Abstract
The main objective of this PhD program is the development of innovative computational read-across methods for predicting engineered nanomaterial (ENM) properties (with emphasis to toxicity-related endpoints), based on experimental data. The read-across methods aim at determining neighbours (similar samples) to the query ENM in a dataset of ENMs with known properties and creating groups of related substances that have similar biological activity or toxic response.An important step in all the developed methodologies is the selection of the properties that are relevant to the endpoint of interest, to reduce the dimensionality of the models, avoid over-fitting and generate interpretable models. The automation of all the modelling parameters, is a key goal in this research project, and the proposed methodologies require the minimum information from the users to produce valid and robust read-across models.Special emphasis was given in the making of the models developed in this program availa ...
The main objective of this PhD program is the development of innovative computational read-across methods for predicting engineered nanomaterial (ENM) properties (with emphasis to toxicity-related endpoints), based on experimental data. The read-across methods aim at determining neighbours (similar samples) to the query ENM in a dataset of ENMs with known properties and creating groups of related substances that have similar biological activity or toxic response.An important step in all the developed methodologies is the selection of the properties that are relevant to the endpoint of interest, to reduce the dimensionality of the models, avoid over-fitting and generate interpretable models. The automation of all the modelling parameters, is a key goal in this research project, and the proposed methodologies require the minimum information from the users to produce valid and robust read-across models.Special emphasis was given in the making of the models developed in this program available through repositories or via user-friendly web applications. Implementation of the models as web tools supports their dissemination and actual use by all stakeholders in real-life applications.To begin with, a novel read-across methodology, related to the prediction of ENMs toxicity was developed. The method selects the most important variables and defines the neighbouring area around the target ENM, using single or multiple similarity criteria. The similarity criteria depend on the available ENM properties (e.g., physicochemical, biological, biokinetics etc.). The read-across prediction is computed as the weighted average of the neighbour ENMs. This novel grouping approach is based on the formulation and the solution of a mixed integer non-linear mathematical programming problem. A specific genetic algorithm scheme was developed to compute an approximate solution, due to the complexity of the problem rendering it practically unsolvable by conventional mathematical algorithms.The second method constructs, a mixed integer-linear optimisation program, which automatically filters out the noisy variables, defines the grouping boundaries based on one of the available properties -which is automatically chosen- and develops specific to each group LASSO linear regression predictive models. The third computational workflow is based on the formulation of a mathematical optimisation methodology that groups the ENMs into regions -according to their endpoint value-, removes the noisy variables, and incorporates the LASSO method for training predictive linear models specific to each region. Finally, k-Nearest Neighbours machine learning methodology was applied for deriving read-across models predicting the cytotoxicity and the biological activity of decorated multiwalled carbon nanotubes using calculated molecular descriptors of their surface ligands, and the zeta-potential of ENMs using geometrical ENMs properties extracted form transmission electron microscopy images.All developed methodologies were applied and validated on benchmark datasets, based on OECD principles, and were compared with methodologies already presented in Literature. They proved to be comparable and, in several cases, outperformed other alternative predictive modelling techniques, illustrating this way their good predictive performance and capabilities. Taking also into account that the grouping, feature selection and model generation steps are fully automated, the proposed methods can be considered as promising new approaches in the field of grouping/read-across modelling.
show more
