Predicting Molecule Activity with Its Structure: Is Deep Neural Network an Evolutionary Solution or Revolutionary One?
Quantitative structure-activity relationship (QSAR) is a technique used in pharmaceutical industry to find out a molecule’s efficacy and safety from its molecule structures using computer models, instead of doing lab tests. A number of independent efforts, including ours, have showed that Deep Neural Network (DNN) can outperform existing QSAR methods. However, the improvements are frequently only marginal. Thus, using DNN can only be claimed as an evolution in QSAR. Our recent efforts shed lights on the reason why some DNN network structures work better. This new insight can further enhance DNN performance. However, in order to make DNN a revolutionary solution for QSAR, the need to investigate into a number of other research directions is discussed.
Dr. Junshui Ma is a Sr. Principal Scientist at Merck Research Laboratories, Merck & Co., Inc. His research focuses on several areas in statistics and machine learning. Since joined Merck in 2005, he has worked on programs across all phases of drug research and development (R&D), from biomarker valiation and preclinical research to early- and late-stage clinical development. His unusual exposure to data-related issues in all aspects of drug R&D allowed him to deliver noval solutions to problems in various drug R&D areas, and to publish in diverse scientific domains. Junshui obtained his Ph.D. from Ohio State University in 2001. He worked in Los Alamos National Labobratories, a bio-tech startup company, and Ohio Supercomputing Center before joining Merck.