Cell Painting morphological profiles can complement QSAR models for rat acute oral toxicity prediction

Small molecule discovery entails the identifying and developing of chemical compounds with optimized safety properties in different target species including laboratory animals, human and environmental species. Early de-risking supports the identification of candidates with improved safety profiles in the development of new chemical compounds. In vivo studies are performed in the early assessment of acute oral toxicity, an important endpoint in the development of crop protection products. Because in vivo studies are long, costly, and raise ethical concerns, non-animal alternatives are needed. Several models were analyzed to classify compounds as highly acutely toxic (LD50 ≤ 60 mg/kg) or not. First, the publicly available QSAR model, CATMoS, was used to classify 630 Bayer Crop Science compounds. This model did not show good results (balanced accuracy of 0.52) as the compounds were in gaps of the model applicability domain. Interestingly, training a K nearest neighbor’s model (equivalent to read-across) using specifically the compound structure information, good classifications were obtained (balanced accuracy of 0.81). Then, this study explored whether morphological profiles obtained using Cell Painting in vitro assay on U2OS cells could assist in the prediction of rat acute oral toxicity and how those predictions could complement those made by QSAR models using chemical structures. Compounds with known acute oral toxicity were selected and a Cell Painting campaign were conducted on 226 compounds at 10 µM, 31.6 µM, and 100 µM. Binary classifiers based on K nearest neighbors, were developed to categorize compounds as highly acutely toxic (LD50 ≤ 60 mg/kg) or not. These classifiers were built, either using the compound chemical structure information (Morgan Fingerprint) or morphological profiles obtained from Cell Painting. Our results showed that the classification of compounds, using a read-across approach, as very acutely oral toxic or not, was possible using chemical structure information, U2OS cell morphological profiles or the combination of both. When classifying compounds structurally similar to those used to train the classifier, the chemical structure information was more predictive (mean balanced accuracy of 0.82). Conversely, when compounds to classify were structurally different from compounds used to train the classifier, the U2OS cell morphological profiles were more predictive (mean balanced accuracy of 0.72). The combination of both models allowed, when classifying compounds structurally similar to those used to train the classifiers, to slightly enhance the predictions (mean balanced accuracy of 0.85).