Search

Kinase inhibitors are widely used in antitumor research, but there are still many problems such as drug resistance and off-target toxicity. A more suitable solution is to design a multitarget inhibitor with certain selectivity. Herein, computational and experimental studies were applied to the discovery of dual inhibitors against FGFR4 and EGFR. A quantitative structure-property relationship (QSPR) study was carried out to predict the FGFR4 and EGFR activity of a data set consisting of 843 and 5088 compounds, respectively. Four different machine learning methods including support vector machine (SVM), random forest (RF), gradient boost regression tree (GBRT), and XGBoost (XGB) were built using the most suitable features selected by the mutual information algorithm. As for FGFR4 and EGFR, SVM showed the best performance with R2test-FGFR4 = 0.80 and R2test-EGFR = 0.75, demonstrating excellent model stability, which was used to predict the activity of some compounds from an in-house database. Finally, compound 1 was selected, which exhibits inhibitory activity against FGFR4 (IC50 = 86.2 nM) and EGFR (IC50 = 83.9 nM) kinase, respectively. Furthermore, molecular docking and molecular dynamics simulations were performed to identify key amino acids for the interaction of compound 1 with FGFR4 and EGFR. In this paper, the machine-learning-based QSAR models were established and effectively applied to the discovery of dual-target inhibitors against FGFR4 and EGFR, demonstrating the great potential of machine learning strategies in dual inhibitor discovery.

Protein kinases are important drug targets in several therapeutic areas ,and structure-based virtual screening (SBVS) is an important strategy in discovering lead compounds for kinase targets. However, there are multiple crystal structures available for each target, and determining which one is the most favorable is a key step in molecular docking for SBVS due to the ligand induce-fit effect. This work aimed to find the most desirable crystal structures for molecular docking by a comprehensive analysis of the protein kinase database which covers 190 different kinases from all eight main kinase families. Through an integrated self-docking and cross-docking evaluation, 86 targets were eventually evaluated on a total of 2608 crystal structures. Results showed that molecular docking has great capability in reproducing conformation of crystallized ligands and for each target, the most favorable crystal structure was selected, and the AGC family outperformed the other family targets based on RMSD comparison. In addition, RMSD values, GlideScore, and corresponding bioactivity data were compared and demonstrated certain relationships. This work provides great convenience for researchers to directly select the optimal crystal structure in SBVS-based kinase drug design and further validates the effectiveness of molecular docking in drug discovery.

Human pharmacokinetics is of great significance in the selection of drug candidates, and in silico estimation of pharmacokinetic parameters in the early stage of drug development has become the trend of drug research owing to its time- and cost-saving advantages. Herein, quantitative structure-property relationship studies were carried out to predict four human pharmacokinetic parameters including volume of distribution at steady state (VDss), clearance (CL), terminal half-life (t1/2), and fraction unbound in plasma (fu), using a data set consisting of 1352 drugs. A series of regression models were built using the most suitable features selected by Boruta algorithm and four machine learning methods including support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), and XGBoost (XGB). For VDss, SVM showed the best performance with R2test = 0.870 and RMSEtest = 0.208. For the other three pharmacokinetic parameters, the RF models produced the superior prediction accuracy (for CL, R2test = 0.875 and RMSEtest = 0.103; for t1/2, R2test = 0.832 and RMSEtest = 0.154; for fu, R2test = 0.818 and RMSEtest = 0.291). Assessed by 10-fold cross validation, leave-one-out cross validation, Y-randomization test and applicability domain evaluation, these models demonstrated excellent stability and predictive ability. Compared with other published models for human pharmacokinetic parameters estimation, it was further confirmed that our models obtained better predictive ability and could be used in the selection of preclinical candidates.

Human ether-a-go-go-related gene (hERG) K+ channel blockage may cause severe cardiac side-effects and has become a serious issue in safety evaluation of drug candidates. Therefore, improving the ability to avoid undesirable hERG activity in the early stage of drug discovery is of significant importance. The purpose of this study was to build predictive models of hERG activity by deep neural networks. For each combination of sampling methods and descriptors, deep neural networks with different architectures were implemented to build classification models. The optimal model M15 with three hidden layers, undersampling method, and 2D descriptors yielded the prediction accuracy of 0.78 and F1 score of 0.75 on the test set as well as accuracy of 0.77 and F1 score of 0.34 on the external validation set, outperforming the other 35 models including 9 random forest models. Particularly, the optimal model M15 achieved the highest F1 score and the second highest accuracy when compared with other five methods from four groups using different machine learning algorithms with the same external validation set. It can be believed that this model has powerful capability on prediction of hERG toxicity, which is of great benefit for developing novel drug candidates.

Data mining methods based on machine learning play an increasingly important role in drug design and discovery. In the current work, eight machine learning methods including decision trees, k-Nearest neighbor, support vector machines, random forests, extremely randomized trees, AdaBoost, gradient boosting trees, and XGBoost were evaluated comprehensively through a case study of ACC inhibitor data sets. Internal and external data sets were employed for cross-validation of the eight machine learning methods. Results showed that the extremely randomized trees model performed best and was adopted as the first step of virtual screening. Together with structure-based virtual screening in the second step, this combined strategy obtained desirable results. This work indicates that the combination of machine learning methods with traditional structure-based virtual screening can effectively strengthen the ability in finding potential hits from large compound database for a given target.

The absorption, distribution, metabolism and excretion properties are important for drugs, and prediction of these properties in advance will save the cost of drug discovery substantially. The ability to penetrate the blood–brain barrier is critical for drugs targeting central nervous system, which is represented by the ratio of its concentration in brain and in blood. Herein, a quantitative structure–property relationship study was carried out to predict blood–brain partitioning coefficient (logBB) of a data set consisting of 287 compounds. Four different methods including support vector machine, multivariate linear regression, multivariate adaptive regression splines and random forest were employed to build prediction models with 116 molecular descriptors selected by Boruta algorithm. The RF model had best performance in training set ( $$ R^{2} $$ = 0.938), test set ( $$ R^{2} $$ = 0.840) and tenfold cross-validation ( $$ Q^{2} $$ = 0.788). Finally, we found that the polar surface area and octanol–water partition coefficient have the greatest influence on blood–brain partitioning. Results suggest that the proposed model is a useful and practical tool to predict the logBB values of drug candidates.