Due to the great biodiversity of the Brazilian flora, a wide variety of species, plant parts and plant products have been consumed as food. They are either wild or cultivated native and exotic plants of which uses can be restricted to certain regions and cultural groups. Many of the species consist of non-conventional food plants or even medicinal remedies, which lacks information regarding their safety. This work was aimed to create models for prediction of the toxicity of natural compounds which can be present in non-conventional food plants. QSTR (Quantitative Structure-Toxicity Relationships) modelling using 139,395 compounds obtained from three different lists of toxics (US EPA DSSTox) and non-toxics (FEMA GRAS™ and FDA GRAS). The structures were drawn, standardized and checked using JChem Suite [1]. Bit fingerprints as defined by Klekota and Roth [2] were calculated in PaDEL software [3] for each compound leading to 4860 bits which were used as independent variables. Data processing with the Caret’s package from R [4] was used to eliminate highly correlated variables and exclude those with near-zero variance, which reduced the number of bits to 166. The dependent variables consisted of the response classes 0 (non-toxic) / 1 (toxic). Classification models were created with both decision trees using J48 algorithm and random trees in the software Weka [5]. Samples were divided in training and test sets, which were composed of 92,001 (66%) and 47,394 (34%) structures, respectively. Models were ranked based on their prediction performance. The best model was obtained with the classifier random tree (scheme: weka.classifiers.trees.RandomTree -K 0 -M 1.0 -V 0.001 -S 1) which yielded true positive rate = 0.966, precision = 0.968, recall = 0.966, and ROC area = 0.849 for the test set prediction. The resulting QSTR model can be used to predict the toxicity of a wide range of compounds and natural products. The knowledge about the toxicity of compounds present in non-conventional food plants should contribute for their safe uses in the future.

Acknowledgements:
This study was supported by FAPESP grant #2014/26866-7, PIBIC-COPES/UFS, and FAPITEC/SE/FUNTEC 03/2017.

References:
1- JChem Suite 16.2.29, 2016, ChemAxon (http://www.chemaxon.com).
2- Klekota J, Roth FP. Chemical substructures that enrich for biological activity. Bioinformatics, 2008;24(21):2518-25.
3- Yap, CW. PaDEL-descriptor: An open source software to calculate molecular descriptors and fingerprints. Journal of computational chemistry, 2011;32(7):1466-74.
4- Kuhn, M. Building preditive models in R using the caret package. Journal of statistical software, 2008; 28(5):1-26.
5- Eibe F, Mark AH, Ian HW. The WEKA Workbench. Online Appendix for Data Mining: Practical Machine Learning Tools and Techniques, Morgan Kaufmann, 4 Ed., 2016; 128p.