatable.TI1erefore, the issue of variation dictates much of the ongoing preference for inbred animals in controlled laboratory environments as tools to understand human health and disease, both from "quality science" and experimental control perspectives.This approach largely neglects the problem of species differences (Olson et al., 2000;Hartung, 2008), when the main purpose of toxicological research is the extrapolation to humans.In modem toxicological test systems, this issue is not fully resolved when conclusions for humans are drawn from zebra fish, drosophila, or cultures of non-human cells.The species extrapolations are made particularly difficult, as "man" is not one genetically defined entity.Large variations, relevant for toxicological and disease outcome, are being mapped by the ENCODE project, for instance.A different approach is taken by the exan1ple technology presented here: improved panem recognition in heterogeneous lnunan data sets, in order to more sharply define human responses on a broad population level.
Variation issues associated with the data produced from human studies are a significant challenge in developing human models suitable as replacement alternatives for animals in biomedical, toxicological, and other health research and investigations.Additional impetus for considering alternatives to animals comes through an increasing focus on the tmreliability of translation from some animal models to htunan disease, and thereafter the development of vaccines and therapeutics (Pound et al., 2004;Hackam and Redehneier, 2006;Khanna and Burrows, 2011).

Human variability
An advantage from modem medical and health systems is the generation of enonnous voltunes of data.Large data sets lend themselves to modem computational approaches, such as machine learning and sophisticated multi-dimensional statistical teclmiques that together comprise the bioinfonnatics field of pattern recognition.Machine learning methods, for example recursive partitioning (decision trees) and support vector machines (SVM), are particularly attractive, since algorithms like these can be trained to identify patt.ems in highly variable data associated with a response (e.g., test for a viral infection, drug side-effects), allowing more accurate predictions from complex data in the subsequent testing phase.These techniques, borrowed from computer science and statistics and combined with access to vast human data sets, provide another avenue through which to approach biological variation and provide an alternative to mice with experimenter-defined genetics and living environment.In short, to overcome the diversity of human subjects that can confound interpretation, sophisticated in silico pattern recognition methods can identify data profiles for sub-populations represented by the predictor variables analyzed and based on a biological or toxicological response of interest.

Machine learning examples: Decision. trees and support vector machines (SVM)
Decision trees have been applied to clinical decision-making involving diverse and complex patient data (Crowley et al., 384 2002), so the application of machine learning has precedent.in the medical knowledge domain.A decision tree is a data classification method generated by asking serial questions on the features associated with data items.This can be simple "yes or no" questions contained within a node, with each node comprising a distinct yes or no outcome.From the top node (the root) a hierarchical path based on yes/no answers continues until a node without further outcomes (a "leaf') is reached; a data item is thereafter assigned to this class (i.e., the terminal classification outcome) (Kingsford and Salzberg, 2008).
Also popular with in.silico pattern recognition studies are SVM.SVM are very powerful for data classification and novelty detection, and they can be used for regression modeling.SVM have the advantage of a simple input space for data entry, with analysis occurring in a high-dimension feature space (<I>) defined by a kernel function.The basis of modeling is the kernel, the simplest example being the linear kernel, which represents the in1ages of two data points (x, x') in the high-dimension feature space (1): ( 1) k(x, x' ) =<<I> x), <I> (x')> (Karatzoglou et al., 2006) Depending on prior knowledge of the data (e.g., distribution, structure) and application, other kernel functions are available (Karatzoglou et al., 2006;Smola and Scholkopf, 2004).
The application of decision trees in tandem with SVMs is being developed to overcome the difficulty of human variability as a battier to replacement alternatives for anin1al studies in biomedical research.This approach provides a pattern recognition tool for large multi-parameter (up to thirty predictor variables) data collections containing thousands of individual cases (i.e., patient results).Decision trees have the advantage of multiple decision boundaries that, when combined with SVMs, provide a powerful means to identify meaningful patterns in complex 1mman health data, either prospectively or retrospectively, which can guide further investigations into a disease mechanism or toxic drug response.
In Silica Pattern Recognition System -Disease, Toxicology A pattern recognition system, using only human data, is central to a new system that employs interaction of the wet laboratory for biological validation with computational pattern models and linkage to modem genetic analytical techniques and databases to further explore the basis of physiological clues detected in the in silica pattern recognition phase of investigations on aggregated pathology data (Fig. 10).
Natural variation in human genetics and environment need not be a barrier to fundamental biomedical research, toxicology studies, or any other aspect of human health research.With the massive accumulation of human data, whether through pathology testing, social and psychological analyses, or epidemiology, data analytic techniques borrowed from computer science and statistics make it possible to expose relevant and unique patterns within highly variable human populations .With such variation successfully harnessed and linked to biological validation strategies and sophisticated human genetic knowledge, viable alternatives to animal studies will emerge, particularly when tethered to modem knowledge of human cell culture systems for mechanistic studies.

Conclusion
Each of the technical approaches and model systems presented here has been developed as a stand-alone method.Most have been developed for a specific purpose and to solve defined problems.Dozens, if not hundreds, of such technologies arealready available, and only a few have been picked to exemplify the progress in the field.We have put forward the hypothesis here that added value may be generated by a combination of such approaches.The approach taken here may, at first glance, look different from or in competition with other new strategies.For instance, the ToxCast progran1, or different approaches that follow the "Tox21" vision take a different starting point.In tlleir extreme form , they rid themselves of the old patchwork of different toxicological models, be they in vivo or in vitro (Hartung and BcBride, 2011;NRC, 2007;Leist et al., 2008a), and put forward a new homogeneous framework, based, for instance, on PoT and systems biology modeling.It is not yet clear, which role assays play that use endpoints that are toxicologically apparently simple but (systems-) biologically highly complex, e.g., cell death, neurite degeneration, or albumin secretion.
Here we take an alternative approach to define an overall scaffold of what information would contribute to an animalfree risk assessment.This scaffold is used to recruit a largely heterogeneous group of assays, providing information at different levels of complexity, with different throughput rates, and possibly with different information value.Combined in a scheme, these assays can fill knowledge gaps and improve the overall risk assessment of chemicals for which little is known.
The framework suggested here is also suited to the incorporation of individual tests and in silico methods developed for Tox21, or even to incorporation of testing strategies at a higher level of integration, as shown by the Altamira example of skin irritancy modeling.Thus, this approach may represent a practical solution for high production volume risk assessment in the intermediate future, while many tests are still under development and no complete test platform on the basis of PoT testing is available.The future will then bring higher throughput assays, better systems biology modeling, better integration of data from omics technologies, and better cell sources.For instance, we envisage that testing in non-transformed cell models, of murine (Sipes et al., 2011a) or preferentially of human origin, will require a further development of stem cell technology, to provide reliable cell sources (Leist et al., 2008c;Zimmer et al., 2011aZimmer et al., ,b, 2012;;Weng et al., 2012;Balmer et al., 2012).

Fig. 2 :
Fig. 2: LUHMES differentiation and neurite outgrowth during differentiation Pro ferat ng LUHMES ce scan be amp tied and eas y converted nto post-m tot c neurons.(A) Schemat c representat on of the 2-step d fferent at on procedure, n t ated by w thdrawa of the cytok ne bas c fibrob ast growth factor (bFGF) and add ton of tetracyc ne.(B) Representat ve scann ng e ectron m croscopy (SEM) mages of und fferent ated (day 0) and d fferent ated (day 5) LUHMES ce s.Marked squares nd cate areas that are shown at h gher magn ficat on to the r ght.(C) Neur te outgrowth was documented by t me-apse m croscopy from day 3 to day 5.A representat ve sequence of merged phase contrast and GFP channe mages, recorded for 8 h on day 4 s shown.Expand ng neur tes w th act ve growth cones are nd cated by arrows, retract ng neur tes are marked by arrowheads.

Fig. 3 :
Fig. 3: Microtissue <MT> formation in hanging drops: overview of different microtissue models produced with the hanging drop technology (A) The ce s accumu ate at the bottom of the drop, form ce -ce contacts and extrace u ar matr x and assemb e to spher ca m crot ssues.Format on of m crot ssues requ res between 2-7 days depend ng on the mode .(B) Automated product on of m crot ssues n the Grav tyPLUSTM p atform.After the format on process, m crot ssues are transferred nto the m crot ssue assay pate (Grav tyTRAPTM) for further cut vat on and assay.(C) Overv ew of d fferent m crot ssue mode s produced w th the hang ng drop techno ogy.

Fig. 6 :
Fig. 6: Activation of EpiDerm-FT'M NFKB reporter by TNFa o r Poly<I:C>The human Ep Derm-FP" sk n mod.e s shown as an examp e for reporter t ssues.Ep Derm t ssue was generated from ce s conta n ng reporter constructs for NFKB (green fluorescent prote n express on under the contro of NF-KB form croscop c assays; uc ferase express on for quant tat ve b ochem ca assays).Contro t ssue (A) d spayed typ ca corn tied squamous ep the a morpho ogy, and on y the basa membrane showed a fluorescence s gna (autofluorescence, 8).After st mu at on w th 50 ng/m TN Fa for 48 h, t ssue morpho ogy appeared norma (C).V sua zat on of the reporter s gna by fluorescence m croscopy showed the act vat on of the NFKB reporter n TN Fa treated t ssue (D).(E) For quant tat ve assessment of Ep Derm-FT"' NFKB reporter act v ty, t ssue was generated w th fibrob asts (FB) conta n ng reporter constructs, or w th kerat nocytes (KC) conta n ng reporter constructs or w th both ce types conta n ng reporter constructs.The t ssues were then st mu ated ether w th TN Fa or w th the to -ke receptor agon st po y nos n c ac d (Po y(I:C)).After 48 h, the t ssue was harvested, homogen zed and used for determ nat on of uc ferase ( uc) act v ty.

Fig. 7 :
Fig.7: Distribution of skin irritation severity effects in the training set D str but on of the sever ty of sk n rr tat on n the tra n ng set accord ng to L'Orea cr ter a: non-rr tant (87 structures).we to erated (42).

Fig. 9 :
Fig. 9: Workflow implementation of "Skin Irritation Predictor"Essent a components of the system configured and a gned through a chem stry-aware p pe n ng techno ogy.
Fig. 10: Extraction of patterns from clinical/epidemiologic data despite large human variabilityA strategy to overcome natura human var at on based on in silica pattern recogn ton stud es, nked to b o og ca va dat on nvest gat ons as a rep acement a ternat ve for nbred an ma s n b omed ca research and tax co ogy assessments.Further va dat on through genet c nvest gat ons a so ass sts the dent ficat on of mechan sms that can be further exp ored by human ce 20 and 30 cu ture systems, such as those descr bed e sewhere n th s rev ew.The nd v dua phases of th s strategy nc ude (1) Data preparat on and in silica pattern recogn ton mode ng of camp ex bu k human data by data m n ng and mach ne earn ng, n tandem w th b o og ca va dat on of data patterns assoc ated w th a defined response (e.g., nfect on, drug tax c ty); (2) I dent ficat on of data patterns assoc ated w th mportant phys o og ca , ce u ar, or b ochem ca processes assoc ated w th d sease or tax c ty response, wh ch w (3) gu de the cho ce of gene express on ch p/pane for deeper b o nformat cs ana ys s and compar son; (4) Invest gat on of response mechan sms through human ce cu ture nvest gat ons, gu ded by find ngs from phase 3. F gure adapted from L dbury eta.(2012).