Data Integration with Self-organising Neural Network Reveals Chemical Structure and Therapeutic Effects of Drug ATC Codes

Abstract

Anatomical Therapeutic Codes (ATC) are a drug classification system which is extensively used in the field of drug development research. There are many drugs and medical compounds that as yet do not have ATC codes, it would be useful to have codes automatically assigned to them by computational methods. Our initial work involved building feedforward multi-layer perceptron models (MLP) but the classification accuracy was poor. To gain insights into the problem we used the Kohonen self-organizing neural network to visualize the relationship between the class labels and the independent variables. The information gained from the learned internal clusters gave a deeper insight into the mapping process. The ability to accurately predict ATC codes was unbalanced due to over and under representation of some ATC classes. Further difficulties arise because many drugs have several, quite different ATC codes because they have many therapeutic uses. We used chemical fingerprint data representing a drugs chemical structure and chemical activity variables. Evaluation metrics were computed, analysing the predictive performance of various self-organizing models.