Timely assessment of compound toxicity is one of the biggest challenges\nfacing the pharmaceutical industry today. A significant proportion of compounds\nidentified as potential leads are ultimately discarded due to the toxicity they\ninduce. In this paper, we propose a novel machine learning approach for the\nprediction of molecular activity on ToxCast targets. We combine extreme\ngradient boosting with fully-connected and graph-convolutional neural network\narchitectures trained on QSAR physical molecular property descriptors, PubChem\nmolecular fingerprints, and SMILES sequences. Our ensemble predictor leverages\nthe strengths of each individual technique, significantly outperforming\nexisting state-of-the art models on the ToxCast and Tox21 toxicity-prediction\ndatasets. We provide free access to molecule toxicity prediction using our\nmodel at http://www.owkin.com/toxicblend.\n