Abstract
Absorption, distribution, metabolism, and excretion (ADME) studies arecritical for drug discovery. Conventionally, these tasks, together with otherchemical property predictions, rely on domain-specific feature descriptors, orfingerprints. Following the recent success of neural networks, we developedChemi-Net, a completely data-driven, domain knowledge-free, deep learningmethod for ADME property prediction. To compare the relative performance ofChemi-Net with Cubist, one of the popular machine learning programs used byAmgen, a large-scale ADME property prediction study was performed on-site atAmgen. The results showed that our deep neural network method improved currentmethods by a large margin. We foresee that the significantly increased accuracyof ADME prediction seen with Chemi-Net over Cubist will greatly accelerate drugdiscovery.