Objective: Early phase of drug discovery studies include a virtual screening phase of detecting active molecules among a large number of small drug molecules. The number of publicly available datasets for drug molecules are growing exponentially every year thanks to the databases, such as PubChem and ChEMBL. Therefore, there is a strong need for analyzing and retrieving useful information from these datasets using automated processes. For this purpose, machine learning algorithms are often used for activity prediction of small drug compounds, since they are faster and comparatively cheaper. Deep neural networks has emerged as a powerful machine learning method with great advantages to deal with high-dimensional big datasets. Material and Methods: In this study, we applied different settings of deep neural networks models to reveal the effects of learning rate, batch size and minority class weight on performance of the network. Results: Small learning rate and large batch size are found to be the most important factors that improve performance of the deep neural network. The best performed model yielded 89% accuracy and 0.78 area under the curve value. Conclusion: Findings of this study is promising for use of deep neural networks in virtual screening of small drug compounds from publicly available databases.
Full Citation: Korkmaz S. Small Drug Molecule Classification Using Deep Neural Networks. Turkiye Klinikleri J Biostat. (Accepted/In press).