Molecular Activity Prediction using Deep Learning Software Library

In order to predict the molecular activities, the molecular fingerprints (chemical descriptor vectors) provided by the “Merck molecular activity challenge” competition and an open source deep learning library Chainer are used. Almost identical increase-decrease tendencies with the correlation Rs2 of the champion group in the competition are reproduced. GPU performance is also examined and the speed gain obtained is more than 11 times than only CPU computation.