Abstract: It is expensive and time-consuming to measure oral bioavailability using traditional methods,and existing machine learning methods show lower accuracy.To solve the problems,a classification method of human oral bioavailability based on stacked autoencoder(SAE) was presented.Filtered features of molecular are combined with the model of SAE to classify human oral bioavailability.Experimental results show that the deep network can study more essential features of molecules comparing with other shallow learning models like support vector machine and artificial neural network,and the combination of screened 2D and 3D molecular features achieves better classification effect of oral bioavailability,with a average accuracy value of 83%,a sensitivity(SE) value of 94% and a specificity(SP) value of 49%.

Key words: Oral bioavailability,Deep learning,Molecular descriptors,Stacked autoencoder,softmax regression