Background: Quantitative Structure Activity Relationship (QSAR) methods based on machine learning play a
vital role in predicting biological effect.
Objective: Considering the characteristics of the binding interface between ligands and the inhibitory neurotransmitter
Gamma Aminobutyric Acid A(GABAA) receptor, we built a QSAR model of ligands that bind to the human GABAA
Method: After feature selection with Mean Decrease Impurity, we selected 53 from 1,286 docked ligand molecular
descriptors. Three QSAR models are built using gradient boosting regression tree algorithm based on the different
combinations of docked ligand molecular descriptors and ligand-receptor interaction characteristics.
Results: The features of the optimal QSAR model contain both the docked ligand molecular descriptors and ligand-receptor
interaction characteristics. The Leave-One-Out-Cross-Validation (Q2 LOO) of the optimal QSAR model is 0.8974, the
Coefficient of Determination (R2) for the testing set is 0.9261, the Mean Square Error (MSE) is 0.1862. We also used this
model to predict the pIC50 of two new ligands, the differences between the predicted and experimental pIC50 are -0.02 and
Conclusion: We found the BELm2, BELe2, MATS1m, X5v, Mor08v, and Mor29m are crucial features, which can help to
build the QSAR model more accurately.