Background

After evaluating the effectiveness of drugs in the process of discovery and development of new drugs, preclinical tests were performed to evaluate their toxicity in animals prior to clinical trials [1]. Preclinical toxicity testing can predict toxic responses in humans, determine safe doses in clinical trials, and monitor toxicity-related symptoms and target organs in patients [2]. After having confirmed the safety of drugs by evaluating various toxicities, such as acute toxicity, nephrotoxicity, cardiotoxicity, reproductive toxicity, and genotoxicity, clinical trials can be performed.

Acute toxicity is defined as deleterious toxicological effects of a chemical from single or multiple exposures over a short duration (usually < 24 h) [3]. Studies on acute toxicity have examined various routes of exposure (e.g., oral, dermal, and inhalation) using rodents, such as mice and rats, to assess lethal doses. During the drug development process, acute toxicity is generally evaluated using acute oral toxicity (AOT) tests that assess acute toxic reactions and the lethal dose 50 (LD50) after a single oral administration to rodents [4, 5]. The LD50 is defined as the dose of the test substance that can kill 50% of animals within 24 h of exposure.

AOT tests in animals are conducted after confirming the efficacy of a drug [5]. Therefore, if drug development fails owing to AOT at this stage, economic loss of development costs incurs. However, as dozens to thousands of compounds are considered drug candidates in the early stages of development, performing AOT tests on all compounds is limited by time and cost issues. Therefore, it is important to evaluate the AOT of such compounds in the early stages of drug development.

To address these issues, various prediction models have been developed based on the results of AOT tests of thousands of compounds [6,7,8]. In addition, various software programs that predict the AOT of small compounds based on machine learning models are currently available [7,8,9,10]. However, several aspects need to be improved for the development of an AOT prediction model. First, it is necessary to develop a computational framework that can predict the AOT in both mice and rats. AOT evaluation of small compounds is still being performed using either mice or rats. Second, it is necessary to further improve the prediction performance of models. One strategy to improve the performance of a model is to build it so that it can properly consider the distribution of the data. For example, using data composed of skewed LD50 values for model training may reduce prediction performance [11].

In the present study, we propose a new computational framework, PredAOT, that predicts the AOT for a given compound in mice and rats. PredAOT is based on multiple random forest models for AOT prediction. For the development of PredAOT, we used a total of 6,226 and 6,238 compounds whose AOT was evaluated in mice and rats, respectively. Moreover, we compared the prediction performance of PredAOT with that of other existing tools. PredAOT is a useful tool for predicting the acute oral toxicity of small compounds during drug development.

Results and discussion

Development of the PredAOT framework

To develop a computational framework (i.e., PredAOT) for accurate prediction of AOT, we first collected data on the AOT of compounds (i.e., LD50) reported for mice and rats. The AOT data for mice for 6226 compounds were obtained from the OCHEM database [7]. The AOT data for rats for 6238 compounds were obtained from the literature [6].

According to the GHS Classification, the AOT level of compounds can be divided into five categories [12] (Table 1). For example, compounds in Category 1 are toxic, while Category 5 compounds are less likely to be toxic. Additionally, we found that the distribution of LD50 values for the compounds was skewed towards categories 4 and 5 in both mice and rats. This data imbalance can adversely affect model training, such as overfitting. To address this issue, we decided to classify the AOT of compounds into two categories (i.e., “toxic” and “less or non-toxic”) instead of five categories (Table 2); compounds with LD50 ≤ 300 mg/kg and compounds with LD50 > 300 mg/kg. Notably, although a compound is classified as less or non-toxic, it does not necessarily have to involve no AOT at all. During the model development, LD50 values were transformed to log10 transformed LD50 values.

Table 1 Number of compounds in datasets for each acute oral toxicity category
Full size table
Table 2 Number of compounds in the “less or non-toxic” and “toxic” datasets
Full size table

We then used the datasets to train a binary classification model, called “AOT classifier,” that predicts AOT as toxic (i.e., LD50 ≤ 300 mg/kg) or less or non-toxic (i.e., LD50 > 300 mg/kg) for a given compound. In addition, as shown in Fig. 1, toxic and less or non-toxic datasets were used to train two regression models called “toxic regressor” and “less or non-toxic regressor”, respectively. In particular, the “toxic regressor” was trained with the toxic dataset (i.e., LD50 ≤ 300 mg/kg) and the “less or non-toxic regressor” was trained with the less or non-toxic dataset (i.e., LD50 > 300 mg/kg).

Fig. 1
figure 1

Overall scheme of PredAOT. PredAOT uses the chemical structure as an input. Thereafter, the molecular fingerprint (i.e., ECFP4) is used as an input feature for models in PredAOT. PredAOT is composed of one classification model (i.e., “AOT classifier”) and two regression models (i.e., “toxic regressor” and “less or non-toxic regressor”). The AOT classifier is used for prediction as “toxic” (LD50 ≤ 300 mg/kg) or “less or non-toxic” (LD50 > 300 mg/kg) for a given compound. If a compound is predicted to be toxic, PredAOT predicts the LD50 of the compound using the toxic regressor trained with compounds with LD50 ≤ 300 mg/kg. If a compound is predicted to be less or non-toxic, PredAOT predicts the LD50 of the compound using the less or non-toxic regressor trained with compounds with LD50 > 300 mg/kg. All these procedures are equally applied to the AOT prediction process in mice and rats

Full size image

To summarize the AOT prediction process in the PredAOT, “AOT classifier” first predicts AOT as “toxic” or “less or non-toxic” for a given compound. Thereafter, if a compound is predicted as toxic, the “toxic regressor” predicts the LD50 of the compound; otherwise, the “less or non-toxic regressor” predicts the LD50 of the compound. All procedures were equally applied to the AOT prediction process in mice and rats.

Optimization and evaluation of AOT prediction models

To build an optimal AOT prediction model, we constructed and evaluated six different machine learning models: a message passing neural network (MPNN) based on graph neural networks, MPNN with molecular fingerprints, MPNN with molecular descriptors, random forest (RF), support vector machine (SVM), and artificial neural network (ANN) models based on molecular fingerprints (see Materials and Methods). The training dataset was used for model training to determine the model with the best prediction performance. A test dataset was used to assess the performance of the final model.

First, we optimized the binary classification model (i.e., “AOT classifier”) used to predict the AOT of a given compound in mice and rats. To this end, we evaluated various hyperparameters using a grid search technique with fivefold cross-validation, and used accuracy as the model performance metric. RF showed the highest accuracy for mice (0.8672) and rats (0.8377; Fig. 2). To further improve the prediction performance of the RF model, we applied an oversampling approach to manage imbalanced data (Table 2). Specifically, we used the synthetic minority oversampling technique (SMOTE), which is a popular algorithm used to generate artificial data [13]. In doing so, RF with SMOTE showed better prediction performance than RF without SMOTE: accuracies of 0.9586 and 0.9335 in mice and rats, respectively (Fig. 3). In addition, we evaluated the prediction performance using the test dataset. The RF model with SMOTE showed an area under the receiver operating characteristic (AUROC) of 0.7778, Matthew’s correlation coefficient (MCC) of 0.5514, positive predictive value (PPV) of 0.6627, and negative predictive value (NPV) of 0.8845 in mice, and an AUROC of 0.7442, MCC of 0.4929, PPV of 0.6435, and NPV of 0.8539 in rats (Table 3). Based on these results, we used RF models with SMOTE in both mice and rats in the PredAOT framework.

Fig. 2
figure 2

Cross-validation prediction performances of the MPNN, MPNN with MF, MPNN with MD, RF, SVM, and ANN models for the AOT classifier using mouse (A) and rat (B) datasets. Each performance metric value was calculated by five-fold cross-validation. MPNN, message passing neural network; MF, molecular fingerprint; MD, molecular descriptor; RF, random forest; SVM, support vector machine; ANN, artificial neural network

Full size image
Fig. 3
figure 3

Cross-validation prediction performances of RF models with and without SMOTE using mouse (A) and rat (B) datasets. MPNN, message passing neural network; RF, random forest; SVM, support vector machine; ANN, artificial neural network

Full size image
Table 3 Prediction performance of classification models in PredAOT on the test dataset
Full size table

As shown in Fig. 1, the AOT of the compound was first predicted to be “toxic” or “less or non-toxic” using an AOT classifier in mice and rats. The LD50 value (mg/kg) was subsequently quantitatively predicted using one of the regression models (i.e., “toxic regressor” or “less or non-toxic regressor”) according to the prediction result of the AOT classifier. Here, we optimized both regressors (i.e., “toxic regressor” or “less or non-toxic regressor”). The prediction performance of these regressors was evaluated through five-fold cross-validation using the training dataset. The root-mean-square error (RMSE) was used as the performance metric for the regression models. Consequently, RF showed the lowest RMSE, i.e., the best performance, in the "toxic regressor” and “less or non-toxic regressor” in both mice and rats. The RF model for both regressors showed an RMSE of 0.2999 and 0.3767 in mice, respectively (Figs. 4A, 5A) and 0.3919 and 0.4984 in rats, respectively (Figs. 4B, 5B). Thereafter, we evaluated the prediction performance using the test dataset. In mice, the toxic regressor showed an RMSE of 0.3806 and an R2 of 0.3557 on the test dataset (Table 4), whereas the less or non-toxic regressor showed an RMSE of 0.2923 and an R2 of 0.3881. In rats, the toxic regressor showed an RMSE of 0.5323 and an R2 of 0.3065 on the test dataset, whereas the less or non-toxic regressor showed an RMSE of 0.3863 and an R2 of 0.2702.

Fig. 4
figure 4

Cross-validation prediction performances of the MPNN, MPNN with MF, MPNN with MD, RF, SVM, and ANN models for toxic regressors using mouse (A) and rat (B) toxic datasets (i.e., LD50 ≤ 300 mg/kg). MPNN, message passing neural network; MF, molecular fingerprint; MD, molecular descriptor; RF, random forest; SVM, support vector machine; ANN, artificial neural network

Full size image
Fig. 5
figure 5

Cross-validation prediction performances of the MPNN, MPNN with MF, MPNN with MD, RF, SVM and ANN models for the non-toxic regressor using mouse (A) and rat (B) less or non-toxic datasets (i.e., LD50 > 300 mg/kg). MPNN, message passing neural network; MF, molecular fingerprint; MD, molecular descriptor; RF, random forest; SVM, support vector machine; ANN, artificial neural network

Full size image
Table 4 Prediction performance of regression models in PredAOT on the test dataset. RMSE and R2 values are calculated using log10 transformed LD50 values
Full size table

Prediction performance of PredAOT compared with other prediction models

We compared the prediction performance of PredAOT with that of other existing tools. First, we compared the prediction performance of PredAOT in mice. Thereafter, we compared the prediction performance of PredAOT with OCHEM Predictor using our test dataset [7]. The OCHEM Predictor predicted the LD50 (mg/kg) result in the same manner as PredAOT, such that predicted values could be directly compared. We compared the prediction performance for each toxicity group in Table 1. PredAOT yields relatively low RMSE values (i.e., improved performance) in groups with relatively little training data (e.g., Categories 1 and 2) (Table 5). Notably, the performance comparison indicates that the test dataset may have been used as training data in OCHEM Predictor.

Table 5 Comparison of RMSE values predicted by PredAOT with OCHEM Predictor on the mouse test dataset
Full size table

Second, we compared the prediction performance of PredAOT in rats with that of BESTox and aiQSAR using our test dataset [9, 10]. The prediction results of BESTox and aiQSAR were not directly comparable with the prediction results of PredAOT ​​because the unit for LD50 was not mg/kg. Therefore, we compared the prediction results using two correlation coefficients: Pearson’s correlation coefficient (Pearson’s r) and Spearman's rank correlation coefficient (Spearman's r). PredAOT showed the highest Pearson’s r of 0.7984 and Spearman's r of 0.7340 compared to BESTox and aiQSAR (Table 6).

Table 6 Comparative performance evaluation of PredAOT with BESTox and aiQSAR on the rat test dataset
Full size table

Although we did not compare LD50 values directly, PredAOT performed comparably or better than existing tools. PredAOT does not outperform other existing tools; however, it can be a useful tool for predicting AOT. In addition, it has the advantage of being able to predict AOT in mice and rats simultaneously.

Conclusions

In the present study, we developed a computational framework called PredAOT, which predicts the AOT of a given compound in mice and rats. PredAOT first classifies the given compound as “toxic” or “less or non-toxic”, and then further qualitatively predicts the LD50 value using a regression model. PredAOT is trained with information on the AOT of 6,226 and 6,238 compounds in mice and rats, respectively. PredAOT has the advantage of predicting AOT in mice and rats simultaneously, and its prediction performance is similar to or better than that of existing tools. The web server for implementing PredAOT is available at https://predaot.netlify.app/ (Fig. 6). PredAOT will be a useful tool for the quick and accurate prediction of the AOT of small compounds in mice and rats for successful drug development.

Fig. 6
figure 6

User interface of the PredAOT web server

Full size image

Methods

Data preparation

Information on acute oral toxicity (AOT) of 6,226 compounds determined in mice was obtained from the OCHEM database [7], and information on AOT of 6,238 compounds determined in rats was obtained from the literature [6]. We defined compounds with lethal dose 50 (LD50) values ≤ 300 mg/kg as “toxic” and compounds with LD50 values > 300 mg/kg as “less or non-toxic”. The dataset was divided into training (80%) and test (20%) datasets. The training dataset was used for hyperparameter optimization, and the test dataset was used for model evaluation (Fig. 1). During the model training and evaluation, we used log10 transformed LD50 values.

Preparation of molecular features

The structures of the compounds were presented in the simplified molecular-input line-entry system (SMILES) format [14]. To train message-passing neural networks (MPNNs), we used the Chemprop Python package [15]. The RDKit Python package was used to calculate the molecular fingerprint. Extended connectivity fingerprints with a maximum diameter parameter of 4 (ECFP4) were used [16].

Optimization of machine learning algorithms

In this study, we tested six different machine learning (ML) algorithms, including MPNN, MPNN with molecular fingerprints, MPNN with molecular descriptors, random forest (RF), support vector machine (SVM), and artificial neural network (ANN) models, to build both classification and regression models for predicting AOT in mice and rats. Here, the RF, SVM, and ANN models were trained using molecular fingerprints as input features. MPNN learns directly from a molecular graph to predict molecular properties [15]. ANN is an ML algorithm inspired by the biological neuronal network of the human brain [17]. The ANN structure consisted of an input layer, hidden layer(s), and an output layer. The ANN learns non-linear relationships from the data. RF is an ensemble learning algorithm that constructs multiple decision trees [18]. The ANN algorithm was implemented using the Keras package (version 2.2.5) with TensorFlow backend (version 2.0.0) [19]. The RF and SVM algorithms were implemented using the scikit-learn Python package [20].

To build the optimal model with the best prediction performance, hyperparameter optimization was performed. For the classification model, we selected the optimal hyperparameter that showed the highest accuracy (ACC) using the grid-search cross-validation method. In addition, for the regression model, we selected the optimal hyperparameter that showed the lowest root mean square error (RMSE) using the grid search cross-validation method.

Five metrics were used to evaluate the performance of the classification model: ACC, area under the receiver operating characteristic (AUROC), Matthew’s correlation coefficient (MCC), positive predictive value (PPV), and negative predictive value (NPV). In addition, to evaluate the regression model performance, two performance metrics were used: the RMSE and R-squared value (R2).