Key words

1 Introduction

Acute systemic toxicity comprises the general adverse effects that occur after a single or multiple exposure of an animal to a substance within 24 h and during an observation period of at least 14 days. The substance may be administered orally, by inhalation, or dermally.

Acute mammalian toxicity tests are often the first in vivo toxicity tests to be performed on a chemical. In recent years there have been considerable efforts to replace, reduce, or refine these animal tests by applying alternative approaches, including both in vitro and in silico models. An increasing number of models are available to predict acute mammalian toxicity. This is partly due to the fact that a reasonable number of datasets are openly available for modeling. However, the reliability of the in vivo data can be highly variable, and the metadata provided is often insufficient to determine the suitability of the data for modeling purposes. Another challenge is related to the multiple mechanisms leading to this complex effect, which is typically expressed as a single numerical value (LD50 for oral and dermal toxicity; LC50 for inhalational toxicity). In addition there are also differences between the routes of administration and species, and different data should be modeled separately [1].

Target organs, such as the liver, kidneys, heart, lungs, and brain, can be affected by exogenous chemicals to the extent that they cease to function. Thus, the use of QSAR models for organ/system specific toxicity would be extremely helpful when predicting acute systemic toxicity. A limited number of QSAR models for specific target organ and tissue effects are available.

The information obtained from acute systemic toxicity studies is used in the hazard assessment of chemicals occurring in food, industrial chemicals, biocides, pesticides, and cosmetics. In this chapter, we give an overview of the regulatory requirements for acute systemic toxicity information in the European Union, the software packages available for assessment of acute systemic toxicity and organ- and system-specific toxicity, as well as the databases available for obtaining such data. Since comprehensive reviews of literature QSAR studies are available elsewhere [2–5], we focus here on some of the more recently published literature models for acute systemic toxicity. Some of these software and literature models are documented in the JRC’s QSAR Model Database (http://qsardb.jrc.ec.europa.eu/qmrf/).

2 Regulatory Context in the European Union

For the assessment of acute systemic toxicity, only in vivo tests are currently accepted by regulatory bodies (Table 1). However, in vivo acute systemic toxicity studies are prohibited for cosmetic substances and products [14].

Table 1 In vivo methods currently available for acute systemic toxicity
Full size table

The endpoint measured in the majority of the standard assays is animal morbidity or death. Evident signs of toxicity (i.e., clear signs of toxicity indicating that exposure to the next highest concentration would cause severe toxicity in most animals within the observation period) are only used in the oral fixed dose procedure (FDP), which causes less suffering and is, therefore, more humane.

The assessment of acute systemic toxicity is one component in the safety evaluation of substances and represents a standard information requirement within several pieces of EU chemicals legislation, including the Regulation on Classification, Labelling and Packaging (CLP) of substances and mixtures [15], the Regulation concerning the Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) [16], the Biocidal Products Regulation [17], the Plant Protection Products Regulation [18], and the Cosmetic Products Regulation [14]. In preclinical drug development [19], however, these studies are no longer required to support first clinical trials in man. The information needed can be obtained from appropriately conducted dose-escalation studies or short-duration dose ranging studies that define a maximum tolerated dose in the general toxicity test species [20, 21]. Further information on the regulatory requirements in the EU is given in Prieto et al. [22].

3 Software for Predicting Acute Systemic Toxicity

Several software tools capable of predicting endpoints related to systemic toxicity are available, as reviewed previously [23]. An updated list is given in Table 2 and some updates on the programs are described below.

Table 2 Software tools for systemic toxicity endpoints
Full size table

Among the commercial software programs covering a broad spectrum of systemic toxicological effects is ACD/Labs Percepta, which is developed and marketed by Advanced Chemistry development Inc. (http://www.acdlabs.com/). The platform has two modules related to systemic toxicity prediction—Acute Toxicity Prediction Module and Health Effects Prediction Module. The Acute Toxicity predictor has been built using experimental data for more than 100,000 compounds extracted from the Registry of Toxic Effects of Chemical Substances (RTECS) and former European Chemical Substances Information System (ESIS) databases. It provides three different software components related to acute mammalian toxicity:

  • LD50—Provides predictions of LD50 values for rats and mice according to various routes of administration. Prior to modeling, the original experimental data were converted to logarithmic form (pLD50) in order to maintain linear relationship with used descriptors. The final prediction results returned to the user are converted back to LD50 values (mg/kg). The predictive model for pLD50 has been derived using GALAS (Global, Adjusted Locally According to Similarity) modeling methodology.

  • Hazards—A knowledge-based expert system that identifies and visualizes hazardous structural fragments.

  • Categories—Classifies compounds into one of five Globally Harmonised System (GHS) categories for acute oral toxicity.

The Health Effects module predicts the probability of a compound having a health effect on a particular organ or organ system (blood, cardiovascular system, gastrointestinal system, kidney, liver, and lungs). The models are based on data collected from chronic, sub-chronic, acute toxicity and carcinogenicity studies with adverse effects reported in particular organs or organ systems.

A common goal of toxicity prediction is to distinguish between toxicologically active and inactive compounds. Since multiple mechanisms are involved in systemic toxicity, this requires the availability of predictive tools that are able to cover a wide region of the activity space. This is the main feature of the expert systems that make assessments on the basis of structural alerts covering a spectrum of structural properties associated with the complex endpoint. One commonly used expert system, developed and marketed by Lhasa Ltd (http://www.lhasalimited.org/), is Derek Nexus which is a development of the former Derek for Windows (DfW). This contains knowledge rules derived from the known relationship between a given substructure and a toxicological effect of the molecule and applies these rules to predict potential toxicological effects of compounds. Derek Nexus generates a prediction by comparing the structural features of the target compound with a toxicophore encoded as structural pattern(s) in its knowledge base. The final predictions are derived from a reasoning scheme which takes into account the presence of a toxicophore in the query structure (‘structural alert’) and a limited number of calculated molecular properties, which, taken together, return an “uncertainty term” for the prediction itself. For some alerts, supporting examples are provided and the system states whether the query compound already exists as an example in the knowledge base. Literature references are also included to enable the user to assess the applicability of the structural alert to the predicted structure and to allow for an expert knowledge assessment. Derek Nexus covers multiple endpoints, including hepatotoxicity, nephrotoxicity, and cardiotoxicity.

CASE Ultra (http://www.multicase.com/) is further development of MCASE methodology and falls in the range of fragment based QSAR expert systems [24]. The CASE Ultra model mainly consists of a set of “positive alerts” (biophores), and “deactivating alerts” (biophobes), i.e., those fragments that are identified as statistically significant for increasing/decreasing the activity. The improvement of CASE Ultra over its predecessor is related to the identified alerts that are no longer limited to linear paths of limited size or limited branching pattern. In addition the training sets can be larger than 8000 molecules. The applicability domains of individual toxicity alerts within the models quantitatively define the necessary structural environment of the toxicity alerts.

The statistically based program TOPKAT (http://accelrys.com/) uses multiple QSARs on small and homogenous sets of data. It is now a part of QSAR, ADMET and Predictive Toxicology module within Biovia Discovery Studio platform. The rat oral LD50 module in TOPKAT comprises 19 regression analyses developed using experimental values of approx. 4000 chemicals from RTECS, including pesticides and industrial chemicals. The rat oral LD50 module in MCASE (named A56) is based on and comprises data for 7920 chemicals from the FDA, WHO, and NTP datasets. Tunkel and coworkers [25] compared the performance of the TOPKAT and MCASE rat LD50 modules against an external test set of 73 organic compounds covering 32 chemical categories retrieved from submissions to the EPA High Production Volume (HPV) Challenge Program (http://www.epa.gov/chemrtk/). The predictive accuracy of each software tool was assessed by applying the EPA’s New Chemical classification approach (http://www.epa.gov/oppt/newchems/index.htm), from the low-concern class (>2000 mg/kg) to the high-concern class (<15 mg/kg). While neither model was able to classify all 73 compounds, TOPKAT correctly classified 67 % of the chemicals, while MCASE classified 70 % correctly. However, it should be noted that the test set used was significantly skewed toward “low concern” chemicals, which both models predicted correctly with a high degree of accuracy (82 % and 100 % correct for TOPKAT and MCASE, respectively). Moreover, a high degree of false negatives was found for moderate and high concern HPV chemicals (TOPKAT, 72 %; MCASE, 100 %), suggesting that these programs are less reliable for the identification of more toxic compounds. The authors also compared the model outputs against the GHS five-tier scheme for classification of rat oral acute toxicants (<5, 5–50, 50–300, 300–2000, and 2000–5000 mg/kg), which is similar to the one adopted by EPA (<15, 15–50, 50–500, 500–2000, >2000 mg/kg). When compared against the GHS scheme, the ability of TOPKAT and MCASE to produce correct classifications was 73 % and 70 %, respectively, for the HPV test set chemicals, thereby changing slightly with respect to the EPA scheme, albeit enough to invert the rank order of these models.

VirtualToxLab is an in silico technology for estimating the toxic potential of chemicals [26] based on an automated protocol that simulates and quantifies the binding of small molecules towards a series of proteins, known or suspected to trigger adverse effects. The interface to the technology allows building and uploading molecular structures, viewing and downloading results and rationalizing any prediction at the atomic level by interactively analyzing the 3D binding mode of a compound with its target protein(s) in real-time. The VirtualToxLab has been used to predict the toxic potential for over 2500 compounds and the free platform, OpenVirtualToxLab, is accessible (in client-server mode) over the Internet. It is free of charge for universities, governmental agencies, regulatory bodies, and nonprofit organizations.

The LeadScope software (http://www.leadscope.com) links chemical and biological data that allows exploration of large sets of chemical compounds, their properties, and biological activities. Chemical structures are organized in a taxonomy of familiar structural features each combined with common substituents—the common building blocks of medicinal chemistry [27]. LeadScope provides QSAR models for diverse physiological adverse effects including cardiological, hepatobiliary, and urinary endpoints.

Other software tools available for predicting acute toxicity (LD50) to rat/mouse are also available, such as TerraQSAR (http://www.terrabase-inc.com/), ADMET Predictor (http://www.simulations-plus.com), Molcode Toolbox (http://molcode.com/). The TerraQSAR software is based on neural network methodology and includes models for predicting both oral and intravenous LD50 values in mice and rats. ADMET Predictor includes a number of in-built models for ADMET, and allows new predictive models to be built from the user’s data. ADMET Predictor’s Toxicity Module provides predictions of various toxicity endpoints including hepatotoxicity, carcinogenicity, acute rat toxicity, and cardiotoxicity. Molcode Toolbox has a range of modules for predicting toxicological endpoints, including intravenous acute LD50 values and in vitro cytotoxicity (IC50 values) (from the Registry of Cytotoxicity). The models are well documented and the underlying experimental data is made available with references and structure files (MDL molfiles).

4 Databases Containing Information on Acute Systemic Toxicity

Sources of rat LD50 values which may be suitable for the development of QSARs are listed in Table 3. Some recent updates are discussed in the section below.

Table 3 Databases containing acute toxicity information
Full size table

In particular, Acutoxbase [29] was developed in the context of the EU FP6 project ‘A-Cute-Tox’ (http://www.acutetox.eu/), which aimed to optimize and “pre-validate” an in vitro testing strategy for predicting acute human toxicity ([30, 31]; Prieto and Kinsner-Ovaskainen 2015). While the database is not available, the in vitro and animal data are published in several publications [30–32].

Recently the COSMOS database has been developed as a part of the COSMOS project (http://www.cosmostox.eu/), one of seven projects forming the Seurat-1 research cluster (http://www.seurat-1.eu/). Version 1 of the COSMOS database (http://cosmosdb.cosmostox.eu/) contains 12,538 toxicity studies for 1660 compounds across 27 endpoints, including acute toxicity data for 1697 compounds tested on different animal species, as well as in vitro data.

The Hazardous Substances Data Bank (HSDB) is a part of NLM’s Toxicology Data Network (TOXNET) [33]. It contains chemical substance information with one record for each specific chemical or substance, or for category of chemicals or substances. HSDB has approximately 5600 chemicals and substances, with information for toxicity and human exposure. All data comes from public scientific sources. HSDB’s content is peer-reviewed by a group of experts.

The Registry of Toxic Effects of Chemical Substances (RTECS) database includes basic toxicity information for: prescription and non-prescription drugs, food additives, pesticides, fungicides, herbicides, solvents, diluents, chemical wastes, reaction products of chemical waste, and substances used in industrial and household situations. It covers six categories of toxicity data including acute toxicity data. In vitro toxicology data has been added as well. Accelrys now produces the RTECS files using existing data selection criteria and rules established by NIOSH (http://accelrys.com/products/databases).

In order to be useful for QSAR development, datasets should be first curated, i.e., the accuracy of the structures should be verified and the quality of biological data should be reviewed. It is useful to provide a reference to the source of the experimental data. In addition, inorganic and organometallic compounds, salts, and compound mixtures are often removed from the analysis. For the development of QSARs, LD50 values should be converted to log[1/(mol/kg)] (if originally expressed as mol/kg or mg/kg). Finally, approximate LD50 values should be converted to discrete values, and multiple LD50 values from different labs/experiments should be converted to a single value. The ChemIDplus and ZEBET databases have been recently employed as data sources for QSAR analyses [34, 35].

5 Prediction of Organ-Specific and System-Specific Toxicity

5.1 Ability to Predict In Vivo Toxicity

Some currently available software tools (e.g., TOPKAT and MCASE) are useful for predicting acute toxicity in categorical terms (e.g., in terms of GHS classifications). The performance of different software tools in predicting acute toxicity has been investigated [36, 37]. In these studies, ACD and T.E.S.T. have performed well.

In the scientific literature, local QSAR models have been generated for sets of congeneric compounds (organophosphates, aromatic amines, anilines, etc.) and are scattered over many original publications. Some of these studies have also explored the use of in vitro data as additional descriptors in the derivation of so-called quantitative structure activity-activity relationships, QSAAR [38]. QSAAR modeling revealed good potential for acute toxicity prediction, particularly in cases when a significant correlation exists between in vivo data (LD50) and in vitro cytotoxicity (IC50), and the additional inclusion of physicochemical parameters serves to improve the correlation. In practical terms, QSAAR could be particularly useful if high-throughput screening methods are used to generate the in vitro data.

Despite their limited applicability when taken individually, local QSAR models might be usefully combined into an expert system for toxicity predictions. As a part of the efforts to develop global QSAR models for acute toxicity Raevsky and coworkers [39] proposed the so-called Arithmetic Mean Toxicity (AMT) modeling approach, which produces local models based on a k-nearest neighbors approach. Arithmetic mean toxicity values of one or more pairs of analogues (nearest neighbors) are considered as the toxicity of the chemical of interest. Recently a classification model based on 436 Munro database chemicals and developed using Dragon descriptors has been proposed as a tool for chemical screening [40]. Kleandrova et al. [3] have developed a multitasking (mtk) QSTR model based on ANN (artificial neural networks) for simultaneous prediction of acute toxicity by considering different routes of administration, different breeds of laboratory animals, and the reliability of the experimental conditions. The model is based on a diverse dataset comprising 1494 chemicals retrieved from CHEMBL (http://www.ebi.ac.uk/chembldb).

A consensus approach has been exploited in some studies where the models are built by using a combinatorial QSAR modeling approach, including multiple descriptors and employing several statistical modeling methods. It has been claimed that the predictive accuracy of consensus QSAR models is superior to the individual ones [34, 41]. In addition, several research studies [35, 42, 43] have demonstrated the ability to improve quantitative predictions for structurally diverse datasets when high throughput bioactivity data are used in combination with traditional molecular descriptors. This can also be regarded as an example of the QSAAR approach. These hybrid approaches and their underlying datasets are publicly available via the ChemBench web portal (https://chembench.mml.unc.edu/).

5.2 Ability to Predict Non-apical Toxicities

The feasibility of using in vitro cytotoxicity data for the prediction of in vivo acute toxicity has been investigated in a number of research programs [28, 44, 45]. Over 70 % correlation has been established between in vitro basal cytotoxicity and rodent LD50 values [46]. The applicability of 3T3 Neutral Red Uptake Cytotoxicity Assay for the identification of substances with an LD50 > 2000 mg/kg has been evaluated by the EURL ECVAM Scientific Advisory Committee (ESAC). It was recommended however that the results should always be used in combination with other information sources. For instance, the assay is recommended as a component of an Integrated Approach to Testing and Assessment (IATA) [47]. A reason for the absence of a clear relationship between basal cytotoxicity and in vivo acute toxicity could be that specific organ toxicity is the most sensitive parameter for acute toxicity. Common sensitive systems and organs include nervous, cardiovascular, immune system, kidneys and liver, lungs and blood. IATA proposed for acute systemic toxicity are a combination of complementary approaches (in vitro, ex vivo, in silico, in chemico) that address functional mechanistic endpoints tied to adverse outcomes of regulatory concern [48].

As summarized in Table 4, there is a limited number of literature models for predicting toxicities at tissue and organ levels. A list of software applications is provided in Table 2. They are based on expert system or regression/categorical QSAR models. In the case of ligand–protein interactions, molecular modeling approaches are mainly used. Among the commonly used software tools, Derek Nexus provides over 500 structural alerts for a range of organ and system-specific toxicities, and other miscellaneous endpoints. Models for predicting liver toxicity are further covered in Chapter 11 (Hewitt et al.).

Table 4 Overview of published organ and system-specific toxicity QSAR models
Full size table

Some of these models are based on the concept of reactivity-based toxicity. The covalent binding of reactive electrophiles to cellular targets (i.e., nucleophilic sites of macromolecules) has the potential to initiate a chain of biological effects (e.g., depletion of glutathione and protein thiols) resulting in specific organ and system toxicities.

Among the few comprehensive studies covering a range of organ toxicities and relying on a broad structural space in the training set are the models published by Matthews et al. [49]. These models were developed for urinary tract toxicities of drugs. For each organ, a number of toxicity endpoints were considered in the QSAR analysis. The investigation utilizes four software programs: MC4PC (versions 1.5 and 1.7); BioEpisteme (version 2.0); MDL-QSAR (version 2.2); Leadscope Predictive Data Miner (LPDM version 2.4). The four QSAR programs were demonstrated to be complementary and enhanced performance was obtained by combining predictions from two programs. The best QSAR models exhibited an overall average 92 % coverage, 87 % specificity, and 39 % sensitivity. These results support the view that a consensus prediction strategy provides a means of optimizing predictive ability.

In the work of Myshkin et al. [51], a detailed ontology of toxic pathologies for 19 organs was created from the literature in a consistent way to capture precise organ toxicity associations of drugs, industrial, environmental, and other compounds. Models for nephrotoxicity and for more specific endpoints related to these organ injuries were developed using a recursive partitioning algorithm. The models performed better at the prediction of distinct organ toxicity subcategories than general organ toxicity, reflecting the well-known tendency of QSAR models to have a better predictive performance for more specific endpoints.

In a more recent study, Lee et al. [50] present QSAR models for three common patterns of drug-induced kidney injury, i.e., tubular necrosis, interstitial nephritis, and tubulo-interstitial nephritis. Binary classification models of nephrotoxin versus non-nephrotoxin with eight fingerprint descriptors were developed based on heterogeneous pharmacological compounds data. Two types of data sets were used for construction of the training set, i.e., parent compounds of pharmaceuticals (251 nephrotoxins and 387 non-nephrotoxins) and their major urinary metabolites (307 nephrotoxins and 233 non-nephrotoxins). Thus the study reflects the fact that the nephrotoxicity of a pharmacological compound is induced by the parent compound as well as its metabolites. The results of a tenfold cross-validation and external validation procedures showed a high accuracy of the models (better than 83 % for external validation sets).

For kidney toxicity, local QSARs have been developed for specific chemical classes, such as the haloalkenes. These high-volume chemicals used in industrial, synthetic, and pharmaceutical applications are common environmental pollutants. Many haloalkenes are known to be nephrotoxic in rodents after bioactivation via the cysteine conjugate beta-lyase pathway, which is triggered by formation of hepatic glutathione S-conjugates, a reaction catalyzed by cytosolic and microsomal glutathione transferases [68]. The study by Jolivette and Anders [53] relates the nephrotoxicity of nine haloalkenes to their lowest unoccupied molecular orbital energies, E LUMO, reflecting their propensity for conjugation reactions catalyzed by glutathione transferase enzymes.

Very few QSAR studies of neurotoxicity have been published. An example is the work of Estrada et al. [57]. Their models are based on the TOPS-MODE approach, which provides a means of estimating the contributions to neurotoxicity in rats and mice of a series of structural fragments.

Organophosphorus (OP) compounds are well-known neurotoxic agents. These chemicals are potent inhibitors of serine esterases, the most critical of which is the widely distributed nervous system enzyme acetylcholinesterase (AChE). This well established mechanism of action underlies the usefulness of molecular modeling approaches like 3D QSAR and pharmacophore modeling to predict the inhibition potency of OPs. Several published models are based on these approaches [54, 55, 58, 63].

Among the commonly used software tools, Derek Nexus estimates neurotoxicity using a number of structural alerts: gamma-diketone or precursor, acrylamide or glycidamide, nitroimidazole, carbon disulfide or precursor, pyrethroid, 1-methyl-1,2,3,6-tetrahydropyridine, lead or lead compound, and organophosphorus ester.

Few studies have been published in relation to other organs/systems. Immunotoxicity can refer to immunosuppression in humans (caused, for example, by benzene and halogenated aromatic hydrocarbons), autoimmune disease (for example the pesticide dieldrin induces an autoimmune response against red blood cells, resulting in hemolytic anemia), and allergenicity (chemicals which stimulate the immune system can cause allergies or hypersensitivity reactions such as anaphylactic shock). Thus, immunotoxicity refers to a wide variety of biological effects, many of which involve complex biochemical networks. Tenorio-Borroto et al. [62] have trained and validated a multi target-QSAR model for high-throughput screening of drug immunotoxicity using TOPS-MODE approach. Yuan et al. [63] have studied the key molecular features of polychlorinated dibenzodioxins, polychlorinated dibenzofurans, and polychlorinated biphenyls for determining binding affinity to the aryl hydrocarbon receptor (AhR)—an intracellular receptor which has been correlated to immunotoxicity, thymic atrophy, weight loss and acute lethality. CoMFA (Comparative Molecular Field Analysis) was applied to generate 3D QSAR models. In a study by Hui-Ying et al. [64], linear relationships between immunotoxicity values (log ED50) and other biological activities of polychlorinated diphenyl ethers and their structural descriptors were established by multiple linear regression. It was shown that the structural descriptors derived from molecular electrostatic potentials together with the number of the substituting chlorine atoms on the two phenyl rings can be used to express the quantitative structure–property relationships of polychlorinated diphenyl ethers.

Evaluation of hematotoxicity is important step in early drug design. Particularly it is a common dose-limiting toxicity associated with anticancer drugs. The first attempt to build in silico models to predict the myelosuppressive activity of drugs from their chemical structure was made by Crivori et al. [65]. Two sets of potentially relevant descriptors for modeling myelotoxicity (i.e., 3D Volsurf and 2D structural and electrotopological E-states descriptors) were selected and PCA (Principal Component Analysis) was carried out on the entire set of data (38 drugs). The first two principal components discriminated the highest from the least myelotoxic compounds with a total accuracy of 95 %. In addition, a highly predictive PLS (Partial Least Squares) model was developed by correlating a selected subset of in vitro hematotoxicity data with Volsurf descriptors. After variable selection, the PLS analysis resulted in a one-latent-variable model with r 2 of 0.79 and q 2 of 0.72.

In contrast to other organ-specific effects, the in silico modeling of cardiotoxicity has been a rather productive field. This is because drug cardiotoxicity is one of the main reasons for drug related fatalities and subsequent drug withdrawals. In recent years, the hERG channel has been extensively investigated in the field of cardiotoxicity prediction as it has been found to play a major role in both cardiac electrophysiology and drug safety. Because hERG assays and QT animal studies are expensive and time consuming, numerous in silico models have been developed for use in early drug discovery. The earliest attempts to identify whether a molecule is a hERG blocker include a set of simple rules based on structural and functional features, but these rules are not always reliable predictors for identifying hERG blockers. In order to give more accurate predictions of hERG blockage, a wide range of QSAR models have been developed based on a variety of statistical techniques and machine learning methods, including multiple linear regression, partial least square (PLS), k-nearest neighbor algorithm (kNN), linear discriminant analysis (LDA), artificial neural networks (ANN), support vector machine (SVM), self-organizing mapping (SOM), recursive partitioning (RP), random forest, genetic algorithm, and naive Bayesian classification (NBC). Most of these QSAR models are classifiers and only a few regression models have been reported.

Pharmacophore modeling has also been employed to develop ligand-based prediction models of hERG channel blockers. Since the crystal structure of the hERG channel is not available, all structure-based studies on its blockage are performed on homology models and are more qualitative and descriptive rather than predictive. For example they have been used for molecular docking, molecular dynamics simulations and free energy calculations to explore the hERG-blocker interactions.

Reviews by [66] and Villoutreix and Taboureau [67] summarize the advances and challenges in computational studies of hERG channel blockage. It is expected that the development of in silico models for hERG-related cardiotoxicity will stay active in the coming years in order to design drugs without undesirable side effects.

6 Conclusions

The modeling of acute systemic toxicity has largely focused on QSARs for predicting LD50 values and for categorizing chemicals according to ranges of LD50 values. For these purposes, which are potentially useful in the regulatory assessment of chemicals, the in silico models seem to perform as well as in vitro cytotoxicity methods. The developments in this field can be attributed to the availability of extensive LD50 datasets and a wide range of machine learning techniques. Many of these datasets, and software tools derived from the datasets, are in the public domain.

The emergence of mechanism-based toxicology (e.g., adverse outcome pathways) is a tremendous opportunity to improve current models with better biological knowledge. Indeed, the time of global (and scientifically dubious) QSARs predicting LD50 based on chemical properties for the whole chemical space is probably coming to an end. Future models should target specific toxicity mechanisms on the basis of current biological knowledge. Historically, this was actually done implicitly by focusing model building on very limited chemical classes (supposedly acting via the same mechanism). According to this approach, global LD50 models would be the sum of a multitude of accurate predictors dedicated to describe well-defined mechanisms of action. In this context, the use of biological (in vitro) descriptors in combination with traditional molecular descriptors provides a promising means of building local QSAARs based on mechanistically based chemical classes.

In general, the modeling of organ-specific and system-specific effects represents an underdeveloped field, ripe for future research but far from regulatory applications, which typically rely on the assessment of lethality. A notable exception concerns the modeling of receptors and ion channels implicated in specific organ pathologies, such as the hERG channel in relation to cardiotoxicity. The development of models for upstream (molecular and cellular) effects represents a more scientifically meaningful exercise which also promises to unify the traditional regulatory distinction between the acute and repeat dose toxicity.

A future research initiative could include, for example, reexamination of the datasets for hepatobiliary and urinary tract toxicities of drugs with a view to developing more accessible models and assessing their applicability to chemicals other than pharmaceuticals. In addition, the concept of reactivity-based toxicity, now established as a plausible mechanism for hepatocyte toxicity, could be further exploited using data from hepatocyte cultures and cell lines. In some areas, such as immunotoxicity, short-term progress seems unlikely. The complexity of such effects probably means that systems biology approaches will be more appropriate.

In general, the development of models for organ-specific and system-specific effects will depend on a new generation of databases, such as the COSMOS database, which contain high quality data that are structured and annotated according to meaningful chemical and biological ontologies.