Abstarct: Abstract In this work a new data reduction method named Multiple LASSO Regression (MLAR) was proposed for the dimension reduction of data in QSAR studies. At the first, data was classified to training and test data and the MLAR models were constructed using the data generated by iterative data splitting in the absence of test data set.The descxriptors were appeared in 60% of a number of 100 LASSO Regression models were selected as the most relavent descxriptors. The number of 9 and 12 descxriptors were selected for data set A and B, respectively. Then, the selected descxriptors were used as input of Artificial Neural Network (ANN) to create QSAR models. In order to build the optimal ANN model with proper efficiency, all effective parameters of ANN modeling were optimized. The ANN model was trained with training data. set under the otimum conditions, and the biological activities of the compounds of the test set were predicted. The coefficient of determination for the test data of two datasets was equal to 0.8 and 0.93. The developed models (Multiple-LASSO-ANN) were also evaluated using Leave-One-Out, applicability domain, diversity, Y-randomization test and evaluation of statistical parameters. The evaluation results prove the generalizability and predictive power of the developed Multiple-LASSO-ANN model. Finally, a number of new chemical compounds were suggested by changing the structure of weak molecules. The biological activity values of the proposed compounds were predicted using the developed Multiple-LASSO-ANN model. In order to check the accuracy of the predicted biological activities of the suggested new compounds, molecular docking study was used, and the interaction between suggested compounds and the active site of the receptor were studied.