Machine Learning-Enabled Drug-Induced Toxicity Prediction

Abstract

Unexpected toxicity has become a significant obstacle to drug candidate development, accounting for 30% of drug discovery failures. Traditional toxicity assessment through animal testing is costly and time-consuming. Big data and artificial intelligence (AI), especially machine learning (ML), are robustly contributing to innovation and progress in toxicology research. However, the optimal AI model for different types of toxicity usually varies, making it essential to conduct comparative analyses of AI methods across toxicity domains. The diverse data sources also pose challenges for researchers focusing on specific toxicity studies. In this review, 10 categories of drug-induced toxicity is examined, summarizing the characteristics and applicable ML models, including both predictive and interpretable algorithms, striking a balance between breadth and depth. Key databases and tools used in toxicity prediction are also highlighted, including toxicology, chemical, multi-omics, and benchmark databases, organized by their focus and function to clarify their roles in drug-induced toxicity prediction. Finally, strategies to turn challenges into opportunities are analyzed and discussed. This review may provide researchers with a valuable reference for understanding and utilizing the available resources to bridge prediction and mechanistic insights, and further advance the application of ML in drugs-induced toxicity prediction.

Keywords: database; deep learning; drug toxicity prediction; machine learning.

Figure 1

The comprehensive framework for drug toxicity prediction using multimodal data and ML: A data‐driven approach to reduce biological experimentation and enhance drug safety. The first part (Figures 1A–C) shows the integration of various data sources‐toxicity data, chemical data, and multi‐omics information (including genomics, proteomics, metabolomics, and single‐cell sequencing) into an ML pipeline designed for accurate drug toxicity prediction. The second part (Figures 1D, E) details the stepwise process of data preprocessing, model fitting, prediction, and evaluation, including preprocessing raw data into relevant features such as chemical properties, toxicity markers, and omics data, followed by model fitting and hyperparameter tuning. The performance of models is evaluated using standard metrics (True Positive, TP; True Negative, TN; False Positive, FP; False Negative, FN) to ensure prediction accuracy and interpretability. The third part (Figures 1F, G) compares traditional toxicity testing methods with this data‐driven ML approach, significantly reducing the need for in vivo biological testing. It highlights how this integrated method can improve the clinical application of drugs and decrease the demand for extensive biological experiments. This framework provides a more effective pathway for identifying potential drug toxicity risks before clinical trials, facilitating faster and safer drug development.

Figure 2

ML and DL Algorithms with Good Performance in Drug Toxicity Prediction. A) RF: An ensemble learning method that constructs multiple Decision Trees (DT) and combines their predictions to improve model accuracy. B) SVM: Maps chemical features to a high‐dimensional space and creates an optimal hyperplane boundary for classification tasks. C) XGB: An optimized implementation of gradient‐boosted decision trees (GBDT), which combines multiple weak learners (e.g., DT) to form a strong model, particularly suitable for handling complex nonlinear relationships. D) DNN: A DL model composed of multiple layers of fully connected neurons, capable of learning complex nonlinear relationships. E) Multi‐Task Deep Neural Network (MTDNN): An extended DNN that supports multi‐task learning, enabling the simultaneous prediction of multiple toxicity endpoints while sharing underlying features and task‐specific layers. F) GNN: A DL model specifically designed to handle graph‐structured data (e.g., molecular graphs). It learns the relationships between nodes (atoms) and edges (bonds) in the molecular structure and captures molecular features through a message‐passing mechanism. G) GAT: A variant of GNN that integrates an attention mechanism to assign different weights to nodes and edges in graph‐structured data, enhancing the model's ability to focus on important molecular fragments. H) GNN and Multilayer Perceptron (MLP) Joint Application: Uses graph embeddings extracted by GNN from molecular graphs as inputs, which are then passed to an MLP for further predictions. I) Long Short‐Term Memory (LSTM): An improved version of recurrent neural RNN, focused on handling sequential data and capable of capturing long‐term dependencies.

Figure 3

Applications of ML in Drug Toxicity Prediction. The figure shows various ML models used for predicting different types of drug toxicity, highlighting specific algorithms used for each toxicity category. It provides a strategic reference for selecting different ML models based on the nature of the toxicity type and the available data. A) Acute Toxicity: RF, SVR, Attentive FP, GAT, and MolFPG performed well in specific acute toxicity prediction studies. B) Clinical Toxicity: MTDNN showed good performance in predicting specific clinical toxicity. C) Carcinogenicity: GBM, Capsule network + GAT, and GNN + MLP performed well in specific carcinogenicity prediction studies. D) Hepatotoxicity: SVM, RF, BN, XGB, DNN, GAN, and SSM performed well in specific hepatotoxicity prediction studies. E) Cardiotoxicity: RF, GAT, and GCN showed good performance in specific cardiotoxicity prediction studies. F) Nephrotoxicity: DNN and LSTM were effective in predicting specific nephrotoxicity. G) Respiratory Toxicity: SVM, XGB, NB, and MPNN performed well in specific respiratory toxicity prediction studies. H) Neurotoxicity: RF, SVM, and CNN showed good performance in specific neurotoxicity prediction studies. I) Hematotoxicity: XGB and Attentive FP performed well in specific hematotoxicity prediction studies. J) Mitochondrial Toxicity: Catboost and RF showed good performance in specific mitochondrial toxicity prediction studies. The bolded italicized portions are algorithms that have achieved good performance in a particular study.

Figure 4

Databases and tools of ML for drug toxicity prediction A) Toxicity Databases provide label data for drug toxicity prediction: a. Universal Chemical Toxicity Database, b. Organ‐Specific Databases, c. Carcinogenicity Database, d. Other Toxicity Databases, e. Toxicity Databases covering rodents, rabbits, birds, pigs, fish, insects, humans, and in vitro cell levels. B) Chemical Databases provide feature data for drug toxicity prediction: f. Universal Chemical and Drug Databases, g. Chemical‐related Target and Interaction Databases, h. Natural Product and Traditional Medicine Databases. C) Multi‐Omics Databases provide feature data for drug toxicity prediction: i. Genomic and Transcriptome Databases, j. Proteomics Databases, k. Metabolomics Databases, l. Multi‐Omics Databases. D) ML Benchmark Databases provide ML benchmark results for drug toxicity prediction: m. ML Benchmark Databases. E) Web Servers and Tools provide rapid visualization, analysis, and quantification of chemical toxicity predictions: n. Chemical Toxicity Prediction Tools, o. Protein and Peptide Toxicity Prediction Tools, p. ADMET Prediction Tools.

Figure 5

Challenges and Opportunities for Future ML in Drug Toxicity Prediction. Although ML has advanced drug toxicity prediction, it still faces challenges and significant opportunities. We summarize four key challenges, with italicized text explaining how to address them. A) Sample Size: The small size of available datasets limits ML performance. Advances in big data accumulation, data augmentation, pre‐trained models, active learning, and meta‐learning, as well as improved data sharing and integration, can help overcome this challenge. B) Data Quality: Poor data quality hampers model reliability. Data cleaning, quality control, standardized global data formats, federated learning, multimodal ML, transfer learning, zero‐shot learning, and ensemble methods are key strategies to improve data quality. C) Optimal Model: Finding suitable ML models for drug toxicity prediction is difficult. Emerging DL models (Transformers and GNNs), interdisciplinary collaboration, creating benchmark databases, and balancing accuracy with interpretability are essential for selecting the right models. D) Model Interpretability: Limited interpretability complicates ML application. Developing customized interpretable artificial intelligence (XAI) methods for different data types, fostering Multidisciplinary cooperation, integrating domain knowledge, and aligning with regulatory frameworks can address challenges in model applicability, the lack of interpretability metrics, and data complexity.