The identification of potential side effects for promising drugs is one of the critical stage in drug development. The costly and time-consuming process in elucidating adverse effects of drugs has become a major and severe bottleneck in drug development. Traditional drug design with one-drug one-target tends to overlook system-wide effects. Recent study on drug side effect prediction has shown the trend from independent analysis to systematic investigation on side effects profiles of drugs. Therefore, developing a new approach which is capable of detecting potential drug side effects systematically is an urgent need for improving the process of side effect identification and at the same time providing efficient evaluation schemes in drug development.In this article, we investigate the relationship between potential side effects of drug candidates and their chemical structures. We propose a novel and efficient model (SR) for drug side-effect prediction. The promising feature of the method lies in the efficiency of obtaining model parameters in training where the closed form solution exists. The primary foundation on regression of the model creates a golden opportunity for evaluating drug side-effect profiles efficiently. An improved version of the model (NSR) with regularization further improves the prediction accuracy. The usefulness of the proposed method is demonstrated in a cross validation setting through prediction of 1385 side-effects in the SIDER database from the chemical structures of 888 approved drugs. Remarkably, our new method exhibits high efficiency as well as good performance in accuracy compared to some of state-of-the-art methods. Theoretical analysis on regularization parameter is conducted to explain the possible role it played in improving both prediction accuracy and efficiency.