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/.

ADMETlab 3.0 is the second updated version of the web server that provides a comprehensive and efficient platform for evaluating ADMET-related parameters as well as physicochemical properties and medicinal chemistry characteristics involved in the drug discovery process. This new release addresses the limitations of the previous version and offers broader coverage, improved performance, API functionality, and decision support. For supporting data and endpoints, this version includes 119 features, an increase of 31 compared to the previous version. The updated number of entries is 1.5 times larger than the previous version with over 400 000 entries. ADMETlab 3.0 incorporates a multi-task DMPNN architecture coupled with molecular descriptors, a method that not only guaranteed calculation speed for each endpoint simultaneously, but also achieved a superior performance in terms of accuracy and robustness. In addition, an API has been introduced to meet the growing demand for programmatic access to large amounts of data in ADMETlab 3.0. Moreover, this version includes uncertainty estimates in the prediction results, aiding in the confident selection of candidate compounds for further studies and experiments. ADMETlab 3.0 is publicly for access without the need for registration at: https://admetlab3.scbdd.com.

As one of the most popular computational approaches in modern structure-based drug design, molecular docking can be used not only to identify the correct conformation of a ligand within the target binding pocket but also to estimate the strength of the interaction between a target and a ligand. Nowadays, as a variety of docking programs are available for the scientific community, a comprehensive understanding of the advantages and limitations of each docking program is fundamentally important to conduct more reasonable docking studies and docking-based virtual screening. In the present study, based on an extensive dataset of 2002 protein–ligand complexes from the PDBbind database (version 2014), the performance of ten docking programs, including five commercial programs (LigandFit, Glide, GOLD, MOE Dock, and Surflex-Dock) and five academic programs (AutoDock, AutoDock Vina, LeDock, rDock, and UCSF DOCK), was systematically evaluated by examining the accuracies of binding pose prediction (sampling power) and binding affinity estimation (scoring power). Our results showed that GOLD and LeDock had the best sampling power (GOLD: 59.8% accuracy for the top scored poses; LeDock: 80.8% accuracy for the best poses) and AutoDock Vina had the best scoring power (r_p/r_s of 0.564/0.580 and 0.569/0.584 for the top scored poses and best poses), suggesting that the commercial programs did not show the expected better performance than the academic ones. Overall, the ligand binding poses could be identified in most cases by the evaluated docking programs but the ranks of the binding affinities for the entire dataset could not be well predicted by most docking programs. However, for some types of protein families, relatively high linear correlations between docking scores and experimental binding affinities could be achieved. To our knowledge, this study has been the most extensive evaluation of popular molecular docking programs in the last five years. It is expected that our work can offer useful information for the successful application of these docking tools to different requirements and targets.

Graph neural networks (GNN) has been considered as an attractive modelling method for molecular property prediction, and numerous studies have shown that GNN could yield more promising results than traditional descriptor-based methods. In this study, based on 11 public datasets covering various property endpoints, the predictive capacity and computational efficiency of the prediction models developed by eight machine learning (ML) algorithms, including four descriptor-based models (SVM, XGBoost, RF and DNN) and four graph-based models (GCN, GAT, MPNN and Attentive FP), were extensively tested and compared. The results demonstrate that on average the descriptor-based models outperform the graph-based models in terms of prediction accuracy and computational efficiency. SVM generally achieves the best predictions for the regression tasks. Both RF and XGBoost can achieve reliable predictions for the classification tasks, and some of the graph-based models, such as Attentive FP and GCN, can yield outstanding performance for a fraction of larger or multi-task datasets. In terms of computational cost, XGBoost and RF are the two most efficient algorithms and only need a few seconds to train a model even for a large dataset. The model interpretations by the SHAP method can effectively explore the established domain knowledge for the descriptor-based models. Finally, we explored use of these models for virtual screening (VS) towards HIV and demonstrated that different ML algorithms offer diverse VS profiles. All in all, we believe that the off-the-shelf descriptor-based models still can be directly employed to accurately predict various chemical endpoints with excellent computability and interpretability.

By using different evaluation strategies, we systemically evaluated the performance of Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) and Molecular Mechanics/Poisson–Boltzmann Surface Area (MM/PBSA) methodologies based on more than 1800 protein–ligand crystal structures in the PDBbind database. The results can be summarized as follows: (1) for the one-protein-family/one-binding-ligand case which represents the unbiased protein–ligand complex sampling, both MM/GBSA and MM/PBSA methodologies achieve approximately equal accuracies at the interior dielectric constant of 4 (with r_p = 0.408 ± 0.006 of MM/GBSA and r_p = 0.388 ± 0.006 of MM/PBSA based on the minimized structures); while for the total dataset (1864 crystal structures), the overall best Pearson correlation coefficient (r_p = 0.579 ± 0.002) based on MM/GBSA is better than that of MM/PBSA (r_p = 0.491 ± 0.003), indicating that biased sampling may significantly affect the accuracy of the predicted result (some protein families contain too many instances and can bias the overall predicted accuracy). Therefore, family based classification is needed to evaluate the two methodologies; (2) the prediction accuracies of MM/GBSA and MM/PBSA for different protein families are quite different with r_p ranging from 0 to 0.9, whereas the correlation and ranking scores (an averaged r_p/r_s over a list of protein folds and also representing the unbiased sampling) given by MM/PBSA (r_p-score = 0.506 ± 0.050 and r_s-score = 0.481 ± 0.052) are comparable to those given by MM/GBSA (r_p-score = 0.516 ± 0.047 and r_s-score = 0.463 ± 0.047) at the fold family level; (3) for the overall prediction accuracies, molecular dynamics (MD) simulation may not be quite necessary for MM/GBSA (r_p-minimized = 0.579 ± 0.002 and r_p-1ns = 0.564 ± 0.002), but is needed for MM/PBSA (r_p-minimized = 0.412 ± 0.003 and r_p-1ns = 0.491 ± 0.003). However, for the individual systems, whether to use MD simulation is depended. (4) both MM/GBSA and MM/PBSA may be unable to give successful predictions for the ligands with high formal charges, with the Pearson correlation coefficient ranging from 0.621 ± 0.003 (neutral ligands) to 0.125 ± 0.142 (ligands with a formal charge of 5). Therefore, it can be summarized that, although MM/GBSA and MM/PBSA perform similarly in the unbiased dataset, for the currently available crystal structures in the PDBbind database, compared with MM/GBSA, which may be used in multi-target comparisons, MM/PBSA is more sensitive to the investigated systems, and may be more suitable for individual-target-level binding free energy ranking. This study may provide useful guidance for the post-processing of docking based studies.