In patients with depression, the use of 5-HT reuptake inhibitors can improve the condition. Topological fingerprints, ECFP4, and molecular descriptors were used. Some SERT and small molecules combined prediction models were established by using 5 machine learning methods. We selected the higher accuracy models(RF, SVM, LR) in five-fold cross-validation of training set to establish an integrated model (VOL_CLF). The training set is from Chembl database and oversampled by SMOTE algorithm to eliminate data imbalance. The unbalanced data from same sources (Chembl) was used as Test set 1; the unbalanced data with different sources(Drugbank) was used as Test set 2 . The prediction accuracy of SERT inhibitors in Test set 1 was 90.7%~93.3%(VOL_CLF method was the highest); the inhibitory recall rate was 84.6%-90.1%(RF method was the highest); the non-inhibitor prediction accuracy rate was 76.1%~80.2%(RF method is the highest); the non-inhibitor predictive recall rate is 81.2%~87.5% (SVM and VOL_CLF methods were the highest) The RF model in Test Set 2 performed better than the other models. The SERT inhibitor predicted accuracy rate, recall rate, non-inhibitor predicted accuracy rate, recall rate were 42.9%, 85.7%, 95.7%, 73.3%.This study demonstrates that machine learning methods effectively predict inhibitors of serotonin transporters and accelerate drug screening.
Download