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

由于候选化合物的不良药代动力学和毒性是药物开发失败的主要原因，因此，吸收、分布、代谢、排泄和毒性(ADMET)的评价应尽早得到评估。在计算机中进行的大量实验中，ADMET评价模型已被开发为辅助药物化学家设计和优化先导物的辅助工具。在这里，我们宣布了ADMETlab 2.0的发布，这是广泛使用的AMDETlab web服务器的一个完全重新设计的版本，用于预测药物动力学和化学品毒性特性，其中支持的admet相关端点的数量大约是上一个版本的两倍。包括17个理化性质，13个药用化学性质，23个ADME性质，27个毒性终点和8个毒性基团规则。采用多任务图注意框架(MGA)，在AdmetLab 2.0中开发强大和准确的模型。批量计算模块是响应于用户的批量请求提供的，并且结果进一步优化了结果的表示。

Figure 1. Workflow scheme of ADMETlab 2.0

1. Comprehensively enhanced ADMET profiles

In this update, the available ADMET profile is extended to 88 related characteristics spanning 7 different categories, roughly twice the number of its predecessor. Compared with the initial version, the number of entries for model training in the current release has almost tripled.

2. Re-engineered modules and batch evaluation support

The functional modules were re-engineered and optimized to improve the user experience. An independent module has been added for supporting batch uploading and downloading. The users could define their own criterion to promising and desirable molecules.

3. Robust and accurate MGA models

The MGA framework was employed to develop classification and regression predictors simultaneously. Deep learning makes multitask learning very natural and the combination leads to improved performance for many modeled endpoints.

4. Practical explanation and guidance

Detailed explanation and optimal range of each property are provided to help the users to get a whole ADMET picture of input molecule. The empirical-based decision states of each property are visually represented with different colored dots (green: excellent; yellow: medium; red: poor).

1. Model data

Table 1. Data information of 53 predictive models

2. MGA framework

An overview of the Multi-task Graph Attention (MGA) framework is shown in Figure 2. As shown in Figure 2, MGA is composed of input, Relation graph convolution network (RGCN) layers, attention layer and fully-connected (FC) layers. In the Input, a node represents the information of an atom, and after passing RGCN layers, the node represents general features of circular substructure centered on the atom. RGCN is an extension of the standard graph convolution network (GCN) by introducing edge features to enrich the messages used to update the hidden states in the network. The propagation rule for each node in RGCN layer is calculated via

where $h_{v}^{(l+1)}$ is the state vector of target node v after l+1 iterations and $N_{v}^{r}$ denotes the neighbors of node v under the relation (edge) $r \epsilon R$. $W_{r}^{(l)}$ is the weight for neighbor node u connecting to node v by an edge attributed with the relation $r \in R$, and $W_{0}^{(l)}$ is the weight for target node v. As can be seen above, the edge information is explicitly incorporated in a RGCN under the relation $r \in R$. The weight $W_{r}^{(l)}$ is a linear combination of basis transformation.

As shown in Figure 2B, attention layers can assign different attention weights to different substructures, and then generate the customized fingerprints (CFP) from the general features for a specific task. The attention weights and customized fingerprints are generated as follows:

where W and bias are the parameters of attention layers learned in model training, N is the number of nodes (substructures), $\omega_{v}$ is the attention weight of node (substructure) v, and $h_{v}$ is the general feature of node (substructure) v.

As shown in Figure 2A, fully-connected (FC) layers predict the corresponding tasks based on the customized toxicity fingerprints. The classification and regression tasks adopt different loss functions (loss_c and loss_r) as follows:

where $X_{n, c}$ is the predict value of molecule n for classification task c, $y_{n, c}$ is the true values of molecule n for classification task c, $p_{c}$ is the weight of positive samples, $x_{n, r}$ is the predict value of molecule n for regression task r, $y_{n, r}$ is the true value of molecule n for regression task r, N is the number of molecules, C is the number of the classification tasks, and R is the number of the regression tasks.

The loss function of MGA is a combination of loss_c and loss_r:

Figure 2. An overview of the Multi-task Graph Attention framework.

3. Model performance

Table 2. Predictive performance of regression models

Table 3. Predictive performance of classification models

Table 4. Results of leave-cluster-out validation of regression models

Table 5. Results of leave-cluster-out validation of classification models

4. Implementation

ADMETlab 2.0 was built using Python web framework of Django and deployed on an elastic compute service from Aliyun running an Ubuntu Linux system. The web access was enabled via the Nginx web server and the interactions between Django and proxy server were supported by uwsgi. This application was developed based on the Model-View-Template (MVT) framework. The model layer maps the business objects to the database objects. The view layer is a business logic layer, responsible for performing the access to the deep learning models, delivering the data to be shown to the template layer, and handling the upload and download of files. The template layer provides the visualization of results, page rendering, integration of documentation, etc. The uploaded and downloaded files, pre-trained models and model predictions are stored in the server. The prediction models were built with the Python programming language. The deep learning packages, PyTorch and DGL, were used in model implementation. Additionally, the RDKit package was employed to provide various cheminformatics support. The server has been successfully tested on the recent version of Mozilla Firefox, Google Chrome and Apple Safari.

Table 6. The development environment of ADMETlab 2.0

5. Browser Compatibility

1. Xiong, G., Wu, Z., Yi, J., Fu, L., Yang, Z., Hsieh, C., Yin, M., Zeng, X., Wu, C., Lu, A., Chen, X., Hou, T., & Cao, D. ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic Acids Res, 2021, doi: 10.1093/nar/gkab255.