New QSAR models to predict chromosome damaging potential based on the in vivo micronucleus test

• Published in: Toxicology Letters Vol. 329; pp. 80 - 84
• Main Authors: Van Bossuyt, Melissa, Raitano, Giuseppa, Honma, Masamitsu, Van Hoeck, Els, Vanhaecke, Tamara, Rogiers, Vera, Mertens, Birgit, Benfenati, Emilio
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.09.2020; Elsevier BV
• Subjects: Animals; Chromosome damage; Chromosomes; Computer Simulation; Databases, Nucleic Acid; Genotoxicity; In silico model; In vivo micronucleus; Micronucleus Tests; Models, Biological; QSAR; Quantitative Structure-Activity Relationship; Sensitivity and Specificity; Software; Genotoxicity; SMILES; EURL ECVAM; MCC; SA; EFSA; (Q)SAR; REACH; OECD; QSAR; ECHA; Chromosome damage; In vivo micronucleus; In silico model
• ISSN: 0378-4274; 1879-3169; 1879-3169
• DOI: 10.1016/j.toxlet.2020.04.016

•Collection of in vivo micronucleus test results of 718 structurally diverse compounds.•Construction of multiple new QSAR models using SARpy model building software.•Overall better performance than existing models, while demonstrating high coverage. A large number of computer-based prediction methods to determine the potential of chemicals to induce mutations at the gene level has been developed over the last decades. Conversely, only few such methods are currently available to predict potential structural and numerical chromosome aberrations. Even fewer of these are based on the preferred testing method for this endpoint, i.e. the micronucleus test. For the present work, in vivo micronucleus test results of 718 structurally diverse compounds were collected and applied for the construction of new models by means of the freely available SARpy in silico model building software. Multiple QSAR models were created using parameter variation and manual verification of (non-) alerting structures. To this extent, the original set of 718 compounds was split into a training (80 %) and a test (20 %) set. SARpy was applied on the training set to automatically extract sets of rules by generating and selecting substructures based on their prediction performance whereas the test set was used to evaluate model performance. Five different splits were made randomly, each of which had a similar balance between positive and negative substances compared to the full dataset. All generated models were characterised by an overall better performance than existing free and commercial models for the same endpoint, while demonstrating high coverage.