Toxicity refers to the extent to which a substance can cause harm to living organisms, including animals, plants, bacteria, and humans (Duffus, 1993; McNaught and Wilkinson, 2025). While many chemicals enhance our quality of life, they can also pose significant toxic risks. To ensure public safety, various regulatory frameworks have been established to mitigate these hazards. Given the potential health risks associated with chemical exposure, thorough evaluation of such substances in the environment is essential. Regulatory standards typically mandate toxicity testing, encompassing hazard identification, dose-response assessment, exposure evaluation, and risk characterization (Krewski et al., 2010). As part of hazard identification, it is necessary to determine the specific toxicity endpoints associated with each chemical. In parallel, in vitro and in vivo studies aim to elucidate the conditions under which these toxic effects may occur in humans, often drawing on epidemiological insights. Dose-response assessments examine the relationship between chemical exposure and adverse effects, using benchmarks such as the no-observed-adverse-effect level (NOAEL), lowest-observed-adverse-effect level (LOAEL), and potential carcinogenicity (NRC, 1994). While this approach focuses on the magnitude of exposure required to produce harmful effects, the adverse outcome pathway (AOP) framework provides a complementary mechanistic perspective (Ankley et al., 2010). AOPs begin with a molecular initiating event, such as a chemical binding to a receptor, and proceed through a series of causally connected key events (KEs) until an adverse outcome (AO) is reached at the organism level (Villeneuve et al., 2014). By linking mechanistic insights with experimental data, AOPs exemplify how diverse information sources can be integrated to better understand chemical toxicity (Villeneuve et al., 2014). This growing emphasis on data integration has also driven the development of AI-based models with both experimental and computational inputs to support early-stage toxicity prediction.

The advent of computational approaches, combined with the growing availability of experimental data, has paved the way for more cost-effective, time-efficient strategies in early-stage drug discovery (Mak and Pichika, 2019; Vamathevan et al., 2019). By incorporating AI-based toxicity prediction models into virtual screening pipelines, compounds likely to exhibit toxicity can be filtered out before in vitro assays. This strategy increases the success rate of candidates advancing through toxicity evaluations, thereby enhancing the overall efficiency of drug development (Figure 1A). AI models can be trained on large-scale public databases such as ChEMBL (Gaulton et al., 2017), DrugBank (Wishart et al., 2018), and BindingDB (Liu et al., 2007), which contain in vitro and in vivo experimental results. In addition to open-source datasets, proprietary data generated from in vitro assays, in vivo studies, clinical trials, and post-marketing surveillance can further enrich these models (Pognan et al., 2023). Integrating AI-based toxicity prediction into virtual screening and then feeding back the experimental outcomes from downstream studies (in vitro, in vivo, and clinical), creates a virtuous cycle. This feedback process includes prospective and external validations, which evaluate model performance using newly generated or independent datasets and are essential for demonstrating generalizability and robustness in regulatory submissions. This continuous feedback loop improves model performance over time and supports more informed decision-making in early toxicity assessment (Pognan et al., 2023).

To develop such models, a systematic workflow is essential, typically consisting of four key stages: data collection, data preprocessing, model development, and evaluation (Figure 1B). The first step involves gathering drug toxicity data from a variety of sources. These data sources, including both public databases and proprietary collections, provide extensive information on chemical structures, bioactivity, and associated toxicity profiles, forming a rich foundation for supervised learning (Pognan et al., 2023). Once the data is collected, preprocessing is carried out to transform raw experimental results into formats suitable for machine learning. This includes handling missing values, standardizing molecular representations (e.g., SMILES strings or molecular graphs), and performing feature engineering such as calculating molecular descriptors (e.g., molecular weight, clogP, number of rotatable bonds) (Wigh et al., 2022). Toxicity labels are also encoded appropriately. These steps ensure data consistency and help extract informative features for training predictive models. The next stage involves selecting and training appropriate modeling techniques. Depending on the data structure and task complexity, a variety of algorithms can be applied, including Random Forest, XGBoost, Support Vector Machines (SVMs), neural networks, as well as more recent approaches such as Graph Neural Networks (GNNs). GNNs align well with the graph-based nature of molecular structures, which contributes to their strong predictive performance in various molecular property prediction tasks (Jiang et al., 2021; Reiser et al., 2022). In addition, they facilitate the identification of substructures or motifs associated with specific biological effects, thereby enhancing both the accuracy and interpretability of toxicity prediction models (Jiang et al., 2021; Reiser et al., 2022; Wu Z. X. et al., 2023). Transformer-based models, originally developed for natural language processing, have also shown strong potential in cheminformatics (Schwaller et al., 2019; Tibo et al., 2024).

In the evaluation phase, performance metrics are selected based on the type of prediction task. For classification models, metrics such as accuracy, precision, recall, F1-score, and area under ROC curve (AUROC) are used to evaluate the model’s ability to correctly distinguish toxic from non-toxic compounds. For regression models that predict continuous values like LD₅₀ or IC₅₀, commonly used metrics include MSE, RMSE, MAE, and R ². In addition to these quantitative measures, interpretability techniques such as SHAP or attention-based visualizations can provide insights into the features driving model predictions, supporting both model validation and decision-making in drug development (Rodríguez-Pérez and Bajorath, 2020; Wang Y. M. et al., 2023).

Driven by the growing need for early toxicity screening, advances in AI model architectures, and the emergence of robust development frameworks, a number of AI-based toxicity prediction models have recently been proposed. These models vary in scope and specificity, often categorized based on the target organ or the type of assay data used for training. This review summarizes representative toxicity prediction models that cover a broad range of toxicological endpoints. In particular, it focuses on models developed for ADMET profiling, hepatotoxicity, cardiotoxicity, neurotoxicity, and mutagenicity/genotoxicity prediction. Each category reflects distinct biological concerns and methodological approaches. Model development within these domains has evolved in response to challenges such as data scarcity, protocol heterogeneity, and class imbalance (Cavasotto and Scardino, 2022; Liu et al., 2023). To address these issues, various strategies have been employed, including multi-task learning, multimodal integration, and active learning. These strategies are discussed in more detail in later sections. In addition, scaffold-based data splitting is also commonly used to evaluate model generalizability across novel chemical structures while minimizing data leakage. In summarizing these models, this review also highlights differences in data sources, input representations, model architectures, and evaluation strategies and interpretability techniques used across toxicity endpoints. These aspects reflect how AI models are tailored to meet the distinct challenges of each toxicological domain.

A wide range of publicly available datasets have been developed to support toxicity prediction using machine learning and deep learning approaches (Table 1). Among the most widely used is Tox21, which comprises qualitative toxicity measurements of 8,249 compounds across 12 biological targets, primarily focused on nuclear receptor and stress response pathways (Richard et al., 2021). A related resource, ToxCast provides high-throughput screening data for approximately 4,746 chemicals tested across hundreds of biological endpoints, offering broad mechanistic coverage for in vitro toxicity profiling (Richard et al., 2016). These datasets are frequently employed as benchmarks for evaluating classification models in predictive toxicology.

To assess clinical toxicity risks, the ClinTox dataset offers labeled data differentiating compounds that were approved by regulatory agencies from those that failed in clinical trials due to toxicity (Gayvert et al., 2016). Several datasets have been curated for evaluating cardiotoxicity associated with the human Ether-à-go-go–related gene (hERG) channel blockade. The hERG dataset (Wang et al., 2016; Karim et al., 2021) includes over 13,000 compounds annotated with binary labels based on a 10 µM inhibition threshold, while the hERG blockers dataset provides a smaller set of 648 compounds (Wang et al., 2016; Karim et al., 2021). A more extensive resource, hERG Central, encompasses over 300,000 experimental records and supports both classification and regression tasks based on various hERG inhibition assays (Du et al., 2011). Liver toxicity is addressed in the DILIrank (Drug-Induced Liver Injury) dataset, which contains 475 compounds annotated for their hepatotoxic potential, an important factor in post-market drug withdrawals (Xu et al., 2015). The SIDER dataset presents multi-label side effect annotations for more than 1,400 marketed drugs, enabling the prediction of clinically observed adverse drug reactions (Kuhn et al., 2016). For dermatological toxicity, the Skin Reaction dataset includes 404 compounds evaluated for their potential to cause skin sensitization (Alves et al., 2015). Genotoxicity is commonly assessed using the AMES dataset, which comprises 7,255 compounds labeled based on the Ames test—a standard assay for detecting mutagenic potential (Xu et al., 2012). The Carcinogens dataset contains 278 compounds classified as carcinogenic or non-carcinogenic, serving as a benchmark for cancer risk prediction (Lagunin et al., 2009). Finally, acute systemic toxicity is represented by the LD₅₀_Zhu dataset, which includes LD₅₀ values for 7,385 compounds and supports regression modeling of lethal dose responses (Zhu et al., 2009). Collectively, these datasets span a broad range of toxicological endpoints and data modalities and have become foundational resources for the development, validation, and comparison of AI-driven toxicity prediction models.

At the same time, their widespread adoption has revealed several practical challenges that impact real-world applications. For instance, data scarcity in certain toxicity endpoints can hinder the performance of machine learning models that depend on sufficient training data. In some cases, limited data may fail to represent diverse chemical scaffolds, reducing model generalizability. When class imbalance is also present, such as a higher proportion of non-toxic compounds, the effects of data scarcity can be further amplified (Cavasotto and Scardino, 2022). Since toxicity labels are typically derived from experimental measurements, inconsistencies across assay protocols often lead to a lack of data uniformity. This protocol heterogeneity can make it difficult to merge datasets from different sources. Furthermore, annotation noise resulting from experimental variability or ambiguous labeling can introduce additional challenges during model training (Liu et al., 2023).

To overcome these issues, expanding datasets through newly generated experimental data and literature-based curation can help improve coverage and diversity. In parallel, standardizing toxicity testing protocols and documentation practices may enhance data consistency and interoperability. These efforts are expected to contribute meaningfully to the development of more robust and reliable AI-based toxicity prediction models in drug discovery.

Several publicly accessible ADMET prediction tools, including ADMETLab 3.0, Deep-PK, ProTox 3.0, Helix-ADMET, FP-ADMET, and admetSAR 2.0 (Yang et al., 2019; Venkatraman, 2021; Zhang et al., 2022; Banerjee et al., 2024; Fu et al., 2024; Myung et al., 2024), provide a wide array of toxicity prediction models, each differing in scope, algorithmic strategy, and coverage. ADMETLab 3.0 offers predictive models for 119 endpoints, including toxicity-related properties such as hERG inhibition, carcinogenicity, and respiratory toxicity. These models are built using directed message-passing neural networks (DMPNNs) and incorporate uncertainty estimation features. The toxicity models, such as the one for hERG inhibition, have demonstrated strong performance with AUROC values approaching 0.94. In terms of interpretability, ADMETLab 3.0 provides uncertainty scores alongside predictions, uses colored indicators to represent empirical decision states, and highlights structural alerts contributing to toxicity (Fu et al., 2024). Deep-PK is a deep learning–based framework that predicts 73 endpoints, including 35 toxicity-related endpoints, 29 other ADMET properties, and 9 general molecular descriptors. While its primary focus lies in pharmacokinetic regression tasks and ADMET optimization, it offers comprehensive support for toxicity assessment through GNN-based pipelines that accept SMILES, SDF, and molecular descriptor inputs. The model also provides interpretability by identifying key molecular subgraphs that contribute to prediction outcomes (Myung et al., 2024). ProTox 3.0 is particularly comprehensive in its treatment of toxicity, providing 61 predictive models covering a broad spectrum of endpoints. These include organ-specific toxicities such as hepatotoxicity, neurotoxicity, cardiotoxicity, and nephrotoxicity, along with models for clinical, immunological, and nutritional toxicities. The platform integrates mechanistic insights through AOPs, molecular initiating events, and target-specific toxicities, and supports ontology-driven, systems-level interpretation (Banerjee et al., 2024). Helix-ADMET is a flexible ADMET prediction platform that combines self-supervised and multi-task learning to enhance generalizability across diverse chemical scaffolds. It supports fine-tuning on user-defined endpoints and classifies toxicity into macro- and micro-level categories (Zhang et al., 2022). FP-ADMET is an open-source tool that focuses on over 50 ADMET-related endpoints, including drug-induced liver injury, hERG inhibition, hemolytic toxicity, mitochondrial toxicity, and cell-specific cytotoxicity. The models are constructed using random forest classifiers trained on 20 different types of chemical fingerprints, enabling broad chemical space coverage and compound exploration (Venkatraman, 2021). admetSAR 2.0 provides 47 curated endpoints, including Ames mutagenicity, carcinogenicity, immunotoxicity, and hERG inhibition. It employs traditional machine learning algorithms such as random forest, SVM, and k-nearest neighbors (KNNs) applied to molecular descriptors and fingerprints (Yang et al., 2019).

The comprehensiveness of these tools not only facilitates broad ADMET screening but also enables prioritization of drug candidates with favorable safety profiles. The development of such general-purpose prediction tools has been largely driven by advances in molecular representations that effectively capture compound features, along with the availability of benchmark datasets annotated with a wide range of ADMET endpoints. On the other hand, tools that focus on specific toxicity types such as hepatotoxicity, cardiotoxicity, nephrotoxicity, neurotoxicity, and genotoxicity/carcinogenicity often require task-specific datasets and tailored feature engineering strategies to enhance predictive performance. The following sections introduce these organ- and mechanism-specific toxicity models and discuss how specialized data and domain-informed approaches contribute to their effectiveness.

Each endpoint is characterized by differences in data properties, sources including databases, and overall data volume. Furthermore, depending on the specific toxicity pathways involved, areas of interest such as the level of interpretability required can also vary. As a result, models for each endpoint have been designed to reflect these unique characteristics, leading to differences in the features used and the methodological approaches adopted (Figure 2 and Table 2). While many of these models share a common foundation in molecular data, it is important to note that the choice of features and modeling techniques is often tailored to the distinct goals and nature of each endpoint.

In hepatotoxicity prediction, physicochemical properties of molecules are known to be influential and are often incorporated into models (Chen et al., 2013a; Kotsampasakou and Ecker, 2017; Lee and Yoo, 2024). Both deep learning and tree-based methods have been used with comparable frequency. For cardiotoxicity, particularly related to hERG channel blockade, the availability of larger datasets has encouraged the use of more data-intensive deep learning approaches. GNNs are frequently applied due to their structural compatibility with molecules and their ability to offer interpretability through substructure-level attention (Jiang et al., 2021; Yang et al., 2024; Lee and Yoo, 2025). In renal or nephrotoxicity prediction, traditional machine learning models are more commonly used, as they tend to perform better than deep learning when data are limited (Xu et al., 2023). In neurotoxicity studies, a single study may develop multiple models to address distinct tasks such as BBB permeability, neuronal cytotoxicity, neural activity interference, and general neurotoxicity, enabling broader predictive coverage (Pang et al., 2025). For genotoxicity and carcinogenicity, multi-task learning has been applied to predict outcomes across several Ames test strains within a single model. This approach outperformed single-task models by leveraging shared parameters across tasks (Martínez et al., 2022). These variations, driven by endpoint-specific requirements, are elaborated in the subsequent sections.

As previously discussed, AI model design is significantly affected by both the characteristics and the volume of data available for training. In the context of toxicity prediction for drug discovery, input data typically comprises molecular structures, physicochemical properties, and task-specific features. However, these tasks are often constrained by the limited availability of labeled data. To address this challenge, a variety of data-efficient learning strategies have been developed to maximize predictive performance under label-scarce conditions.

In data-scarce settings, transfer learning strategies use pre-trained parameters to boost toxicity prediction performance. For instance, HelixADMET employs a three-stage training framework that incorporates self-supervised pretraining on large-scale unlabeled molecular data, followed by multi-task and fine-tuning stages to transfer learned chemical knowledge to various ADMET endpoints, significantly improving extrapolation to novel chemical scaffolds (Zhang et al., 2022). Multimodal models ingest diverse data types (e.g., chemical structures, omics profiles, and bioactivity assays) simultaneously to capture complementary information. For example, M2REMAP is a multimodal deep learning framework that predicts drug indications, mono-drug side effects, and drug–drug interaction side effects by integrating molecular chemical structures with clinical semantic embeddings derived from large-scale electronic health records (EHR). By learning joint representations across these heterogeneous modalities, M2REMAP achieves superior predictive accuracy and generalizability over unimodal baselines (Wen et al., 2023). Martínez et al. developed multi-task deep neural networks to simultaneously predict Ames mutagenicity across multiple Salmonella typhimurium strains (Martínez et al., 2022). They demonstrated that shared representations improved performance, especially on the strains with limited training data. Active learning enhances data efficiency by strategically selecting the most informative samples, enabling high model performance even with limited labeled data. For example, the muTOX-AL framework integrates structure-based and activity-based selection strategies to guide experimental toxicology, significantly improving model performance with fewer labeled compounds compared to random sampling (Xu et al., 2024). Federated learning enables multiple institutions to collaboratively train a global toxicity prediction model on decentralized datasets in which each party keeps its raw data locally and only shares model updates, thus preserving data privacy and regulatory compliance while benefiting from a much larger, heterogeneous training pool. The MELLODDY project exemplifies this approach, demonstrating that federated QSAR models trained across ten pharmaceutical companies achieved comparable or superior predictive performance to local models, while maintaining strict data confidentiality (Heyndrickx et al., 2023).

In parallel, interpretability techniques are also advancing to better inform and guide decision-making in drug discovery based on model predictions. SHAP estimates the contribution of each input feature to the output, providing insight into which molecular properties influence model decisions (Lundberg and Lee, 2017). For graph-based models, methods like EdgeSHAPer extend this concept by identifying important substructures within molecular graphs (Mastropietro et al., 2022). Attention-based visualizations, commonly used in transformer and graph neural network models, highlight which parts of the input the model focuses on during prediction (Ying et al., 2019; Zheng et al., 2019). For example, in SMILES-based models, attention heatmaps can reveal which atoms or functional groups are most influential in predicting toxicity. Counterfactual explanations, on the other hand, offer intuitive and sparse insights by showing the smallest alteration to input features that would change a model’s prediction, particularly useful for understanding how minimal molecular changes affect outcomes. In drug design, small structural modifications can often result in counterfactual cases with significant impact on activity or toxicity, leading to growing interest in counterfactual explanation methods to better capture such subtle yet meaningful variations (Wellawatte et al., 2022).

Building on these recent advances, the next phase of toxicity prediction may be driven by foundation models and large-scale language-based systems. Looking ahead, emerging foundation models such as MoleculeGPT (Liu et al., 2024), BioT5 (Pei et al., 2023), and ChemCrow (Bran et al., 2024) could be applied to toxicity prediction. Even before the advent of large language models (LLMs), Papamokos and Silins demonstrated that integrating QSAR modeling with text mining improved the mechanistic understanding of carcinogenicity and helped compensate for limited structure–activity data on non-genotoxic compounds (Papamokos and Silins, 2016). By linking chemical structures with literature-derived modes of action, their hybrid approach offered more biologically meaningful interpretations to support mechanism-based toxicity evaluation. Today, with the advent of powerful LLMs, such strategies can be further scaled and generalized. Fine-tuning these large, pre-trained models enables researchers to integrate broad and transferable chemical and biological knowledge into downstream toxicity prediction tasks, while producing mechanistically interpretable results even in data-scarce domains. These AI-driven methods not only improve predictive accuracy in data-scarce scenarios but are also continuing to advance rapidly, expanding the possibilities for mechanism-informed toxicology.

The use of AI-based models in regulatory toxicology is drawing growing interest, especially as agencies seek alternatives to animal testing. Yet, adoption remains limited due to the absence of clear validation standards and acceptance criteria. ICH guidelines, including M7 (R2), S2 (R1), and S1B(R1), provide frameworks for using in silico approaches such as AI-based models and advanced QSAR tools (ICH, 2022; 2023). These can support mutagenic impurity screening, genotoxicity testing, and carcinogenicity assessment, provided they are properly justified and validated. In this context, the FDA NCTR’s AI4TOX program is specifically aimed at applying AI to toxicology to develop new tools that support FDA regulatory science and strengthen the safety review of FDA-regulated products (An FDA Artificial Intelligence, 2024). It focuses on leveraging AI for tasks like developing virtual animal models, evaluating toxicological endpoints, and analyzing complex data from FDA documents and histopathology. For broader adoption, AI models must align with regulatory expectations, demonstrate consistent performance, and offer interpretability.

As the use of AI in regulatory toxicology continues to expand, it becomes increasingly important to consider how existing validation principles can be adapted or extended to ensure these models meet regulatory standards. To enhance the reliability and regulatory acceptance of AI-based toxicity prediction models, it is useful to apply the OECD QSAR validation principles (OECD, 2014). Originally developed for traditional QSAR models, these principles outline key elements such as defined endpoints, transparent algorithms, applicability domains, performance metrics, and mechanistic interpretation when possible. While these criteria remain broadly relevant, the guidance was established before the advent of modern AI techniques. Given the rapid development of AI and its increasing integration into the drug discovery process, there is a growing need for updated validation frameworks that explicitly address the unique challenges and opportunities presented by AI-based modeling approaches.

The efficacy and safety of chemical compounds are fundamental considerations in drug discovery, with toxicity representing a key determinant of clinical success or failure. AI-based prediction models have emerged as powerful tools for toxicity assessment during the early stages of drug discovery. As databases continue to grow, computational resources become more accessible, and AI architectures evolve, these models have significantly advanced beyond traditional computational methods and enable reliable predictions across various toxicological endpoints, including hepatotoxicity, cardiotoxicity, nephrotoxicity, neurotoxicity, and genotoxicity. This review has systematically examined both general-purpose ADMET prediction tools and endpoint-specific toxicity models, highlighting rapid progress, increasing methodological sophistication, and expanding diversity within the field of computational toxicology. Nevertheless, several critical challenges persist. First of all, despite significant advances, current AI models frequently struggle with accurately predicting complex and rare toxicity events due to intrinsic biological complexities. The scarcity of high-quality labeled data, particularly data that accurately reflects clinical outcomes or rare toxicological events, severely constrains model training and validation. Also, generalizability to novel chemical scaffolds remains uncertain, limiting confidence in AI predictions for structurally diverse or innovative drug candidates. Finally, interpretability also remains a crucial bottleneck; although advanced AI models offer powerful predictive capabilities, their complex inner workings often limit the clarity and transparency required by regulatory bodies and clinical practitioners.

To overcome these limitations, future research can be focused on the integration of diverse data types, including detailed chemical structures, comprehensive biological assay outcomes, multi-omics profiles, and real-world clinical datasets. Such integration will enable AI models to capture the multifaceted nature of toxicological responses in a better way. Harmonizing toxicity annotations across multiple databases will also significantly enhance data interoperability, enabling more extensive and efficient utilization of available data resources. In parallel, fostering deeper cross-disciplinary collaboration among computational scientists, toxicologists, medicinal chemists, clinical pharmacologists, and regulatory experts is essential. Such collaborations can facilitate the development of predictive models that are not only robust and accurate but also practically interpretable, ensuring that model insights can directly inform discovery and regulatory decisions.

As AI technologies continue to evolve, it would be definite that they hold significant potential for enhancing early-stage decision-making, substantially reducing late-stage drug development failures, and accelerating the delivery of safer, more effective therapeutic solutions to patients. To fully employ this potential, it is crucial to foster a deeper understanding of the real-world implications and limitations of predictive outcomes. Practical integration requires not just technological advances but also a comprehensive awareness of pharmaceutical, clinical realities and regulatory standards. Thus, ongoing dialogue and knowledge-sharing between computational developers, experimental toxicologists, clinical researchers, and regulatory stakeholders will be indispensable in shaping the next-generation of AI-driven predictive toxicology tools that meaningfully improve drug discovery outcomes in both academic research and industry practice.