We utilized publicly available datasets that were collected either from previously published papers or from bioactivity databases, such as ChEMBL, BindingDB, and Metrabase, for model building and evaluation (Mak et al., 2015; Gilson et al., 2016; Jiang et al., 2020; Zdrazil et al., 2024). For the 12 transporters used in this work, we collected separate data for inhibitors and substrates. Supplementary Table S1 provides a detailed list of data sources for each transporter in this study. We pre-processed the simplified molecular-input line-entry system (SMILES) using the ChEMBL structure pipeline, which is comprised of three functions, namely, checker, standardizer, and salt strip, to check for validity of chemical structure, format compounds to standardized conventions, and strip salts, respectively (Bento et al., 2020). We created a final pre-processed dataset for model building after evaluation and removal of duplicate compounds. The final set of pre-processed data used for model building is provided on the GitHub page (https://github.com/bhsai/monstrous).

We utilized Chemprop, an open-source software, to develop the GCNN models (Yang et al., 2019; Heid et al., 2024). This program uses a directed message-passing neural network (D-MPNN) to generate molecular features (Heid et al., 2024). The D-MPNN models the molecular structure as a graph, where atoms are represented as nodes and bonds as edges. Each node and edge is assigned a feature vector that captures the characteristics of the corresponding atom and bond. The Chemprop approach involves two phases: the message-passing phase and the readout phase (Yang et al., 2019). In the message-passing phase, the D-MPNN iteratively refines the atomic- and bond-level features based on information from neighboring nodes and edges during each convolution operation. Ultimately, in the readout phase, the compound’s learned representation is generated using an aggregation function that combines the final updated features at both the atom and bond levels (Heid and Green, 2022). This learned representation is then fed into a feed-forward neural network, which uses it as the input feature vector to predict the compound’s activity. In this approach, the molecular representation is automatically learned by the program, eliminating the need for pre-defined chemical fingerprints.

The program processes a list of SMILES strings and their corresponding activity values, which are provided as 1 for active molecules and 0 for inactive ones, in a CSV format. In this study, we selected “classification” as the modeling type and used five-fold and 10-fold cross-validations for the “number of folds” option. Each run involved developing 10 ensemble models over 30 epochs. For the remaining model development parameters, we used the default values: a depth value of three, i.e., the number of message-passing steps in D-MPNN, the ReLu activation function, 300 hidden neurons, and two layers for the feed-forward neural network.

We carried out five-fold and 10-fold cross-validations as well as a scaffold split-based validation to understand model performance. In the five-fold cross-validation procedure, we split the dataset into five groups and left one group out; subsequently, we used the model built from the compounds in the remaining four groups to predict the compounds in the left-out group. Once we completed this prediction cycle by leaving out each of the five groups, we calculated the model evaluation parameters: the receiver operating characteristic (ROC) area under the curve (AUC) and the Matthews correlation coefficient (MCC). This approach of leaving out groups is commonly used during model development to evaluate the model’s ability to predict the activity of compounds not seen during training and assess its performance on unseen data. In the 10-fold cross-validation, we repeated the same process by dividing the data into 10 groups and leaving out one group. The results presented here are averages across cross-validation folds. Such a cross-validation study summarizes generalizability and robustness of the model. We also performed a scaffold split and created training and test sets. We developed the model, as described above, using the scaffold-split training data and evaluated the performance using the scaffold-split test data. We calculated the following metrics: sensitivity (also known as the recall or true positive rate), the ability to correctly predict positive results; specificity (also known as the true negative rate), the ability to correctly predict negative results; accuracy, the total percentage correctly predicted; and MCC. These parameters are defined as follows (Equations 1–4): Sensitivity = TP TP + FN (1) Specificity = TN TN + FP (2) Accuracy = TP + TN TP + TN + FP + FN (3) MCC = TP . TN   ‐   FP . FN TP + FP . TP + FN . TN + FP . TN + FN (4) where TP represents true positive, TN denotes true negative, FP represents false positive, and FN denotes false negative. We calculated balanced accuracy, which is the mean of sensitivity and specificity. We generated ROC and precision-recall (PR) curves and calculated the ROC-AUC and PR-AUC, respectively.

We developed GCNN models for transporters with sufficient data and implemented a similarity-based screening approach for the other transporters. The 2D chemical similarity approach uses atom connectivity/fingerprints and has been shown to perform well in retrieving related compounds (AbdulHameed et al., 2021). We evaluated the performance of the 2D similarity approach by examining whether the reference set of compounds for a particular target was able to identify chemicals already known to interact with the target when mixed with inactive compounds. We collected the reference set of compounds for 10 different targets from the DrugBank database and the corresponding external test dataset for each target from the ChEMBL database. We downloaded the largest available bioactivity data of the same type (IC₅₀ or EC₅₀ or K_i) for each target. Most of the targets had IC₅₀ values. We labelled compounds with activity values ≤1 μM as actives and ≥10 μM as inactives. Such binary thresholds are commonly used in cheminformatics studies (Chen et al., 2023). We pre-processed the data and removed duplicate compounds as well as overlapping compounds that were present in the reference set for that target. We performed screening for each target and calculated the performance using ROC-AUC values. Higher AUC values represented improved screening, whereas an AUC closer to 0.5 indicated that the approach was not able to separate active from inactive compounds.

In this work, we selected the known inhibitors or substrates of each transporter as the reference set of compounds for that transporter. Then, we calculated the similarity between each query compound and the reference set of compounds. Finally, we used the maximum similarity score (MAX) between them to represent the potential of the query compound to interact with the transporter.

We previously developed an applicability domain method, the sum of distance-weighted contributions (SDC), that uses a distance-to-training set approach (Liu and Wallqvist, 2019). Basically, this class of approaches evaluates the chemical space for which the model will make reliable predictions. It uses the weighted distance between the query compound and all the molecules in the reference set to define the applicability domain (Liu and Wallqvist, 2019). SDC is defined as in Equation 5 SDC = ∑ i = 1 n e − 3 T D i 1 − T D i (5) where TD _i represents the Tanimoto distance (TD) between a target molecule and the ith training molecule and n represents the total number of training molecules (Liu and Wallqvist, 2019). We calculated the TD between two molecules using the RDKit Morgan fingerprint with a radius of 2 (RDKit, 2024). The TD value between two compounds ranges from 0 to 1, and the higher the TD value, the lower the similarity between the compounds. We implemented the Python version of SDC and utilized it to define the applicability domain for all 24 transporter models developed so far. The script used to calculate SDC is provided on the GitHub page (https://github.com/bhsai/monstrous).

The MONSTROUS web application runs on an Apache Tomcat server and utilizes a multi-tiered architecture consisting of a model inference engine, a front-end, a database, and a controller. At the core of the application is the model inference engine implemented as Python scripts that runs the GCNN models and the similarity-based screen. The user can submit up to 10,000 compounds at a time as SMILES through the front-end user interface (UI). The UI consists of forms for user inputs and pages to display model information and graphical outputs. The UI is built on the Flutter framework and is compatible with most modern web browsers and devices of varied screen sizes, including smartphones. The application back-end is a relational database hosted on PostgreSQL platform. The Java-based controller ties the front-end, the back-end, and the model inference engine together by providing pipes for efficient dataflow across the application. Additionally, the controller manages user authentication, session management, and job scheduling. Users can view the compounds they submitted and download the computed results for up to 2 weeks, after which users’ compounds and results are deleted from the database. None of user data, including compounds and results, is used to train new models or shared with anyone else at any point of time. The model inference engine along with the models developed in this work and the raw data used to train the models are available for download as a command-line tool at https://github.com/bhsai/monstrous. This command-line tool provides a batch screening option without any limit on the number of input query compounds. The MONSTROUS web application is publicly available at https://monstrous.bhsai.org/.

We collected publicly available datasets associated with 12 transporters recommended for screening by regulatory agencies, including five transporters from the ABC family and seven transporters from the SLC family (Table 1). Overall, BCRP and Pgp, which are known to be associated with cancer-induced multi-drug resistance, had the greatest data availability, with more than 2,000 compounds with associated transporter inhibition data (Figure 2). The amount of available data associated with transporter substrates was significantly lower than that for inhibitors (Figure 2). BSEP is a well-known transporter associated with bile transport, and its inhibition leads to cholestasis. While we were able to collect some known substrates for this transporter, we were not able to obtain compounds that are non-substrates. BCRP and Pgp had the largest number of compounds with substrate activity data. Overall, the sparse publicly available bioactivity data for transporters highlights the challenge in creating a computational tool for rapid transporter liability screening. In this work, to address this bottleneck, we developed machine learning models for those transporters with sufficient bioactivity data and implemented a 2D similarity-based screening approach for the others. Table 1 lists the 12 transporters used in this work along with information on whether the web tool uses machine learning models or the similarity-based approach to make predictions on new query compounds. Our final set includes 24 transporter models, nine of which are GCNN models and 15 of which utilize the similarity-based screening approach.

Machine learning models are widely used for predicting various absorption, distribution, metabolism, excretion, and toxicity endpoints (Wang et al., 2021). Typically, to develop such models, the structure of the compounds is represented using fingerprints or molecular properties. Fingerprints capture the presence of predefined sets of chemical substructures/functional groups to represent the chemical structure. More recently, GCNN-based approaches provide an alternative approach as they allow us to learn the representation of chemical structures in an automated manner (Chuang et al., 2020). Others have performed a detailed analysis of benchmark datasets and have shown that GCNNs perform better than other machine learning approaches for predicting the bioactivity of compounds (Yang et al., 2019). Although the use of GCNNs has been reported to be a powerful approach, so far they have not been widely used in predicting transporter interactions. In this work, we developed six transporter inhibitor models and three transporter substrate models using the GCNN approach. Our inhibitor GCNN models include four ABC transporters (Pgp, BCRP, MRP1, BSEP) and two SLC transporters (OATP1B1, OATP1B3). We developed three substrate GCNN models for Pgp, BCRP, and MRP1. We used five-fold and 10-fold cross-validation analyses and obtained performance evaluation metrics for these models. Figure 3 provides the five-fold and 10-fold cross-validation performance metrics for the inhibitor models. All six transporter inhibitor models had ROC-AUC >0.7 and MCC >0.2 in both cross-validations. The ROC-AUC values in the 10-fold cross-validation analysis are on par with the performance reported for the Vienna LiverTox web tool. As mentioned in the Introduction, Vienna LiverTox is a web-based tool for screening transporter interactions related to liver injury. Figure 4 provides the five-fold and 10-fold cross-validation performance metrics for the substrate models. Similar to the inhibitor models, all three substrate models had ROC-AUC >0.7 and MCC >0.2 in both cross-validations.

Finally, we evaluated model performance by creating an external test set using the scaffold-split approach. Scaffold split-based analysis provides an alternate way to evaluate model performance. Figure 5 shows the scaffold split-based evaluation results for the inhibitor and substrate models. All the inhibitor and substrate models demonstrated reasonable performance, with average ROC-AUC >0.7 and balanced accuracy >0.5. Among the inhibitor models, the MRP1 and BCRP models exhibited the highest balanced accuracy, whereas the OATP1B1 and OATP1B3 models had the lowest, with values closer to 65%. Similarly, for the substrate models, the Pgp model achieved the highest balanced accuracy, while the MRP1 model showed the lowest. The inhibitor models overall performed slightly better than the substrate models, likely due to the availability of more training data for inhibitors. The final models were built using all available data and integrated into the MONSTROUS web tool to facilitate easy access. Additionally, we implemented the SDC approach to define the applicability domain and provide clear guidance on where predictions are most reliable.

Overall, we utilized GCNN, a well-established approach for property prediction, to develop models for transporters recommended by regulatory agencies and integrated them into an easy-to-use web interface. As a limitation, we acknowledge that this study primarily focused on tool development and has exclusively used the GCNN approach. Future work will explore comparisons with other available methods to further enhance the tool’s predictive performance.

Many transporters do not have sufficient data, particularly with regard to compounds that are inactive as inhibitors or substrates. In such cases, it is not possible to develop machine learning models. In order to provide a comprehensive tool that covers all regulatory-relevant transporters, we implemented a 2D similarity-based screening tool. First, we tested whether such an approach is reasonable using a test evaluation of 10 targets with sufficient data. Figure 6 and Supplementary Table S3 provide a list of the 10 targets evaluated. We evaluated the ability of a small set of known reference compounds (active compounds) for these 10 targets to screen an external set of active and inactive compounds from ChEMBL using the 2D similarity approach. We calculated the MAX Tanimoto scores between the reference and query compounds. Figure 6 shows the ROC-AUC values for these targets. We found that eight of the 10 targets had ROC-AUC >0.6, indicating that the 2D similarity approach was able to retrieve active compounds from inactive compounds and can serve as a screening tool for transporters with low publicly available data. As highlighted in Table 1, we implemented this 2D similarity approach in inhibitor mode for six transporters (MRP2, OAT1, OAT3, OCT2, MATE-1, MATE-2K) and in substrate mode for nine transporters (MRP2, BSEP, OATP1B1, OATP1B3, OAT1, OAT3, OCT2, MATE-1, MATE-2K). Similarity-based models rely on the presence of structurally similar molecules in the reference set, which may lead to less reliable predictions for novel or unique chemical scaffolds not represented in the reference set. However, the hybrid approach of combining GCNN and similarity-based models is practically useful, as it enables users to screen against all 12 transporters within a single tool. As more data become available, we plan to upgrade the models in future versions of the tool to enhance prediction accuracy.

One of the main goals of this work was to create an easy-to-use, publicly available, web-based computational tool for rapid screening and prediction of the potential of chemicals to interact with transporters. Here, we created one such web interface that allows easy access to the deep learning models and similarity-based screening tools as a means to rapidly identify chemicals with potential to interact with transporters either as an inhibitor or a substrate.

The MONSTROUS website is publicly available online at https://monstrous.bhsai.org/. The login page (Figure 7) allows users to create a new account or log in to an existing or guest account. After login, users can submit query compounds in SMILES format by typing directly, uploading a CSV file, or using the Marvin chemical drawing tool (Chemaxon, Boston, MA). Job status can be monitored on the homepage. Figure 8 shows a snapshot of the output, which provides separate views for transporter inhibition and substrate predictions as well as the ability to toggle between them. Users also have the option to view the results with the applicability domain. The output is color-coded for easy interpretation: red indicates active (inhibitor or substrate) and green indicates inactive. The predictions outside the applicability domain are shown in striped red and green cells. Selecting the applicability domain option will show these output cells in gray. Users have the option to download the results as an Excel heatmap or as raw data. Both options provide a table with separate sheets for inhibitor and substrate predictions, with and without the applicability domain. The heatmap is color-coded to represent the potential of query chemicals to interact with the transporter (red) or to have no interaction (green).

In addition to the web application, users with computational expertise can utilize the command-line version of the tool at https://github.com/bhsai/monstrous. Users can download, install, and locally run this command-line tool on their infrastructure. The web application allows batch screening of up to 10,000 compounds at a time, while the command-line version provides batch screening with no limit on the number of input query compounds.

In order to demonstrate the utility of MONSTROUS, we conducted an example analysis using a dataset of 95 well-characterized kinase inhibitors (Supplementary Table S2). These compounds were part of the IDG-DREAM Drug-Kinase Binding Prediction Challenge, where their interactions with 295 kinases were experimentally profiled and made publicly available (Cichońska et al., 2021). FDA-approved kinase inhibitors, such as imatinib and gefitinib, are known to interact with ABC transporters, highlighting the importance of evaluating transporter liability. While the 95 compounds from the IDG-DREAM challenge have been extensively studied for off-target kinase interactions, their transporter liability remains unknown. We anticipate that in silico tools like MONSTROUS will help bridge this gap. Figure 9 shows the prediction results for these 95 compounds using the MONSTROUS tool. Our predictions indicate that of the 95 compounds, 13 are potential inhibitors of MRP1, three inhibit OATP1B1, and 11 are substrates for BCRP. This analysis highlights how MONSTROUS can provide insights on potential transporter liability for query compounds of interest.