In Silico Cytotoxicity Prediction of Nanoparticles by Artificial Neural Networks Gabriel Engcong (1, 2, a), Irelie Ebardo (1, 2, b), Phoebe Sombilon (1, c), Mikee Joy Rodriguez (1, d), John Riz Bagnol (3, e), Lilibeth Coronel (4, f), Mary Joy Latayada (5, g), Jay Michael Macalalag (5, h), Farlley Bondaug (1, i), Arnold Alguno (6, j), and Rey Capangpangan (1,4, k)
Author Affiliations
1) REY Laboratories, Research Division, Mindanao State University at Naawan, Naawan, 9023, Philippines
2) Department of Science and Technology - Science Education Institute, General Santos Avenue, Bicutan, 1631, Taguig City
3) The University of Southeastern Philippines, F. Inigo St., Barrio Obrero, Davao City, 8000, Philippines
4) College of Marine and Allied Sciences, Mindanao State University at Naawan, Naawan, 9023, Philippines
5) Caraga State University, Ampayon, Butuan City, 8600, Philippines
6) Department of Physics, Mindanao State University-Iligan Institute of Texhnology, Tibanga, 9200, Iligan City
Author Emails
a) gabriel.engcong[at]msunaawan.edu.ph
b) irelie.ebardo[at]msunaawan.edu.ph
c) phoebe.sombilon[at]msunaawan.edu.ph
d) mikeejoy.rodriguez[at]msunaawan.edu.ph
e) jrbagnol[at]usep.edu.ph
f) lilibeth.coronel[at]msunaawan.edu.ph
g) mrlatayada[at]carsu.edu.ph
h) jrmacalalag[at]carsu.edu.ph
i) farlley.bondaug[at]msunaawan.edu.ph
j) arnold.alguno[at]g.msuiit.edu.ph
k) rey.capangpangan[at]msunaawan.edu.ph
Abstract
Information on the toxicity behavior of nanomaterials (NMs) is of vital importance to ensure that its delivery to biological targets causes no adverse effects. Some engineered NMs present new and unusual hazards. However, there is very little information on how their toxicity can be identified, assessed, and even controlled. NMs toxicity evaluation is normally evaluated through in vitro toxicity or animal testing. However, these modes of toxicity evaluation, aside from being expensive, require long analysis time and are being questioned for ethical considerations. Furthermore, the reliability and reproducibility of animal studies within species are also questionable even if the experiment is being done under rigorous protocol. Hence, our research is driven towards the development of in silico predictive models in assessing the cytotoxicity of the different NMs. Artificial neural networks (ANNs) are a class of machine learning models that have been successfully applied in the field of chemical sciences to develop quantitative structure-activity relationship models for chemical activities of compounds. This paper applied ANN to build predictive models for nanoparticle (NP) cytotoxicity. The models are trained on a heterogeneous dataset on NP cytotoxicity obtained from a comprehensive literature review. Neural network models for binary classification and regression are both considered and are separately investigated. Shallow network architectures with up to 2,048 units per hidden layer were explored for each model. The selected models for the binary classification task showed prediction accuracies in the range of 80% - 93% while the models for the regression task showed mean absolute errors (MAE) in the range of 14% - 23%. Results show that excessively large datasets are not necessary to obtain
optimal predictive performance and that the choice of the selected features for the models greatly impact predictive performance on both the classification and regression tasks.
