Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2020.10.02.322917v1?rss=1 Authors: Green, A. J., Mohlenkamp, M. J., Das, J., Chaudhari, M., Truong, L., Tanguay, R. L., Reif, D. M. Abstract: There are currently 85,000 chemicals registered with the EPA under the Toxic Substances Control Act, but only a small fraction (~5%) have any measured toxicological data. To address this data gap, high-throughput screening (HTS) methods are vital. As part of one such HTS effort, embryonic zebrafish were used to examine a suite of morphological and mortality endpoints at six concentrations from over 1,000 unique chemicals found in the ToxCast library (phase 1 and 2). We hypothesized that by using a conditional Generative Adversarial Network (cGAN) and leveraging this large set of toxicity data, plus chemical structure information, we could efficiently predict toxic outcomes of untested chemicals. CAS numbers for each chemical were used to generate textual files containing three-dimensional structural information for each chemical in the Protein Data Bank (PDB) file format. Utilizing a novel method in this space, we converted the 3D structural information into a weighted set of points while retaining all information about the structure. The in vivo toxicity and chemical data were used to train two neural network generators. The first used regression to train a generator (Go-ZT) to produce toxicity data while the second utilized cGAN architecture to train a generator (GAN-ZT). Our results showed that both Go-ZT and GAN-ZT models produce similar results, but the cGAN achieved higher sensitivity (SE) value of 85.7% vs 71.4%. Conversely, Go-ZT attained a higher specificity (SP), positive predictive value (PPV), and Kappa results of 67.3%, 23.4%, and 0.21 compared to 24.5%, 14.0%, and 0.03 for the cGAN, respectively. By combining both Go-ZT and GAN-ZT, our consensus model improved the SP, PPV, and Kappa, to 75.5%, 25.0%, and 0.211, respectively, resulting in an area under the receiver operating characteristic (AUROC) of 0.663. Considering their potential use as efficient prescreening tools, these new models can provide in vivo toxicity predictions and insight into as-yet untested areas of chemical space to prioritize compounds for HT testing. Copy rights belong to original authors. Visit the link for more info