Predicting Organ-Specific Toxicity of Selective Androgen Receptor Modulators, using Transfer Learning on Graph Convolutional Networks

Abstract

Novel Quantitative Structure-Activity Relationship (QSAR) models were constructed using Graph Convolutional Networks (GCNs), to predict Drug-Induced Liver Injury (DILI), Drug-Induced Renal Injury (DIRI) and Drug-Induced Cardiotoxicity (DICT) of Selective Androgen Receptor Modulators (SARMs) - an emerging class of performance-enhancing drugs. Prior to training on DILI, DIRI and DICT datasets, the GCN QSAR models were first pre-trained on a variety of unrelated biomedical assay datasets, as an attempt to improve model performance via transfer learning. The success of the transfer learning was mixed; model performances were measurably improved via pre-training on certain datasets, by statistically weak increases. The optimal final QSAR models achieved overall accuracy scores of 68% for DILI (no significant improvement via ensemble modelling), 76% for DIRI (improved to 77% via ensemble modelling) and 65% for DICT (improved to 67% via ensemble modelling). Application of the most optimal singular models to a dataset of 25 SARMs predicted that 21 of the 25 SARMs are either DILI-positive, DIRI-positive, or both - which raises concern, given the rising use of SARMs. All SARMs except for one were predicted as DICT-negative. A novel definition of the Applicability Domain (AD) was used, intended for close relevance to the models, via generating three-dimensional graph embeddings, for each model. Convex hulls were fitted around training data embeddings, with a {+/-}10% buffer, defining the AD as the region of embedded chemical space covered by the convex hull, for a given model. Subsequent analysis found that a majority of DILI, DIRI and DICT testing data lay within the AD, alongside a majority of the SARMs - adding consensus to the reliability of the predictions.