An open source multistep model to predict mutagenicity from statistical analysis and relevant structural alerts

Abstract

BackgroundMutagenicity is the capability of a substance to cause genetic mutations. This property is of high public concern because it has a close relationship with carcinogenicity and potentially with reproductive toxicity. Experimentally, mutagenicity can be assessed by the Ames test on Salmonella with an estimated experimental reproducibility of 85%; this intrinsic limitation of the in vitro test, along with the need for faster and cheaper alternatives, opens the road to other types of assessment methods, such as in silico structure-activity prediction models.A widely used method checks for the presence of known structural alerts for mutagenicity. However the presence of such alerts alone is not a definitive method to prove the mutagenicity of a compound towards Salmonella, since other parts of the molecule can influence and potentially change the classification. Hence statistically based methods will be proposed, with the final objective to obtain a cascade of modeling steps with custom-made properties, such as the reduction of false negatives.ResultsA cascade model has been developed and validated on a large public set of molecular structures and their associated Salmonella mutagenicity outcome. The first step consists in the derivation of a statistical model and mutagenicity prediction, followed by further checks for specific structural alerts in the "safe" subset of the prediction outcome space. In terms of accuracy (i.e., overall correct predictions of both negative and positives), the obtained model approached the 85% reproducibility of the experimental mutagenicity Ames test.ConclusionsThe model and the documentation for regulatory purposes are freely available on the CAESAR website. The input is simply a file of molecular structures and the output is the classification result.

Mutagenic and carcinogenic structural alerts and their mechanisms of action

“…CAESAR meets all five principles of the Organisation for Economic Co-operation and Development (OECD). It has good predictive capabilities for mutagenicity and carcinogenicity but, unfortunately, it does not include prediction models for genotoxicity (59).…”

In Silico Prediction of Chemically Induced Mutagenicity: How to Use QSAR Models and Interpret Their Results

“…The algorithm of the model is described in Ferrari and Gini [ 23 ]. CAESAR-VEGA automatically calculates chemical descriptors for the chemicals of interests and contains a subset of Toxtree rules ( see previous paragraph) to enhance the sensitivity of the model.…”

“…The CAESAR model [ 23 ] was developed on the basis of 4204 chemicals (2348 mutagenic and 1856 non-mutagenic) extracted from the Bursi data set [ 24 ]. This initial set was then split into training set (3367 chemicals, 80 % of the entire data set) and external test set (837 chemicals, 20 % of the entire data set) [ 24 ].…”

Novel Uses of In Vitro Data to Develop Quantitative Biological Activity Relationship Models for in Vivo Carcinogenicity Prediction

“…[1,2] QSAR models have been continuously improving with new machine learning algorithms, molecular descriptors and training databases. [3][4][5] However,s everal studies show that they are still not very predictive for mechanistically complex endpoints like carcinogenicity. [6,7] These limitations are primarily due to the multiple mechanisms of action associated with more complex toxic endpoints.…”