This paper explores the use of deep neural networks (DNNs) for quantitative structure–activity relationships (QSAR) in the pharmaceutical industry. DNNs, which have gained popularity in recent years due to advancements in overfitting prevention, efficient training algorithms, and improved computer hardware, are compared to random forests (RFs) on a set of large, diverse QSAR data sets from Merck's drug discovery efforts. The study finds that DNNs can consistently outperform RFs in most cases, with a mean \( R^2 \) improvement of 10% over RFs. The paper also demonstrates that a single set of recommended parameters can achieve better performance across most data sets, making DNNs a practical method for QSAR in industrial drug discovery. Additionally, the study investigates the impact of various DNN parameters, such as network architecture, activation functions, and unsupervised pretraining, and provides insights into their effects on predictive capability. The recommended DNN parameter settings are shown to be effective on additional data sets not used in the calibration, further validating their utility. The paper concludes by discussing the computational efficiency of DNNs and the potential for further improvements in parameter tuning and joint DNN training.This paper explores the use of deep neural networks (DNNs) for quantitative structure–activity relationships (QSAR) in the pharmaceutical industry. DNNs, which have gained popularity in recent years due to advancements in overfitting prevention, efficient training algorithms, and improved computer hardware, are compared to random forests (RFs) on a set of large, diverse QSAR data sets from Merck's drug discovery efforts. The study finds that DNNs can consistently outperform RFs in most cases, with a mean \( R^2 \) improvement of 10% over RFs. The paper also demonstrates that a single set of recommended parameters can achieve better performance across most data sets, making DNNs a practical method for QSAR in industrial drug discovery. Additionally, the study investigates the impact of various DNN parameters, such as network architecture, activation functions, and unsupervised pretraining, and provides insights into their effects on predictive capability. The recommended DNN parameter settings are shown to be effective on additional data sets not used in the calibration, further validating their utility. The paper concludes by discussing the computational efficiency of DNNs and the potential for further improvements in parameter tuning and joint DNN training.