Matrix factorization-based data fusion for drug-induced liver injury prediction

Traditional studies of liver toxicity involve screening compounds through in vivo and in vitro tests. They need to distinguish between compounds that represent little or no health concern and those with the greatest likelihood to cause adverse effects in humans. High-throughput and toxicogenomic screening methods coupled with a plethora of circumstantial evidence provide a challenge for improved toxicity prediction and require appropriate computational methods that integrate various biological, chemical and toxicological data. We report on a data fusion approach for prediction of drug-induced liver injury potential in humans using microarray data from the Japanese Toxicogenomics Project (TGP) as provided for the contest by CAMDA 2013 Conference. Our aim was to investigate if the data from different TGP studies could be fused together to boost prediction accuracy. We were also interested if in vitro studies provided sufficient information to refrain from studies in animals. We show that our recently proposed matrix factorization-based data fusion provides an elegant computational framework for integration of the TGP and related data sets, 29 data sets in total. Fusion yields a high cross-validated accuracy (AUC of 0.819 for in vivo assays), which is above the accuracy of the established machine learning procedure of stacked classification with feature selection. Our data analysis shows that animal studies may be replaced with in vitro assays (AUC = 0.799) and that liver injury in humans can be predicted from animal data (AUC = 0.811). Our principal contribution is a demonstration that analysis of toxicogenomic data can substantially benefit from data fusion with directly and circumstantially related data sets.