ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties

Guo-Li Xiong, Zhenxing Wu, Jiacai Yi, Li Fu, Zhi-Jiang Yang, Chang-Yu Hsieh, Mingzhu Yin, Xiangxiang Zeng, Chengkun Wu, Aiping Lu, Xiang Chen, Tingjun Hou, Dong-Sheng Cao

TL;DR

ADMETlab 2.0 is a significantly upgraded web server that provides comprehensive and accurate predictions of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties for drug candidates. This new version doubles the number of predicted endpoints, incorporates a novel multi-task graph attention framework for improved model performance, and offers a more user-friendly interface with batch processing capabilities, aiming to accelerate drug discovery and development.

What Problem Does This Paper Address?

Drug development is notoriously challenging and expensive, with a high failure rate often attributed to undesirable pharmacokinetic properties and toxicity of candidate compounds. Evaluating these ADMET properties early in the discovery process is crucial to avoid costly late-stage failures. While in silico methods offer a valuable tool to assist medicinal chemists, existing platforms often suffer from limitations such as a limited number of endpoints, redundant modules, and suboptimal user experience. ADMETlab 2.0 aims to address these issues by providing a more integrated, accurate, and user-friendly platform for comprehensive ADMET prediction.

Methods

The core of ADMETlab 2.0's predictive power lies in its novel Multi-task Graph Attention (MGA) framework. This framework treats molecules as graphs, with atoms as nodes and bonds as edges. It leverages graph convolutional networks (GCNs) to learn atom and substructure representations, followed by an attention mechanism to assign varying importance to different substructures. This allows the model to generate customized "fingerprints" for specific tasks. A key innovation is its multi-task learning capability, enabling it to simultaneously learn both regression and classification tasks for ADMET prediction. This approach forces the sharing of parameters across tasks, improving generalization and allowing tasks with less data to benefit from the broader chemical space covered by tasks with more data. The MGA framework is implemented using PyTorch and DGL, with cheminformatics support from RDKit.

Datasets

The paper mentions that ADMETlab 2.0 was trained on a high-quality experimental ADMET dataset comprising approximately 0.25 million entries. This dataset was curated from various sources, including open-access bioactivity databases (ChEMBL, PubChem, OCHEM), peer-reviewed literature, and the Toxicity Estimation Software Tools (TEST) developed by the U.S. Environmental Protection Agency. Specific examples of data expansion are highlighted, such as a tenfold increase in molecules for the logD 7.4 dataset. The data underwent rigorous curation, including filtering of organometallic compounds, isomeric mixtures, neutralization of salts, and removal of molecules with more than 128 atoms.

Key Insights and Ideas

  • The power of multi-task learning for ADMET: By training a single model to predict multiple ADMET endpoints simultaneously, ADMETlab 2.0 achieves improved accuracy and efficiency compared to traditional single-task models. This is particularly beneficial for tasks with limited data.
  • Graph attention for nuanced molecular representation: The MGA framework's ability to learn and weigh substructural features allows for a more sophisticated representation of molecules, leading to better prediction of complex ADMET properties.
  • Integration is key: ADMETlab 2.0 consolidates various ADMET-related properties, including physicochemical, medicinal chemistry, ADME, and toxicity endpoints, into a single, unified platform, streamlining the drug discovery workflow.
  • User experience matters: The inclusion of batch evaluation, an intuitive interface, and clear result explanations significantly enhances the usability of the platform for medicinal chemists.

Key Results

ADMETlab 2.0 offers predictions for 88 ADMET-related parameters, nearly double the number in its predecessor. The MGA models demonstrate robust performance, with classification models achieving an average AUC of 0.863 and regression models achieving an average R² of 0.783. The platform also boasts significantly improved computation speed, with a task for 1000 molecules taking approximately 84 seconds, a substantial improvement over the previous version. The paper also highlights the successful validation of the models, including leave-cluster-out validation, indicating good generalization to novel chemical structures.

Limitations and Future Work

The authors acknowledge that while the MGA models generally outperform traditional methods, there were a few classification and regression endpoints where XGBoost models showed slightly better performance. They also note that some endpoints, like Clearance and LC50DM, had limited historical data for training, leading to underperformance of their corresponding models. Future work could involve further expanding the dataset, exploring new deep learning architectures, and potentially integrating more advanced features for drug design and optimization. The paper also mentions that the built-in database from the previous version has been removed, which might be a limitation for users who relied on its similarity searching capabilities.

Summary updated

Abstract

Because undesirable pharmacokinetics and toxicity of candidate compounds are the main reasons for the failure of drug development, it has been widely recognized that absorption, distribution, metabolism, excretion and toxicity (ADMET) should be evaluated as early as possible. In silico ADMET evaluation models have been developed as an additional tool to assist medicinal chemists in the design and optimization of leads. Here, we announced the release of ADMETlab 2.0, a completely redesigned version of the widely used AMDETlab web server for the predictions of pharmacokinetics and toxicity properties of chemicals, of which the supported ADMET-related endpoints are approximately twice the number of the endpoints in the previous version, including 17 physicochemical properties, 13 medicinal chemistry properties, 23 ADME properties, 27 toxicity endpoints and 8 toxicophore rules (751 substructures). A multi-task graph attention framework was employed to develop the robust and accurate models in ADMETlab 2.0. The batch computation module was provided in response to numerous requests from users, and the representation of the results was further optimized. The ADMETlab 2.0 server is freely available, without registration, at https://admetmesh.scbdd.com/.
BibTeX
@article{xiong2021admetlab,
  author = {Guo-Li Xiong and Zhenxing Wu and Jiacai Yi and Li Fu and Zhi-Jiang Yang and Chang-Yu Hsieh and Mingzhu Yin and Xiangxiang Zeng and Chengkun Wu and Aiping Lu and Xiang Chen and Tingjun Hou and Dong-Sheng Cao},
  title = {{ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties}},
  journal = {Nucleic Acids Res.},
  year = {2021},
  doi = {10.1093/nar/gkab255},
}
Figure