Anatomical Therapeutic Chemical Classification (ATC) With Multi-Label Learners and Deep Features

Abstract

Automatic anatomical therapeutic chemical (ATC) classification predicts an unknown compound's therapeutic and chemical characteristics. Predicting the organs/systems an unidentified compound will act on has the potential of expediting drug development and research. That a given compound can affect multiple organs/systems makes automatic ATC classification a complex problem. In this paper, the authors experimentally develop a multi-label ensemble for ATC prediction. The proposed approach extracts a 1D feature vector based on a compound's chemical-chemical interaction and its structural and fingerprint similarities to other compounds, as defined by the ATC coding system. This 1D vector is reshaped into 2D matrices and fed into seven pre-trained convolutional neural networks (CNN). A bidirectional long short-term memory network (BiLSTM) is trained on the 1D vector. Features extracted from both deep learners are then trained on multi-label classifiers, with results fused. The best system proposed here is shown to outperform other methods reported in the literature.