Quantitative systems toxicology modeling in pharmaceutical research and development: An industry‐wide survey and selected case study examples

Abstract

Quantitative systems toxicology (QST) models are increasingly being applied for predicting and understanding toxicity liabilities in pharmaceutical research and development. A European Federation of Pharmaceutical Industries and Associations (EFPIA)‐wide survey was completed by 15 companies. The results provide insights into the current use of QST models across the industry. 73% of responding companies with more than 10,000 employees utilize QST models. The most applied QST models are for liver, cardiac electrophysiology, and bone marrow/hematology. Responders indicated particular interest in QST models for the central nervous system (CNS), kidney, lung, and skin. QST models are used to support decisions in both preclinical and clinical stages of pharmaceutical development. The survey suggests high demand for QST models and resource limitations were indicated as a common obstacle to broader use and impact. Increased investment in QST resources and training may accelerate application and impact. Case studies of QST model use in decision‐making within EFPIA companies are also discussed. This article aims to (i) share industry experience and learnings from applying QST models to inform decision‐making in drug discovery and development programs, and (ii) share approaches taken during QST model development and validation and compare these with recommendations for modeling best practices and frameworks proposed in the literature. Discussion of QST‐specific applications in relation to these modeling frameworks is relevant in the context of the recently proposed International Council for Harmonization (ICH) M15 guideline on general principles for Model‐Informed Drug Development (MIDD).

BACKGROUND

Quantitative systems toxicology (QST) models provide a quantitative approach to predict multiscale biological responses, often using predicted drug exposure, and evaluate pharmaceutical drug safety risks based on predicted adverse outcomes in response to exposure or target modulation assumptions. QST models can explore both on‐target and off‐target toxicities. In recent years, the development of several new QST models stemmed from initiatives driven by regulatory agencies, academia, and private and public partnerships. The TransQST (http://transqst.org) project was launched in 2017 by the European Innovative Medicines Initiative. ¹ It focused on developing QST models in the cardiovascular, liver, kidney, and gastrointestinal tract/immune organ systems. DILIsym ² is a QST model predicting drug‐induced liver injury liability for new drug candidates and is now an integral part of decision‐making within some pharmaceutical companies with findings being included in regulatory communications and new drug application submissions. ³ Another initiative, focused on the prediction of the proarrhythmic potential for new drugs, is the comprehensive in vitro proarrhythmia assay (CiPA), jointly sponsored by the United States Food and Drug Administration (FDA), the Health and Environmental Sciences Institute (HESI), and the Cardiac Safety Research Consortium (CSRC). ⁴ , ⁵ Alongside such efforts, literature focused on frameworks for building and validating computational models continues to emerge. ⁶ , ⁷ , ⁸ , ⁹

Considering these advances in systems modeling for predicting safety endpoints, it is timely to bring together the EFPIA community of QST modelers. The EFPIA QST working group was established through the EFPIA Safety Reflections initiative to (i) understand the use of QST models across EFPIA companies, (ii) share learnings from incorporating QST models for use in decision‐making, and (iii) reflect on the experience of developing and validating QST‐specific examples in the context of proposed modeling frameworks. These aims are relevant in the context of the International Council for Harmonization (ICH) M15 guideline on general principles for Model‐Informed Drug Development (MIDD). ¹⁰ , ¹¹

To address these aims, an EFPIA‐wide survey on QST model use was conducted in Q4 2022. The survey was circulated to EFPIA Preclinical Development Expert Group (PDEG) representatives who directed it to relevant colleagues within their companies. A copy of the survey questions is included in the Supplementary Material. The survey was available for responses for 9 weeks. Responses were collected via Microsoft Forms and were anonymized by the EFPIA Secretariat before sharing with the EFPIA QST working group for analysis. Multiple responses per company were welcomed and received from some companies. 25 responses were collected including representatives from 15 EFPIA companies. These survey findings provide a QST‐specific perspective to complement Quantitative Systems Pharmacology (QSP) focused and other modeling surveys conducted previously. ¹² , ¹³

This article describes the results of the EFPIA‐wide survey on QST model use as well as the discussion of QST case study applications within EFPIA companies.

Frameworks for mechanistic pharmacokinetic and pharmacodynamic model building and model credibility assessment

As QST model use becomes more prevalent in the pharmaceutical industry, it is crucial to establish assessment criteria to appropriately evaluate confidence and communicate model risks before use for decision‐making. Recent foundational papers ⁶ , ⁷ , ⁸ , ⁹ and references therein, propose credibility frameworks and considerations for models used in drug development, incorporating risk‐based verification, validation, and uncertainty quantification similar to that described for medical devices (ASME V&V40‐2018) in https://www.fda.gov/media/154985/download.

Kuemmel et al. ⁸ presented a model credibility framework for Physiologically‐Based Pharmacokinetic (PBPK) models, emphasizing the importance of defining the question, context of use (COU), and potential decision consequences. Model risk is established by considering a matrix of (i) model influence, that is, how much weight the model predictions have on the decision, and (ii) potential decision consequences, that is, what consequences a wrong decision based on the model outcome could have. Model acceptance criteria are then defined, including verification of the software code and calculations, and validation of the model's predictive ability against defined quantitative and qualitative observations. Finally, model credibility is assessed considering the aforementioned factors. ⁸

Musuamba et al. ⁶ proposed a “credibility matrix” for assessing the credibility of mechanistic in silico models with the feasibility assessed through application to case studies. The authors acknowledge that the framework is not intended as a definitive recommendation for regulatory evaluation, but it is considered broadly applicable to facilitate scientific communication between modelers, both for model application and when preparing models for regulatory submissions. Regulatory impact is included in the credibility matrix, and it allows an indication of the model's role against other data supporting the decision. This article also includes a QST case study for cardiac safety. ⁶

Braakman et al. ⁹ consolidated QSP model evaluation methods into a comprehensive framework. Friedrich ¹⁴ proposed a QSP model qualification framework focusing on four categories: model relevance to the biological question of interest, model uncertainty, model variability, and ability of the model to reproduce data. Cucurull‐Sanchez et al. ⁷ offered best practices for maximizing QSP model use, covering topics including model context, through defining input data and model assumptions, and finally, model validation and verification.

These frameworks could inform future regulatory guidance (ICH MIDD roadmap). ¹¹ The ICH MIDD M15 Expert Working Group is developing an overarching MIDD guideline ¹⁰ , ¹¹ which focuses on a risk‐based assessment of MIDD approaches used in regulatory decision‐making. We reflect on specific QST case study examples described in this manuscript, as well as the findings from the EFPIA survey on QST use, in light of proposed frameworks (in particular, Kuemmel et al. ⁸ ), and published recommendations for mechanistic model building and application.

SURVEY FINDINGS

Survey methods and responder demographics

Survey results were analyzed using MATLAB (versions 2021b and 2022b). Most respondents (36%) were scientists within modeling departments, followed by drug safety/toxicology (32%) (summarized in Figure S1). Among those who themselves build, develop, or apply QST models (either directly or through those reporting to them), 44% were from a dedicated modeling department (Figure S2). Most respondents were involved in modeling, either as hands‐on modelers (52% as expert or intermediate) and/or managers of modelers (44%) (Figure S3). Non‐modelers mostly identified as modeling advocates (32% of total responses). QST was the most popular modeling type performed (18% responses), followed by translational modeling and PK/PD (16% each), and QSP (12%) (Figure S4).

Definition of quantitative systems toxicology

Various definitions have been presented in the literature to define QST including describing QST as a prototype for toxicity assessment with a focus on understanding drug adverse effects across multiple scales of biological organization to link the phenotypic observations. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸

Responses to the EFPIA survey (summarized in Figure 1) identified “mechanistic,” “integrates experimental data,” “includes toxicity endpoints,” and “quantitative” as “essential” QST model features. “Multiscale,” “excludes pharmacology endpoints,” “includes population variability,” and “includes uncertainty quantification” were indicated as “preferred” features. “Includes PK/PBPK” was identified by responders equally as “essential” and “preferred.” A model being translatable was a popular additional feature offered as “essential” by responders.

FIGURE 1.

Characteristics of a QST model. Model features are considered “essential” and “preferred” characteristics of QST models. The number of responders indicating an option as “essential” is shown in blue and as “preferred” is shown in red. PK, Pharmacokinetics; PBPK, Physiologically‐based pharmacokinetics.

QST is sometimes identified as a form of QSP and other times as a discipline in its own right. ¹⁶ The overlap between QST and QSP has been discussed in the past, ¹⁶ with both indicated as providing quantitative descriptions of the mechanism of drug action. This aligns with the high number of survey respondents who deemed “quantitative,” “mechanistic,” and “integrates experimental data” as essential features of QST models. 56% of responders suggested that QSP and QST are not interchangeable, 40% that they are, and 4% were unsure (Figure S6). This close split may be due to different QST approaches and respondents' experiences. When QST models focus on on‐target toxicity predictions, they may be considered more interchangeable with QSP models. QST and QSP modelers could be part of the same department or separate, and these factors may also contribute to the perception of interchangeability between QSP and QST. The distinction between QSP and QST may primarily be the output ranges for clinical biomarkers and endpoints, with QSP examining safe ranges and QST focusing on toxicity outcomes when predicted outputs are outside the normal range or supratherapeutic doses are being explored.

Responders suggested the main differences between QST and QSP models are “type of modeling endpoints,” “ways in which a model can be applied,” “potential impact,” and “way of use as part of a regulatory submission” (Figure S7). Recognizing these differences and the focus of previous similar surveys being primarily on QSP applications, with more limited reference to applications of systems modeling to predict safety outcomes, ⁶ it is timely to add a QST perspective to the literature.

Overview of survey results for QST model use

The survey revealed that QST models are more commonly used in larger companies (>10,000 employees), with 73% using QST models, than smaller ones (100–10,000) employees, with only 25% using QST models (Figure 2a). Figure 2b shows usage and interest in specific organ system QST models. To avoid bias from multiple responses about the same model from one company, respondents were asked to have the most appropriate colleague answer questions about specific organ system QST models. Despite this, repeat responses from a company could potentially concern the same or different models for the same organ system.

FIGURE 2.

QST model use. (a) Use of QST models according to company size of the responder (red indicates a response of “yes” QST models are used in the responder's company and blue indicates a response of “no”), (b) Overview of the level of interest in different organ system QST models by survey responders.

Cardiac electrophysiology, liver, and bone marrow/hematology were the organ system QST models with the highest number of EFPIA companies indicating active use (Figure S8). Cardiac mechanics, cardiovascular hemodynamics, kidney and immune‐related were the models the highest number of companies indicated are currently in development. Two companies indicated interest in lung models, with one company having a model in development. Interest was indicated in a skin model, but currently with no such QST model in use within that company.

Most respondents expressed interest in organ system QST models but currently lacked them, potentially indicating the expected trajectory of growth of QST modeling applications. The interest level was highest for CNS (7 companies), kidney (8 companies), and cardiac mechanics models (5 companies). Only a few organ system QST models are used routinely including cardiac electrophysiology (4 companies), liver (3 companies), and bone marrow/hematology (2 companies), potentially reflecting the maturity of the modeling field for those organ systems relative to others where models are not yet used routinely (for example CNS, gastrointestinal and immune‐related). For every organ system, at least one respondent indicated no interest or model, with five companies indicating this response for CNS. Elements of these results are contradictory—for example with CNS simultaneously having the highest percentage of responders indicating interest but no QST model and no interest and no QST model. However, other findings are as expected, for example, the organ systems which more responders indicated as being actively used within their companies (cardiac electrophysiology, liver, and bone marrow/hematology) have comparatively fewer responders indicating they have no model. These results may be influenced by the company's portfolio, existing safety strategies, and the perceived value of QST by responders.

Details of use of organ‐specific QST models

Figure 3 summarizes survey responses related to the use of QST models including frequency and timing of use, toxicity mechanisms represented, model inputs and outputs, and identification of modeling opportunities for the three most used organ system models: liver, bone marrow/hematology, and cardiac electrophysiology. The complete results for all organ systems are shown in Figures S9–S13.

FIGURE 3.

Use of organ‐specific QST models. Details of QST models currently used within EFPIA companies. Questions asked of the responders are shown in dashed circles around each organ system. Responses to each question are represented in individual pie charts with the corresponding numbered rectangle. Percentages in the pie charts represent the number of responders choosing the predefined or responder provided answers to these questions. An overview of survey results related to the use of QST models for (a) cardiac electrophysiology, (b) liver, and (c) bone marrow/hematology are highlighted as these were the models indicated as being in active use by the largest number of companies.

Responders identified the toxicity mechanisms represented in QST models where they were in active use. For cardiac electrophysiology models, five companies selected ion channel block. Liver models had a range of different mechanisms represented including cytotoxicity, disruption of bile acid homeostasis, inflammation, mitochondrial dysfunction, and oxidative stress. For bone marrow/hematology models, five companies reported cytotoxicity as a mechanism represented in the model.

Various strategies are used to identify opportunities to apply QST models. These range from ad hoc, indicated by four companies for bone marrow/hematology models, to being determined based on compound indication indicated by three companies for cardiac electrophysiology models, to all lead compounds considered for modeling indicated by three companies for liver. Some companies apply QST models to all lead compounds, 2 companies for cardiac electrophysiology models and 1 company for liver models. The strategy employed will likely vary based on the toxicity type the model explores—some on‐target toxicities may be more applicable for specific indications (e.g., cytotoxicity or disruption in cell cycle from oncology compounds) while other off‐target toxicities may have the potential to be triggered by any small molecule compound (e.g., disruption to cardiac electrophysiology through ion channel block).

QST models are used in varying frequencies both within and across organ systems. Three companies indicated that they apply cardiac electrophysiology models to over 10 compounds per year. QST models are used for 2–10 compounds per year by 4 companies for liver and 3 companies for bone marrow/hematology. 2 companies for both liver and bone marrow/hematology apply the models to 0–1 compounds per year. The frequency is likely linked to the strategy for QST model application—for QST models applied to all lead compounds there will likely be more model applications per year than models which are only applied to compounds with a specific indication or where opportunities are identified by chance.

The survey indicated QST model applications in both preclinical and clinical stages of drug development. QST models were suggested as mainly used in preclinical and discovery stages, with 60% for cardiac electrophysiology, 61% for liver, and 60% for bone marrow/hematology. Only cardiac electrophysiology models were indicated as being used to support a compound post‐approval. Immune‐related QST models were identified as being mostly used to support clinical development (Figure S11). The survey distribution method may have influenced these results, while it was asked that the survey was distributed to appropriate individuals across responder companies, regardless of the stage of research and development they work in, the survey was disseminated via PDEG members, which may favor preclinical applications. Or these results simply could reflect that there are more QST model applications in a preclinical setting—explained by a strong intent to gain as much understanding as possible of patient safety liability before a drug enters clinical development, recognizing that QST models are promising tools for this.

Figure S15 shows that QST models are most used for internal decision‐making (50–100% of responses, depending on the organ system). Clinical decisions are also informed by QST models, with 10% of responses for liver models. Regulatory decisions are informed by cardiac electrophysiology QST models (22% responses), bone marrow/hematology QST models (20% responses), and liver QST models (10% responses). The relative maturity of the organ‐specific QST models may influence their usage, with mature models like cardiac electrophysiology, liver, and bone marrow/hematology impacting clinical and regulatory decisions. In contrast, models introduced more recently, like gastrointestinal and immune‐related, currently used for internal decisions, may support clinical and regulatory decisions as they mature and gain acceptance.

A variety of inputs have been reported for organ system QST models (Figure S13): QSAR model inputs (0–12%), pharmacokinetic information (11–27%), literature/public data (9–19%), clinical data (6–29%), in vivo biomarkers/endpoints (6–27%), high‐throughput data (0–12%), and in vitro data (9–29%). It is noted that the different data types are used in similar frequencies in QST model applications. Omics data are used only for gastrointestinal (9%) and liver (4%) models, its usage may increase as more QST modeling applications using omics data are published and as these approaches are adopted within pharmaceutical settings. ¹ , ¹⁹ , ²⁰

A variety of QST model outputs have been reported (Figure S13) including time course and concentration‐response biomarker predictions and predicted toxic dose. Except for cardiac electrophysiology and kidney models, outputs also included the likely affected population percentage.

The survey showed diverse software used for QST model development (Figure S17). MATLAB and SimBiology were commonly used across organ systems. These findings are not dissimilar to the survey findings exploring software used for QSP modeling applications reported in ¹² . Some companies use executables for model deployment, which may indicate the provision of models for use by non‐modelers. Other software not listed as options but indicated as used included proprietary, SAS and ADAPT (for bone marrow/hematology models) and Heta (for immune‐related models). Some responders did not know the software used, possibly due to indirect model interaction or lack of knowledge of the technical details.

The survey revealed diverse origins of QST models (summarized in Figure S18). Models were built internally (17–50% responses depending on the organ system) and developed as part of consortia for most organ systems, the highest percentage being for gastrointestinal (43% responses) and kidney (40% responses). Some models were developed through academic collaborations, with the only organ systems indicating this was not a means for model development being CNS, kidney, and immune‐related systems. Some models were bought or licensed from external companies, the highest percentage of which was 24% responses for liver, and some, including bone marrow/hematology (7%) and CNS (50%), were obtained through Contract Research Organization (CRO) contracts.

Challenges associated with QST model use

The survey explored challenges in implementing QST models, identifying obstacles to wider use of each organ system QST model (summarized in Figure 4).

FIGURE 4.

Obstacles to QST use and impact. Obstacles to wider use and impact of QST models across organ systems as selected by survey responders.

Promisingly, at least one respondent for each organ system found no obstacles to wider QST model use or impact. Common obstacles identified included resource limitations and/or modeling expertise to fulfill all QST modeling opportunities, lack of internal models, and data needed for model input and/or validation being limited or unavailable. Lack of safety needs was another obstacle indicated for all organ systems except the liver. Additional obstacles provided by responders included prohibitive costs and time for data generation and reporting, poor model prediction, misalignment between clinical and preclinical modelers, and limited model knowledge in therapeutic areas.

The time taken to prepare a QST model for the first application varies across different organ systems and among different responders for the same organ system (Figure S16). For cardiac electrophysiology models, arguably the most mature QST models, it took 2–5 years for 60% of responders to use the models in decision‐making. For liver models, it took less than a year for 43% of responses and kidney models over 5 years for 50% of responses. Differences in time could reflect the relative maturities of the various organ system models. Many responders were unsure of the time frame from initial interest to decision‐making use, possibly due to personnel changes.

To complement the insights into QST use gained from the survey, we next discuss four case studies of QST model use within EFPIA companies represented by manuscript authors. More details on these features of the case studies can be found in Tables S1 and S2.

CASE STUDIES OF QST MODELS USED IN EFPIA COMPANIES

Drug‐induced crystalluria

Drug‐induced crystalluria (DICU) results from an interplay between drug plasma exposure, ²¹ physicochemical properties of a drug, and physiological properties of the kidney. ²² When exposure exceeds its solubility limits, a drug may precipitate in nephrons, leading to obstruction, physical damage, and thus kidney toxicity. ²¹ , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ To mitigate the risk of DICU and help with chemical design at the early stages of drug discovery, a physiologically‐based in silico model that predicts DICU in rats, dogs, and humans was developed. ²² , ²⁵ The model incorporates key features of kidney physiology with species‐specific parameterizations, uses physicochemical properties of a drug, and maximal in vivo plasma concentration of a drug as model inputs to predict the risk of DICU. The model has been calibrated using urinary excretion data for 42 drugs collected in rats, dogs, and humans. The forward predictability of the calibrated model was evidenced by a high alignment between model predictions and experimental observations for 13 compounds known to be associated with DICU (e.g., sulfadiazine and acyclovir) and 7 compounds that served as negative controls (e.g., amitriptyline and miglitol). The model was used retrospectively to analyze if a drug and/or its metabolite(s) will cause DICU in animal studies or human trials. The model was also used prospectively around the lead optimization stage to understand the risk of DICU, to design experiments for an in vivo validation, and to prioritize compounds based on these results. Full model details are available in ²² , ²⁵ .

Drug‐induced changes in cardiac electrophysiology

Perturbations to normal cardiac electrophysiology function, due for example, to block of cardiac ion channels, can lead to the risk of life‐threatening arrhythmias. Single‐cell cardiac electrophysiology models can be used to predict changes in cellular cardiac electrophysiology arising from drug block of key cardiac ion channels (hERG, NaV1.5, CaV1.2, plus IKs if data are available). Input data take the form of IC₅₀ values quantifying drug inhibition of ion channels and predictions take the form of changes in cardiac action potential at different drug concentrations. The model integrates information on ion channel block to predict changes in cardiac cellular electrophysiology. A rabbit‐specific cardiac electrophysiology model ²⁸ was validated, with predicted changes in action potential at different concentrations compared with experimental measurements of QT interval from the ex‐vivo rabbit ventricular wedge assay. Model predictions were evaluated using input ion channel data from high‐throughput and manual patch clamp ion channel assays and making different assumptions related to the use of available input data. Full model details are reported in ²⁹ , ³⁰ . The model output is used as part of a body of nonclinical evidence when compounds are considered for progression. The model outputs can also be used to inform prioritization of compounds to balance resource. The model outputs may be confirmatory of understanding and mechanisms (supporting evidence from in vivo and/or in vitro assays), or, if contradictory, may suggest the repetition of assays or inform the experimental screening cascade for further investigative studies.

Drug‐induced neutropenia

For drug targets where neutropenia is a known on‐target mechanistic toxicity, it is important to determine the probability and severity of neutropenia in drug discovery/early clinical development to influence internal decisions and early clinical trial designs. To support this objective, a translational, physiological model of granulopoiesis and its regulation was developed that utilizes drug‐specific in vitro human bone marrow mononuclear cell proliferation data and relevant clinical pharmacokinetic profiles (predicted or measured) to predict the time course of absolute neutrophil count (ANC) and the incidence of neutropenia. ³¹ The ODE model was built from previous publications that informed on granulocyte disposition and neutrophil margination in the presence of endogenous granulocyte colony‐stimulating factor (G‐CSF) and exogenous G‐CSF therapies (for example ³² ). The model includes a representation of bone marrow progenitor cell cycle, allowing for a mechanistic representation of the action of relevant anticancer drugs based on in vitro results specific to each drug of interest. Initially, the model was evaluated for the cell cycle inhibitor palbociclib, using an in vitro system of human bone marrow mononuclear cells to quantify the action of palbociclib on proliferating progenitor cells. The model results were compared with neutropenia observed in three previously reported clinical trials with palbociclib and the model was able to predict grades 3 and 4 neutropenia with 86% accuracy. The model was further validated with additional chemotherapeutics and combination therapies. ³³ Building a translational model that links clinical outcomes to a preclinical relatively high‐throughput assay allows teams to evaluate the likelihood of achieving efficacious and safe concentrations in humans with preclinical compounds, which aids in internal resource prioritization decisions. It also has been used for compound selection and risk/benefit analysis. Clinically, it can inform early clinical study designs and be used to explore neutropenia risk for drug combinations or alternative dosing regimens.

Drug‐induced liver injury

DILIsym® (SimulationsPlus) is a “middle‐out,” multiscale representation of drug‐induced liver injury. It includes key liver cell populations (e.g., hepatocytes, Kupffer cells), bile duct cells (cholangiocytes), intracellular biochemical systems (e.g., mitochondrial function), and whole‐body dynamics (e.g., drug distribution and metabolism). ³⁴ DILIsym was used to investigate liver enzyme increases observed during challenge trials of ferroquine against blood‐stage malaria. And 2 of 8 subjects treated with 800 mg ferroquine experienced significant increases in alanine aminotransferase (ALT) and aspartate aminotransferase (AST) within a few days of treatment. ³⁵ Both subjects also received acetaminophen during the treatment period. DILIsym was used to investigate the potential interaction between ferroquine and acetaminophen and to identify patient‐specific risk factors for liver injury. Input data for modeling included basic molecular properties of the drugs, pharmacokinetics (PK) and metabolism data, animal biodistribution, and in vitro mitochondrial function and transporter activity. Modeling indicated that increases in reactive oxygen species by both compounds in combination could lead to increased aminotransferase levels, exacerbated by increased hepatocellular apoptosis due to malaria infection and low body weight and nutritional deficit in treated patients. Modeling further demonstrated that the risk of severe liver injury was low and could be mitigated. The model output was used to provide confidence in the clinical safety of ferroquine, to demonstrate the synergistic effect of acetaminophen coadministration leading to increased liver enzyme levels, and to adjust clinical trial protocols to further protect patient safety by limiting the dose of acetaminophen to less than 40 mg/kg/day.

QST model use in EFPIA companies

QST models can be used at various stages of drug discovery and development and their use can be triggered in different ways as summarized in (Figure 3 and Figure S9) and (Figure 5 and Table S1). They can be used as standard screening models to assess safety liabilities for many compounds (typically of a particular modality) once data are available (e.g., high‐throughput ion channel screening data for the cardiac electrophysiology case study). They can be used to investigate specific questions after an observation (e.g., a clinical safety finding triggering model application for the liver case study). Model use can also be triggered by a drug intended to treat a particular indication which leads to an on‐target toxicity concern (e.g., neutropenia case study) or specific triggers related to compound properties or observations in experimental studies (e.g., crystalluria case study).

FIGURE 5.

Examples of QST model uses. Selected examples of how, when, and why QST models can be used in pharmaceutical discovery and development including stages at which models are applied, the trigger for the use of the model, types of decisions the model is used to inform, extent of validation of the model and the influence the model has. Model influence examples can occur at multiple stages during the drug development process but are only included once in this figure for clarity.

Screening models, like the cardiac electrophysiology case study, are used early in drug development, requiring high‐throughput data for inputs which are readily produced as part of standard drug development pipelines. Their building and validation may differ from models exploring a specific on‐target toxicity, like the neutropenia case study, or which are built responding to a safety finding, as for the liver case study. Screening‐type QST models are often validated using large compound sets, which is feasible if the input data and an appropriate comparator for predictions are readily available. Models applied to identify or understand a particular safety mechanism for individual compounds focus validation on molecules with the same target or same mechanism. Models applied in clinical development stages need high confidence in output for that application, not broad predictivity but where lower confidence may be acceptable such as for screening‐type models applied in discovery or preclinical development.

For QST models like the neutropenia case study, focusing on a specific on‐target toxicity, building, and validation might be expected to align with that of typical QSP models, since the aim is to model target‐specific behavior of the drug. The key difference is that QST models emphasize toxicity mechanisms and endpoints over efficacy.

For QST model applications triggered by concerns related to the physical chemical properties of a compound or a tissue‐specific exposure concentration (for example, the crystalluria case study), there may be more emphasis placed on ensuring that the PK or PBPK model concentration input to the QST model is as representative and predictive as possible, or at least ensure that any uncertainties in that component are appropriately explored.

QST MODEL VALIDATION, IMPACT, AND DECISION CONSEQUENCE

Before using a QST model for compound safety decisions, it must be sufficiently validated for the decision it is being used to inform. This aligns with risk‐based modeling frameworks proposed in the literature (such as Kuemmel et al. ⁸ ). Our case studies and survey responses highlight impacts of QST models in preclinical and clinical settings and the consequences of incorrect model predictions. For the case studies, the required validation level before the model's first application is described in Table S2. Tables S1 and S2 form an adapted version of the credibility matrix presented by, ⁶ outlining the models' COU, applicability domain, decisions informed, and validation steps. The categories selected to present case studies in this manuscript include elements from publications such as those described above. We chose to document one category on model evaluation and assessment criteria. Following a risk‐based model analysis approach we chose to separately document model influence and decision consequence. We chose also to include details about the model, including its trigger for use, the stage of drug development the model is applied and technical details of the model including mathematical model type and toxicity mechanisms represented. We do not include the category of regulatory impact as suggested in ⁶ since for several of the case study examples, it is not simple to define the regulatory impact due to the timing of their application (this may however be realized at a later stage).

The liver case study used a commercially licensed QST model (DILIsym), validated through consortium and EFPIA company activities. Other case studies, like the cardiac electrophysiology model, were developed through industry‐academic collaborations, requiring external data sharing for validation. Some models, like those for the crystalluria and neutropenia case studies, were developed internally by EFPIA companies. QST models are typically acquired from various sources, each requiring unique considerations for model building and validation. As summarized in Figure S20, models acquired from consortia or external companies often rely on literature or commercial activities for validation. For all sources of QST models, models are typically tested with reference drugs and at least one proprietary compound.

The validation level varies most in internal applications, indicating diverse internal uses and requirements, or possibly nonsignificant findings due to limited data. Figure S21 summarizes validation levels for all organ system models. The survey findings did not suggest specific validation levels for decision types, but responses along with the case studies provide examples to be considered as this is further explored in future modeling guideline development. The case studies illustrate how QST models inform drug discovery and development decisions. They predict potential safety issues early in development (crystalluria case study), which may result in the reduction of animal use or optimize safety screening strategies (cardiac electrophysiology case study), prioritize resources, and influence clinical trial design (neutropenia). A central use of QST models, as shown in the neutropenia case, is the translation of in vitro or in vivo toxicity observations to humans.

The case studies demonstrate the ability of QST models to explain toxicity mechanisms following clinical (liver case study) or preclinical (crystalluria case study) observations when used retrospectively. They can also contribute to compound termination decisions due to safety, impact preclinical and clinical study design, provide 3Rs benefits (reduction, refinement, replacement), and identify patient risk factors in clinical settings.

Various consequences of incorrect QST model predictions are evident in the case studies. For the liver case study, a wrong prediction could lead to adverse human outcomes, the most severe consequence of incorrect QST model prediction. However, QST predictions often contribute supportive evidence rather than being the sole basis for decision, reducing the impact of incorrect predictions, especially when applied in early drug discovery. Incorrect predictions can also lead to significant practical outcomes, such as wasted resources or delays in medicines reaching patients due to pursuing compounds with a lower probability of success. Other consequences include suboptimal resource, time, or animal use allocation, misinformation, or unnecessary protocol alterations. These risks can be business‐critical. The model's influence and the consequences of incorrect predictions are tied to their contribution to the overall evidence, which should be considered in a risk‐based modeling framework.

While each case study was independently applied without following a specific modeling framework, they implemented many best practices described in the literature, ⁶ , ⁷ , ⁸ , ⁹ including defining the model's COU, conducting credibility assessments including code verification and model validation, independent training and test data sets and applying model parameterization where required, clearly outlining model assumptions, documenting and publishing model and validation results, and sharing code for reproducibility. Some of the indicated best practices were performed to an extent, however, on reflection, more formal guidelines for model verification, risk assessment, uncertainty quantification, and reproducibility could be beneficial.

Aligning the credibility and validation rigor appropriately with the level of influence and consequence of the model predictions is critical.

The case studies placed more emphasis on model selection and context‐specific validation approaches than existing literature. Although it is acknowledged that a section is dedicated to model selection in Braakman et al. ⁹ Three of the four case studies highlight model influence in preclinical settings, contrasting with the literature's clinical focus. Future guidelines could benefit from exploring requirements for coupling PBPK and QST models for tissue‐specific toxicity predictions as these considerations may differ from QSP model applications for example.

CHALLENGES

As with other MIDD disciplines, QST model use in pharmaceutical research and development presents challenges. As QST and related fields like PBPK and QSP are still maturing, there is a basic challenge of inconsistent terminology across these disciplines, especially regarding terms like virtual and plausible patients or populations, which can lead to methodological and technical inconsistencies.

Practical challenges often relate to data—including structure and storage, and accessibility of data both for model inputs as well as validation. Ideally, automated workflows transfer input data directly from its primary location to the model. However, difficulties in locating and understanding data structure, along with incomplete data sets in digital format, can delay establishing efficient data flows. Having the necessary data inputs and sufficient compound examples for model calibration and testing is crucial. Another challenge is finding suitable validation data, especially when predicting clinical outcomes with limited available data. This is particularly relevant for QST models predicting rare adverse outcomes.

Some QST methods, translating in vitro or in silico data to human, may require an intermediate validation step. If a preclinical in vivo outcome is predicted, the model's accuracy can increase confidence in human translation. This approach may be appropriate when no clinical data are available for model validation, in a prospective use case, and where species differences are accounted for in the model.

The survey further explored challenges in implementing QST models in EFPIA companies, asking respondents to identify obstacles to the wider use of each organ system QST model, with resource constraints indicated as a primary reason. A 2019 survey by Ermakov et al. ¹² highlighted similar obstacles.

Building and implementing QST models is a time‐consuming process with survey findings indicating that some models take more than 5 years from interest to first application of a QST model and use in decision‐making.

LEARNINGS

Learnings from the case study examples and reflections on generalized modeling frameworks

The four case studies (cardiac electrophysiology, crystalluria, neutropenia, and liver) did not follow a specific QST model framework as none existed at the time of implementation. They adhered to good modeling principles and implemented several elements identified in the recent publications ⁶ , ⁷ , ⁸ , ⁹ relating to modeling frameworks and best practices. This included defining the model's use context and validation against experimental data (with the intended model use in mind). However, they did not explicitly perform sensitivity analyses, identifiability analyses or uncertainty quantification, which could be a limitation as such analyses can be critical. Most survey responders considered uncertainty quantification a “preferred” rather than “essential” feature of a QST model, however. Model risk assessment was also not formalized. Instead, an informal judgment was made, and model assessment was performed accordingly to ensure that there was sufficient confidence in the model prediction for the decision informed. Future modeling guidance could benefit from not only focusing on clinical outcomes but also considering the impact on business decisions throughout research and development, as these decisions affect the delivery of safe and effective medicines. By sharing the challenges in implementation of QST case studies, we aim to inspire solutions for these issues and facilitate smoother integration of future QST models in pharmaceutical research and development.

Learnings from the EFPIA‐wide survey on QST model use

The survey indicated broad interest in QST models across EFPIA. Cardiovascular electrophysiology, liver, and bone marrow/hematology models are actively used by most companies. Over 50% of responders showed interest in CNS and kidney models, despite not currently having them, indicating potential areas for future QST model development. Additional interest was noted in lung and skin models.

QST models support decisions in both preclinical and clinical stages of pharmaceutical development. Models are used with varying frequency, opportunities for use are identified in different ways, and model inputs and outputs take various forms.

Resource constraints, including personnel and infrastructure, currently limit the broader use and impact of QST models, suggesting a need for further industry investment. This also presents an opportunity for educational providers to prepare students for quantitative, multidisciplinary industrial work through tailored programs, potentially industry‐sponsored, to equip future quantitative systems toxicologists.

RECOMMENDATIONS

The survey and case studies provide key insights for QST modelers on practices and recommendations for future model development in the pharmaceutical industry. Common use of MATLAB and SimBiology were highlighted, which may encourage further code sharing in these languages. However, it is important to highlight that a wide range of different software has been adopted across the industry over the years, ¹² and based on our collective industry experience, it may be that we see other historically less used software types becoming more frequently used in coming years, particularly those which are open source. ³⁶ As noted from the survey, the demand for QST models surpasses the available resource and the ability to implement them. Investment in QST resource and education is crucial for strengthening the QST community and ensuring the field's future success and impact. Reflections on validation approaches and the relevance of modeling frameworks to QST applications can guide the community toward well‐accepted practice as it develops and applies QST models within the pharmaceutical industry. While many aspects of published modeling frameworks ⁶ , ⁷ , ⁸ , ⁹ apply to QST models, we highlight some QST‐specific aspects for consideration and incorporation as part of future QST modeling frameworks or guidelines:

• (i)QST models, used for predicting toxicities and impacting various drug discovery stages, can predict on‐ and/or off‐target toxicities (with different mechanistic representations within the models), can take various forms (ODE‐based, agent‐based, flux balance, partial differential equation (PDE)‐based as well as others) and may be used for different purposes and impact at different stages of drug discovery and development pipeline. This flexibility means each model type may need unique building and validation approaches.
• (ii)QST is an end‐to‐end approach with the potential for significant preclinical or clinical influences. Safety is a primary contributor to compound attrition, therefore, timely identification of safety concerns early on in drug development is hugely beneficial from a business perspective. A single QST model may serve distinct purposes at different stages, requiring independent validation approaches.
• (iii)QST aims to predict toxicity outcomes, and as a result, some uncertainty may be acceptable (although it should be appropriately explored). QST approaches typically adopt a conservative approach to prediction—final predictions should represent worst‐case scenarios (which are the ones leading to a higher level of toxicity) due to the safety implications of incorrect predictions, which may be a different approach to other mechanistic modeling fields.
• (iv)Similarly, while an incorrect model prediction of efficacy is undesirable, an incorrect model prediction of toxicity can be unacceptable. Therefore, one may place more emphasis on model validation for QST models in a preclinical setting before predicting human outcomes than for other mechanistic modeling applications. However, the business consequence of an over‐prediction of toxicity may lead to the termination of a compound (although this decision would likely be taken considering additional evidence and not the modeling prediction alone).
• (v)Validating QST models predicting rare toxic side effects can be challenging due to infrequent occurrences or being idiosyncratic in nature, and this should be considered in the validation strategy.
• (vi)Model reproducibility is crucial for transparency and trust. QST models can be adapted or expanded to other indications or organ systems or incorporate additional toxicity mechanisms.

Outlook for the future

Results of an EFPIA‐wide survey confirm the growth of QST as a field, with increasing applications in pharmaceutical research and development. While case studies presented focus on small molecules, we anticipate more QST models for different modalities, as well as hybrid approaches with AI/ML, and increased omics data integration. QST models could expand into digital twins for drug safety and precision medicine, and efforts may shift toward connecting multiple organ systems. An NC3Rs CRACK‐IT challenge is currently exploring this in the form of a virtual dog for drug safety testing. ³⁷

FUNDING INFORMATION

No funding was received for this work.

CONFLICT OF INTEREST STATEMENT

Authors are employees (and may be shareholders) of pharmaceutical research and development companies which may have a potential commercial interest in the subject matter of this manuscript.

Supporting information

Data S1:

Data S2:

Beattie KA, Verma M, Brennan RJ, et al. Quantitative systems toxicology modeling in pharmaceutical research and development: An industry‐wide survey and selected case study examples. CPT Pharmacometrics Syst Pharmacol. 2024;13:2036‐2051. doi: 10.1002/psp4.13227