If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Yusuf Hamied Department of Chemistry, University of Cambridge, Lensfield Road, CB2 1EW, Cambridge, United KingdomMilner Therapeutics Institute, University of Cambridge, Puddicombe Way, CB2 0AW, Cambridge, United Kingdom
Various sources of information can be used to better understand and predict compound activity and safety-related endpoints, including biological data such as gene expression and cell morphology. In this review, we first introduce types of chemical, in vitro and in vivo information that can be used to describe compounds and adverse effects. We then explore how compound descriptors based on chemical structure or biological perturbation response can be used to predict safety-related endpoints, and how especially biological data can help us to better understand adverse effects mechanistically. Overall, the described applications demonstrate how large-scale biological information presents new opportunities to anticipate and understand the biological effects of compounds, and how this can support predictive toxicology and drug discovery projects.
Understanding the risk of potential adverse effects of compounds is crucial across all chemistry-related industries, especially in the pharmaceutical industry, and has hence also been studied across a broad chemical space [
Chemical toxicity prediction for major classes of industrial chemicals: is it possible to develop universal models covering cosmetics, drugs, and pesticides?.
]. Thereby, adverse effects manifest through different levels of biological organization. While we can model the binding of compounds on the molecular level relatively well, effects are more difficult to predict on higher levels as these can generally be induced through multiple molecular initiating events and depend on the broader cell and tissue context. Molecular initiating events refer to the early interactions between a molecule and a biological system which are then causally linked to the later outcome. Our understanding of these and our ability to measure these has advanced comparatively slower. Generally, effects are first induced on the cellular level, either through direct or indirect action of the compound and then lead to effects on the tissue or organ level influencing their physiological function. These can then eventually result in clinically relevant phenotypes, such as liver failure or other adverse outcomes (Fig. 1). Across all levels of biological complexity, there are generally multiple lower-level events that can induce the higher-level event. As an example, compounds can cause mitochondrial dysfunction through interaction with multiple protein targets, and this is in turn only one cellular mechanism leading to hepatotoxicity [
In this review, we will discuss in vitro models and how these can help us to anticipate the in vivo adverse effects of new compounds. We will first introduce the types of information that can be derived from chemical structure, in vitro and in vivo models, and will subsequently introduce different types of in vitro assays and the rationale for employing them in the context of toxicology. We will discuss how in vitro readouts can be modelled in combination with chemical and/or in vivo information to understand the observed properties of compounds or to anticipate biological properties of yet unseen compounds.
Types of information
Chemical structures: If we assume that the toxicity (or safety-related properties) of a molecule is linked to its structure, we may be able to predict the same based on the chemical structure. However, this requires a numerical or machine learning-friendly representation that may be obtained by some logical or mathematical procedure. Chemical features which are machine friendly can contain a wide range of chemical information about a molecule ranging from atomic to structural, geometrical, and even physicochemical properties. A useful feature retains sufficient and necessary information from the structure, is different in value for dissimilar molecules while also being distinguishable from other features. In the following section, we discuss various ways to encode chemical information into features.
Chemical structure can be described in different dimensions, ranging from 1D, such as a compound's molecular weight, to 2D, which describes the connectivity between atoms, to 3D, which additionally accounts for the orientation of atoms and bonds [
]. Descriptor vectors can be composed of bits, counts or continuous variables, such as computationally predicted physicochemical properties, and examples for these are summarised in Table 1. Thereby, descriptors can capture various types of information on the molecule which may be relevant in different contexts. For example, physicochemical properties such as partition coefficient, water solubility, and topological polar surface area are highly relevant for the pharmacokinetics of a chemical. In contrast, chemical structure and fragments are more commonly used in current quantitative structure-activity relationship (QSAR) studies. Although 3D molecular descriptors and quantum mechanical properties can be used to represent a compound, studies have shown that 2D fragments were overall more popular due to better predictivity and computational cost [
]. More recently, chemical structures have also been represented by molecular graphs which outperformed molecular fingerprints by being able to leverage deep learning techniques [
Weighted holistic invariant molecular descriptors. Part 2. Theory development and applications on modeling physicochemical properties of polyaromatic hydrocarbons.
Having derived chemical descriptors, the general assumption is then that a structurally similar compound is likely to show similar properties and may exhibit the same pharmacological effects, e.g. by binding to the same target protein, and hence chemical information can be valuable to anticipate compound properties [
]. Using compound descriptors, we can calculate the similarity between two compounds, as shown in Fig. 2, and can predict the relevant compound properties.
Fig. 2Principal Component Analysis (PCA) of a subset of compounds in Tox21 based on fragment-based fingerprints in DataWarrior (https://openmolecules.org/datawarrior/). The PCA shows the global structure of the chemical space covered, while the Tanimoto similarity describes the similarity of each compound to the query compound.
However, one potential limitation of models built with chemical information is that these are restricted by limited variation of the chemical space of the training data, which defines the model's applicability domain. For example, a model trained on one or several analogue series, irrespective of the dataset size, is likely not predictive for compounds that have a different scaffold from the training data. Instead, only structurally similar compounds are likely confidently predicted. However, also two structurally very similar chemicals can display large differences in binding affinity, which is referred to as an “activity cliff”. This may result from the lack of invariance of chemical space, which may be associated with the selected molecular descriptors and the training data [
], but can also simply be a consequence of the complexity of biological processes. For example, adding a methyl group to Sudoxicam can cause a different metabolic pathway, resulting in hepatotoxicity [
]. For these kinds of problems, information on the biological response to perturbation may be required to reveal the mode of action of a chemical and to predict the effects of compounds in vivo.
In vitro information: One approach to identifying functional similarity is to derive compound representations that are biologically more relevant than chemical structure. From the experimental side, this can be achieved by characterising the response of biological systems to perturbation with compounds in vitro using various readouts. In this context, different model systems are used with different levels of physiological relevance, and the underlying assumption is that these can ideally capture early convergent effects despite different molecular interactions with the biological system [
The most commonly used in vitro assays are ligand binding assays which measure the binding of the compound to a target molecule, which can e.g. be a protein or mRNA, and can give insights into which biomolecules a compound likely interacts with [
]. If the given target has a strong mechanistic link to a phenotype, the assay can then help to identify potentially unsafe compounds earlier. However, in the case of more complex phenotypes, such as systems-level adverse effects, the information may be limited as binding alone does not imply that the ligand has a functional effect, and perturbation of an individual protein might simply not be sufficiently informative. An alternative approach to measuring individual targets is hence to detect intermediate phenotypes in simple model systems. The simplest model systems for large-scale perturbations are immortalised cell lines which are well-established, scalable and cost-efficient. However, the relevance of these cell cultures is limited by the fact that both the cells and their surrounding conditions are artificial [
]. At the same time, it should be noted that also more complex in vitro model systems exist and are actively being developed which mimic physiological properties more closely and hence are promising tools to derive more relevant readouts. For instance, it was found that HepaRG cells that retain drug metabolism capabilities are better at detecting hepatotoxic compounds than HepG2 cells which lack these, while both immortalised liver-derived cell lines are less sensitive than primary human hepatocytes [
Characterization of primary human hepatocytes, HepG2 cells, and HepaRG cells at the mRNA level and CYP activity in response to inducers and their predictivity for the detection of human hepatotoxins.
Beyond mono cell cultures, more advanced in vitro systems should be mentioned which also aim to emulate whole tissues or organs. The two main directions in this regard are organoids, which aim to mimic organ function in vitro with high accuracy, and microfluidics-based “organ-on-a-chip” models which instead aim to capture relevant microphysiological aspects in a miniaturised manner [
]. As an intermediate between in vitro and in vivo models, patient-derived organoids should be mentioned which allow the same analyses as in vitro models but using cells and tissues derived from patients which hence closely resemble the in vivo physiology. As example, patient-derived tumour organoids have been used to screen drug sensitivity for 240 drugs in 5 days and are able to guide decision-making for the given patient and tumour shortly after surgery [
]. However, these are of course only able to capture information up to the organ level, which may still be insufficient to describe in vivo phenotypes such as adverse effects which manifest over longer periods and on a whole-organism level. Organ-on-a-chip models are particularly interesting for compound screening due to their focus on scalability. Assays to study drug-induced toxicity are being developed for multiple organ systems, e.g. for liver, kidney and heart, and are being further extended to multi-organ or whole-body human-on-a-chip assays [
]. Beyond drug safety, these models can also help to estimate a compound's pharmacokinetic properties, and can help to study the spatiotemporal effects of a compound at higher resolution than in vivo models as well as simpler in vitro models [
]. While the prospects and development in this field are encouraging, no larger-scale perturbation screens are yet available, which is why these organ-on-a-chip models will not be further discussed in the sections below.
From the introduced model systems, different types of readouts can be derived. These can give a broad overview on changes in biological processes including gene expression and imaging or can be aimed at specific physiological phenotypes, such as calcium transients which are used as an indicator for cardiomyocyte contractility [
] (Table 2). The currently available large-scale perturbation screens will be introduced below. A general limitation of these in vitro models is that it is difficult to extrapolate perturbation response to in vivo systems given that exposure is critical but generally not characterised in vitro. While there are efforts to better estimate the biologically effective dose in vitro [
], these are not currently widely employed. Consequently, most in vitro models generally only characterise hazard, the potential to cause harm, instead of risk, the corresponding likelihood. However, more and more advanced model systems are being developed which address this challenge of in vitro-in vivo translation, including multi-organ chips which model physiological pharmacokinetics, or more precisely first-pass drug absorption, metabolism and excretion through linked liver, gut and kidney chambers [
Curation and analysis of clinical pathology parameters and histopathologic findings from eTOXsys, a large database project (eTOX) for toxicologic studies.
Notoriety bias in a database of spontaneous reports: the example of osteonecrosis of the jaw under bisphosphonate therapy in the French national pharmacovigilance database.
In vivo information: In toxicology as well as other research areas, most in vivo information is derived from studies in animals, while this is complemented by clinical studies in the pharmaceutical industry. From the respective subjects, a wide range of information can be derived including information on the response to compound treatment from the cellular to the whole-organism level. The DrugMatrix [
] databases for instance contain transcriptomics, clinical chemistry, haematology and histopathology data from multiple organs in rats treated with 627 and 170 compounds, respectively, across multiple timepoints and doses (Table 2). These are invaluable resources in the field of toxicogenomics [
] as they contain multiple types of information within the same animal through which it is possible to derive mechanistic links, e.g. between gene expression and histopathology, while accounting for the variability between biological replicates.
Given that in vivo studies are costly and need to be ethically justified [
], these extensive studies across larger sets of compounds and data types are generally rare. One approach to tackle this is to integrate data from distinct studies, e.g. ToxRefDB [
] aimed to integrate different types of pre-clinical data from multiple pharma companies. Additionally, various databases exist which gather omics data, including in vivo perturbation data, such as Gene Expression Omnibus (GEO) [
Notoriety bias in a database of spontaneous reports: the example of osteonecrosis of the jaw under bisphosphonate therapy in the French national pharmacovigilance database.
]. Generally, these efforts are often challenged by missing standardisation of expert-guided annotations, both concerning terminology and severity scales as well as batch effects for measured readouts. However, if these challenges are overcome, the integration of studies might provide better information on phenotypic effects on a population level.
One general challenge in all in vivo studies is that there is not only a single response pattern to a compound, but instead, it is highly dependent upon the route of administration [
] are known to affect drug response in patients, but also behavioural and environmental factors can confound drug response. With respect to mapping to chemical space, this means that some level of abstraction is necessary to summarise biological response to something which can be modelled and ideally accounts for biological variability at the same time. One outstanding example for this is the FDA's work curating information on Drug-Induced Liver Injury (DILI). This resulted in the DILIst [
] databases which have explicitly been established to allow better evaluation of predictive models. However, it is debatable whether such complex information can meaningfully be reduced to a classification problem [
] which summarise frequency and severity of histopathological findings across replicates as well as previous work by Galeano et al. who modelled drug side effect frequencies demonstrating that this semi-quantitative information can be modelled as ordered classes [
In vitro approaches to characterise biological perturbation response
In vitro assays can be separated into two main classes based on whether or not the assay aims to measure a specific readout of interest, so is hypothesis-based, or instead measure general changes in the model system and is hence hypothesis-free (Fig. 3). Hypothesis-based assays are currently more established in drug discovery and toxicology, as the readout is clearly indicative e.g., for binding to a specific target or a specific cellular phenotype with established or at least anticipated in vivo relevance. In contrast, hypothesis-free assays have the potential to be informative for a wide range of endpoints, but the usefulness of the assay for a given application needs to be evaluated in the first place.
Fig. 3Hypothesis-based and hypothesis-free assays. Hypothesis-based assays aim to measure a known, and often low-dimensional, readout which is mechanistically linked to an in vivo endpoint and hence generally interpretable. In contrast, hypothesis-free assays measure biological response broadly. Consequently, their interpretation is not straightforward, however, they can potentially be informative for a wide range of endpoints.
Compound profiling using different bioassays is commonly used to predict the potential risk of adverse effects as well as to evaluate potentially desired effects, such as in target-based drug discovery [
]. For instance, secondary pharmacology screening is conducted in the early stages of drug discovery to identify potential off-targets of the compound and generally includes proteins associated with clinical adverse drug reactions [
], and these can be classified into five categories. 1) G protein-coupled receptors (GPCRs) are targeted by approximately 1/3 of all FDA-approved drugs in current medicines [
] and adenosine, adrenergic, cannabinoid, dopamine, opioid, muscarinic, 5-hydroxytryptamine and vasopressin receptors are prominently represented on secondary pharmacology panels. 2) Ion channels such as hERG have long been associated with undesired effects as they control cellular excitability and hence neuron and muscle function. Other potassium voltage-gated channels (Kv7.1, MinK), the sodium channel Nav1.5 and calcium channel Cav1.2 are also screened as they are related to cardiovascular risks. 3) Enzymes such as cyclooxygenase 2 (COX2) and monoamine oxidase (MAO) are associated with cardiovascular risks and hence are included in the screening panel [
]. The minimum panel includes dopamine, noradrenaline and serotonin transporters, which have a high hit rate and are associated with abnormal blood pressure and abuse liability. 5) Nuclear receptors include two targets, the androgen receptor and glucocorticoid receptor.
Because the mechanisms of adverse effects are not fully understood and target-based assays may be insufficient to estimate the biological effects of a compound, also a variety of hypothesis-based phenotypic assays are employed to profile the biological response to compound treatment. For example, cytotoxicity [
], which may be indicative of neurodegenerative disorders, stroke, myocardial infarction etc., cellular stress response assays, genotoxicity and carcinogenicity assays [
Hypothesis-free assays measure general changes in the model system and are not targeted at a specific readout. This means that the applicability of the assay needs to be evaluated for each endpoint. As for all in vitro models, the interpolation of hypothesis-free assays to in vivo effects is not a trivial task as effects can depend on the exact model system used, dose, timepoint of administration and other factors. While modelling of phenotypic screening data will be discussed in the subsequent section 3, we will first introduce the background of common types of large-scale technologies for hypothesis-free compound profiling.
One class of hypothesis-free assays, focussing on “-omics”, measures the abundance of specific biological entities (Fig. 4). Thereby, different technologies measure different types of biomolecules and these are continuously improving in terms of coverage, scalability and resolution [
]. One advantage of omics technologies is that there is already a large pool of public data, e.g. signatures characterising diseases, drugs or targeted perturbations of individual genes or pathways, using which it is possible to identify similar or complementary signatures [
], and knowledge about the measured entities, e.g. in the form of pathways or interaction networks, in particular for genes and gene products. Consequently, these can be easily mapped to a broader biological context providing mechanistic insight into the cellular state. Furthermore, the same methodology can be used to characterise different cell-based model systems, as well as living organisms, allowing versatile comparisons not only of perturbation response between compounds but also between biological systems. This enables studies on in vitro to in vivo extrapolation (IVIVE) which can provide insight into how well a model system resembles the in vivo response it tries to mimic [
]. For instance, it was shown that cancer cell lines showed similar transcriptomic responses to primary hepatocytes, which are frequently used as in vitro model for DILI, while concordance was poorer in comparison to liver gene expression profiles derived from repeat-dose studies in rats [
While different kinds of omics can be derived from all kinds of cell-based systems (Table 2, Table 3), we here focus on transcriptomics from in vitro cell cultures as these approaches are particularly cost-efficient and hence also particularly suited for large-scale compound screening. The two key technologies in this regard are microarrays, which quantify the abundance of a predefined set of mRNAs, and RNA-Sequencing, RNA-Seq, where captured mRNAs are instead sequenced so that in theory all sequences can be detected [
]. While microarrays have been the more popular technology of choice in the early 2000s, RNA sequencing has overtaken more recently due to continuously decreasing costs. For both technologies, cheaper high-throughput versions have been developed which are employed for compound screening. In this regard, the array-based L1000 platform should be highlighted which measures the expression of 978 landmark genes and has been used to characterise 19,811 compounds [
] making it the most comprehensive resource for uniformly generated transcriptomic data on chemical perturbations. One major limitation, however, is the limited number of measured genes although these were picked to be non-redundant so that 9,196 genes can be predicted with high fidelity. More recently, multiple sequencing-based platforms have been published including the DRUG-seq platform which overcomes this limitation and measures whole transcriptome changes for 2-4$ per sample [
]. Beyond increasing throughput and coverage, omics technologies are also improving in terms of resolution, providing measurements of single cells and in a spatial context which is particularly interesting for more complex systems modelling cell-cell interactions. While this is yet less explored in the context of chemical perturbations, the sci-Plex method should be mentioned which allows affordable measurement of single-cell transcriptomics from multiple experimental conditions through massive multiplexing and has been used to characterise 188 perturbations [
]. Hence, transcriptomics is now an affordable and widely implemented technology which allows profiling at a wide range of resolution and throughput levels.
Table 3Summary of datasets from different platforms used to profile compounds on different biological levels.
Cox MJ, Jaensch S, Van de Waeter J, et al. Tales of 1,008 small molecules: phenomic profiling through live-cell imaging in a panel of reporter cell lines. doi:10.1101/2020.03.13.990093
A more recent type of cell profiling that is distinct from omics is derived from microscopy imaging of cell cultures, which represents changes in cell morphology (Table 3). One assay for deriving high-dimensional cell morphology data which has been heavily researched is the Cell Painting assay used by the Broad Institute [
] for screening compound perturbations in human osteosarcoma (U2OS) cells. In this assay, the effects on the nucleus, nucleoli, cytoplasmic RNA, endoplasmic reticulum, Golgi apparatus, plasma membrane and the actin cytoskeleton are measured using live cell staining with respect to neutral DMSO control.
As shown in Fig. 5, the Cell Painting assay output can be thought of as a versatile descriptor of cellular response and characterises compounds based on the phenotypic changes in the stained cell organelle. The retrieved morphological traits include shape, and adjacency statistics, as well as intensity, texture, microenvironment, and context features. For example, texture features identify periodic variations in cell intensity and are e.g. used for single-cell analysis and to assess the intensity regularity in images.
Fig. 5Diverse phenotypes across compounds as captured in U2OS cells in the Cell Painting assay compared to neutral control DMSO. In the Cell Painting assay, cells are perturbed with chemicals in a multi-well plate setup and the key organelles and sub-compartments are stained using a combination of six well-characterised fluorescent dyes: Hoechst 33342 (DNA), concanavalin A (endoplasmic reticulum), SYTO 14 (nucleoli and cytoplasmic RNA), phalloidin (actin) and WGA (Golgi and plasma membrane), and MitoTracker Deep Red (mitochondria). For example, microtubule stabilisers podophyllotoxin and paclitaxel show giant, multinucleated cells. These images were extracted from https://idr.openmicroscopy.org/webclient/?show=screen-1952.
]. Furthermore, since these features are more computational and derived from images (correlations, adjacency, etc. between objects and images), their relevance in biological processes is not very well understood. There is some work towards interpretability, for example, it was found that features related to radial distribution and intensity in mitochondria object could be related to mitochondrial death [
As the assays are generated to address a wide range of applications instead of having a clear target hypothesis in mind, an additional challenge is to determine how much signal a given readout contains for a property or a compound of interest. The applications section on hypothesis-free assays below discusses how these assays can and have been applied to model diverse endpoints.
Computational approaches to model biological response
Data and knowledge resources: Mapping between chemical structure, in vitro and in vivo readouts can be helpful to anticipate the properties of unseen compounds with predictive models, and to understand the critical events linked to the phenotype with mechanistic models, which will both be discussed in the subsequent sections. In either case, it is necessary to integrate data types describing different levels of information (Table 2). As the compound coverage between these data sources varies widely, we provide a summary of the overlaps between compounds, compared with standardised InChI, in Fig. 6 for the datasets previously introduced in Table 3.
Fig. 6Compound overlaps between in vitro and in vivo datasets. (A) Compound overlap between bioactivity assays from Klaeger et al.
Cox MJ, Jaensch S, Van de Waeter J, et al. Tales of 1,008 small molecules: phenomic profiling through live-cell imaging in a panel of reporter cell lines. doi:10.1101/2020.03.13.990093
Besides compound-specific data, it is additionally possible to take advantage of prior knowledge on biological entities and their relations between each other to put the available data into a broader known biological context. For instance, on the cellular level, the Gene Ontology [
] can provide useful information on shared biological functions which can be formalized as networks. While our understanding of higher biological scales is less complete, some databases link these to the molecular and cellular level based on known or statistical associations [
Smalheiser NR, Bonifield G. Two similarity metrics for medical subject headings (MeSH): an aid to biomedical text mining and author name disambiguation. doi:10.1101/039008
] which can be a prior knowledge-driven approach to quantify compound similarity or, e.g. in the context of drug combinations, also complementarity.
Predicting toxicity and other endpoints: applications of hypothesis-based assays
If enough compounds have been measured using a given assay, this information can be used to predict the activity of novel compounds using so-called Quantitative Structure-Activity Relationship (QSAR) models which are based on supervised learning (Fig. 7). In these models, the chemical descriptors, as outlined in Table 1, are used to characterise the compound and the already existing data is used to train a predictive model. While this approach is generally feasible for any kind of assay and endpoint, it works well for common assays with abundant public data, such as in case of assays used in secondary pharmacology screening. The advantage of QSAR is that only chemical structure is required to predict a new compound while also the cost of computing is far lower than for approaches based on molecular modelling and simulation. Furthermore, various kinds of properties can be modelled, e.g. chemical structure cannot only be linked to target proteins but also to more complex readouts characterising biological responses such as those described in Table 2 [
]. However, there are also drawbacks to QSAR models, which are fundamentally linked to the fact that the model only predicts what was already seen in the training data: for instance, data imbalance might affect model performance [
], and only modes of action seen in the training set can be successfully predicted.
Fig. 7Supervised learning of properties (Y) from compound descriptors (X). Machine learning algorithms, such as Random Forests or Support Vector Machines can learn how to predict compound properties (Y), e.g. in vivo phenotypes or hypothesis-based assays, from sets of descriptors (X), e.g. based on chemical structure or assay readouts.
In addition to QSAR models, which use chemical structure to model compound activity against a certain protein, it is also possible to use already known chemical-protein associations to predict new potential connections[
]. An advantage of these kinds of approaches is that negative samples are not compulsory because chemical-protein associations are sparse (most of the bioactivities between the pairs are missing) and we can assume that most compounds are inactive for most target proteins. The basic concept underlying this method class is that if two compounds have similar bioactivities in target assays, it is likely that they have more common targets than those already measured. This approach is particularly useful when the bioactivity of a compound is already well-known, e.g. from previous screening panels. However, one limitation of the current approaches is that they cannot quantitatively assess the binding affinity, which may reduce the impact of the models compared to QSAR-based models.
Also, bioactivity readouts themselves can be used as compound descriptors if multiple targets are known to be linked to the phenotype. For example, Liu et al. used high-throughput screening data from ToxCast to predict the five most frequent chronic organ-level adverse outcomes (liver, kidney, adrenal gland, and lung) and found that bioactivity descriptors produced more accurate classifications than chemical descriptors in supervised models [
]. However, it should be noted that bioactivity data needs to be generated in the first place, and may not be available, especially for new chemicals. In this case, computationally predicted bioactivity can then be used to fill the missing values [
]. This integration of computational bioactivity fingerprints with QSAR was for instance previously found to be beneficial for scaffold hopping when compared to chemical descriptors[
Predicting toxicity and other endpoints: applications of hypothesis-free assays
Biological features from hypothesis-free assays can be used for a wide variety of purposes in the context of compound profiling, including unsupervised identification of functionally similar compounds and supervised prediction of specific properties. While the overall aim in computational toxicology is to predict adverse effects, in vivo data is generally rare which also limits how well the predictivity of an assay or descriptor can be evaluated. In contrast, in vitro data, e.g. from secondary pharmacology screening, is much more abundant and can hence provide a useful proxy to establish signals in assays, e.g. in comparison to chemical structure. We will hence discuss how hypothesis-free assays introduced above, namely transcriptomics and image-based cell profiling, compare to ligand-based descriptors and how these have been applied to model in vitro and in vivo endpoints.
Among transcriptomic signatures from compound perturbations, a significant similarity was found in 20% of cases for two structurally similar chemicals [
Relating chemical structure to cellular response: an integrative analysis of gene expression, bioactivity, and structural data across 11,000 compounds.
]. This shows that while chemical similarity indeed indicates a more likely similar functional effect, there are discrepancies which are not fully understood and may be a result of both technical noise and biological signal in the data. It was previously found that a drug-drug similarity network based on transcriptional signatures was able to identify ATC class enriched modules indicating that gene expression response is informative for a compound's mechanism of action [
The L1000 platform has also been used in the context of lead identification and optimization with the underlying assumption that it is possible to compare and prioritise compounds with respect to their expression signature instead of their chemical structure. Janssen profiled 31K compounds of the primary screening deck using the L1000 platform and found that these expression profiles were able to detect active and chemically diverse compounds in a case study on activity against the glucocorticoid receptor and the Hsp90 chaperone. Consequently, it has been suggested as a way for scaffold hopping [
]. Furthermore, it was explored as readout guiding lead optimisation comparably to QSAR and was found to be particularly useful in flagging off-target effects [
]. While the transcriptome itself describes a broad biological response and hence not a specific phenotype, it's often found that only a small subset of genes is needed to predict endpoints [
]. These predictive genes cannot only give insights into changes in biological processes but can practically lead to the development of targeted gene panels which aim to predict concrete endpoints at even lower cost.
High-throughput imaging data describing cell morphology can also be interpreted as a biological fingerprint that is characteristic for a compound, at least to the extent to which activity is visible in readout space. One of the first studies using morphological data [
] derived features from microscopy-based screens that were specifically built for glucocorticoid receptor nuclear translocation. Machine learning models were trained using cell morphological data to predict assay-specific biological activity. For two drug discovery projects, their findings showed a 60- to 250-fold improvement in hit rates over the initial high-throughput screens showing that cell morphology data can describe biological information not always contained in ligand structure. This has been further shown by Trapotsi et al. [
] who directly compared ECFP and cell morphology (Cell Painting) information for bioactivity prediction. Out of 224 targets, they could predict around 45% of targets using ECFP and 40% using image data with high AUC-ROC (>0.80). They showed that both descriptors worked better in other cases showing the partially complementary nature of cell morphology features to ligand structure. Furthermore, morphological features have also been shown to predict cell health phenotypes [
As both, transcriptomics and cell morphology, describe general biological response, the question arises whether both describe complementary information or whether there might be clear links between them. Previous studies on the relationship between changes in gene expression and cell morphology found that they could predict changes in cell morphology based on a transcriptomic query hence confirming that transcriptomic changes are related to morphological changes [
]. Way et al. showed that for the analysis of mechanism of action (MOA) classes, Cell Painting was more reproducible across profiles and could weakly cluster together a greater number of MOA classes while gene expression was better at predicting MOA and could cluster together a lesser number of MOA classes with a stronger signal [
]. When combining hypothesis-free data, either gene expression or cell morphology was able to produce a predictive model for many MOAs with AUC exceeding 0.70 suggesting they capture different aspects of the cell's response to chemical perturbations [
]. Overall, these studies show that, although transcriptomic changes are related to morphological changes, both are not redundant in terms of contained information [
Additional complementary information can be provided by ligand structure or chemical fingerprints, which is distinct from the biological information in cell morphology or transcriptomics data. Moshkov et al. compared chemical structures, cell morphology (Cell Painting), and gene expression profiles (L1000) for 270 assays from drug discovery projects where a combination of the feature spaces of morphology, gene expression, and chemical structure could predict 21% of all assays with high accuracy, compared to 6-10% when using single feature spaces [
]. Further they showed that compound bioactivity predictions could be improved further when combining predictions from phenotypic profiles and chemical fingerprints compared to using only one source of information. Overall, these investigations provide support for the use of gene expression and cell profiling data to advance areas such as target identification, mechanistic analysis, and toxicity prediction in conjunction with structural data.
Mechanistic adverse outcome pathways
Beyond prediction, data can also be used to derive mechanistic relevant events, e.g., on the molecular or cellular level, which are informative for adverse effects on the organism- or population-level. In the context of toxicology, for instance, the Adverse Outcome Pathway (AOP) framework was created which aims at describing event cascades describing the relation from Molecular Initiating Events (MIE) to the Adverse Event (AE) through Key Events (KE) on different biological levels [
]. Mechanisms can be formalised as directed event cascades or networks, which describe how upstream perturbations result in downstream phenotypic changes [
]. In this context, for instance, target binding assays can give insight into MIEs, while -omics data provides insight into KEs on the cellular level (Fig. 8). However, it should also be noted that to estimate the risk of adverse effects in vivo it is not only necessary to understand potential MIE but also whether these are likely to take place at the given treatment dose which is strongly dependent on the compound's ADME properties.
Fig. 8Adverse Outcome Pathways (AOP). AOPs describe the event cascades from the first interaction of a compound with the biological system, termed Molecular Initiating Event or MIE, to Key Events (KEs) on different biological levels to the Adverse Event (AE). Practically, AOPs can help to anticipate AEs by identifying useful intermediate events, which can be measured or estimated using suitable assays, and then in turn are informative for the likelihood of the subsequent events including the AE itself.
Mechanistically understood MIE or KE can in turn guide the development of assays, such as the hypothesis-based assays discussed above, which help to anticipate potential risks for adverse effects (Fig. 8). In this context, it should be noted that knowledge on relevant targets is largely incomplete and is often only identified after side effects of drugs targeting a particular protein are observed [
]. For instance, COX2 is part of the secondary pharmacology panel of all four companies studied by Bowes et al. and has only been linked to cardiovascular risks after rofecoxib, a selective inhibitor targeting COX2, was voluntarily withdrawn from the market due to potential risks of heart arrest and stroke [
When studying mechanisms of pathogenesis, the goal is to identify not only statistical associations but causal relations which can help rationalize decision-making. Different strategies exist to derive evidence for causality, e.g. in the AOP development guidelines, evidence for causality is defined based on the Bradford Hill criteria [
] and includes time concordance, dose concordance and incidence concordance as empirical criteria as well as biological plausibility and essentiality [
]. While this is also true to some extent for the other criteria, biological plausibility is particularly suited to be evaluated based on the different available data types and the available prior biological knowledge [
]. For example, Oki et al. integrate compound-disease associations from the Comparative Toxicogenomics Database (CTD) and compound-gene associations, either based on ToxCast or CTD, using frequent itemset mining to identify associations between genes and diseases [
], while Bell et al. derive compound-gene associations from ToxCast which are then mapped to compound-pathway associations and combined with pathway-phenotype interactions to a computationally predicted AOP network [
]. Jointly, these kinds of knowledge graphs can depict the biological understanding at a given time and thereby can support hypothesis generation, e.g. in expert-driven AOP development. In the context of mechanistic hypotheses, particularly prior knowledge on causal interactions should be highlighted which can be used to infer upstream mechanisms resulting in transcriptomic changes [
To derive practically useful event relationships from computationally inferred mechanisms, these need to be further characterized with respect to their applicability domain, the strength of evidence and predictive performance, which are some of the aims of quantitative AOPs (qAOPs) [
]. For instance, mechanistic qAOPs evaluate dose-response and time-course behaviour using experiments targeting individual event relationships. However, once an AOP is confidently established this information can be used to predict AEs from earlier MIEs or KEs. For instance, Burgoon et al. predicted a compound's probability to cause steatosis from in vitro targets representing known MIE using an AOP-based Bayesian network [
]. Overall, this shows how computational approaches can, both, support the identification of mechanistic knowledge and help to leverage this knowledge practically in the form of predictive models.
Conclusions and future directions
While there are already in vitro model-based strategies to link compound structure to in vivo safety, it is often unclear which specific model, assay or readout should be included to detect toxic drug candidates as early as possible in drug discovery. However, in vitro assays are advancing on the experimental and technical side resulting in more relevant model systems, such as organoids and microfluidic systems, as well as better readouts with higher throughput, e.g. achieved through assay multiplexing. Consequently, it can be anticipated that more in vitro data will become available to characterise the biological effects of compounds, hence also resulting in better models allowing to bridge chemical space to complex biological endpoints. This will not only help to better predict biological properties of new compounds but also will help untangle how the underlying biology functions, and consequently also how these effects can potentially be mitigated in drug discovery projects.
Funding
A.L. received funding from GlaxoSmithKline. S.S. acknowledges funding from the Cambridge Commonwealth, European and International Trust, Boak Student Support Fund (Clare Hall), Jawaharlal Nehru Memorial Fund, Allen, Meek and Read Fund, and Trinity Henry Barlow (Trinity College). S.S. acknowledges support with funding from the Cambridge Centre for Data Driven Discovery and Accelerate Programme for Scientific Discovery under the project title “Theoretical, Scientific, and Philosophical Perspectives on Biological Understanding in the Age of Artificial Intelligence”, made possible by a donation from Schmidt Futures. H.Y. acknowledges support from the Cambridge Alliance on Medicines Safety (CAMS).
Declaration of Competing Interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Anika Liu reports financial support was provided by GlaxoSmithKline. Srijit Seal reports financial support was provided by Cambridge Commonwealth, European and International Trust, Boak Student Support Fund (Clare Hall), Jawaharlal Nehru Memorial Fund, Allen, Meek and Read Fund, and Trinity Henry Barlow (Trinity College). Srijit Seal reports financial support was provided by Cambridge Centre for Data Driven Discovery and Accelerate Programme for Scientific Discovery made possible by a donation from Schmidt Futures. Hongbin Yang reports financial support was provided by Cambridge Alliance on Medicines Safety (CAMS). Anika Liu reports a relationship with Boehringer Ingelheim Pharma GmbH & Co. KG that includes: employment. Andreas Bender reports a relationship with Healx Ltd that includes: equity or stocks. Andreas Bender reports a relationship with PharmEnable Ltd that includes: equity or stocks. Andreas Bender reports a relationship with Pangea Botanica that includes: employment. Srijit Seal reports a relationship with AbsoluteAI Ltd that includes: employment.
Acknowledgments
The authors would like to thank members of the Bender Group (University of Cambridge) for their valuable input during the preparation of this work.
References
Alves VM
Muratov EN
Zakharov A
Muratov NN
Andrade CH
Tropsha A.
Chemical toxicity prediction for major classes of industrial chemicals: is it possible to develop universal models covering cosmetics, drugs, and pesticides?.
Weighted holistic invariant molecular descriptors. Part 2. Theory development and applications on modeling physicochemical properties of polyaromatic hydrocarbons.
Characterization of primary human hepatocytes, HepG2 cells, and HepaRG cells at the mRNA level and CYP activity in response to inducers and their predictivity for the detection of human hepatotoxins.
Notoriety bias in a database of spontaneous reports: the example of osteonecrosis of the jaw under bisphosphonate therapy in the French national pharmacovigilance database.
Curation and analysis of clinical pathology parameters and histopathologic findings from eTOXsys, a large database project (eTOX) for toxicologic studies.
Cox MJ, Jaensch S, Van de Waeter J, et al. Tales of 1,008 small molecules: phenomic profiling through live-cell imaging in a panel of reporter cell lines. doi:10.1101/2020.03.13.990093
Smalheiser NR, Bonifield G. Two similarity metrics for medical subject headings (MeSH): an aid to biomedical text mining and author name disambiguation. doi:10.1101/039008
Relating chemical structure to cellular response: an integrative analysis of gene expression, bioactivity, and structural data across 11,000 compounds.
13714-7&id=gr1.jpg){kind=link}
13714-7&id=gr2.jpg){kind=link}
13714-7&id=gr3.jpg){kind=link}
13714-7&id=gr4.jpg){kind=link}
13714-7&id=gr5.jpg){kind=link}
13714-7&id=gr6.jpg){kind=link}
13714-7&id=gr7.jpg){kind=link}
13714-7&id=gr8.jpg){kind=link}