Toxicity Classification of Oxide Nanomaterials: Effects of Data Gap Filling and PChem Score-based Screening Approaches

Abstract

Development of nanotoxicity prediction models is becoming increasingly important in the risk assessment of engineered nanomaterials. However, it has significant obstacles caused by the wide heterogeneities of published literature in terms of data completeness and quality. Here, we performed a meta-analysis of 216 published articles on oxide nanoparticles using 14 attributes of physicochemical, toxicological and quantum-mechanical properties. Particularly, to improve completeness and quality of the extracted dataset, we adapted two preprocessing approaches: data gap-filling and physicochemical property based scoring. Performances of nano-SAR classification models revealed that the dataset with the highest score value resulted in the best predictivity with compromise in its applicability domain. The combination of physicochemical and toxicological attributes was proved to be more relevant to toxicity classification than quantum-mechanical attributes. Overall, by adapting these two preprocessing methods, we demonstrated that meta-analysis of nanotoxicity literatures could provide an effective alternative for the risk assessment of engineered nanomaterials.

Practices and Trends of Machine Learning Application in Nanotoxicology

“…For example, Ban et al [80] used curve-fitting to calculate missing ages based on the age-weight relationships of different species. While assessing data quality and completeness, nano-specific filling in of missing values using manufacturer's specifications and/or estimations [60,64] was suggested within the Safe and Sustainable Nanotechnology (S2NANO) (http://portal.s2nano.org/ (Webpage accessed autumn 2019)) database. Furxhi et al [72] investigated the robustness of several ML tools on generated versions of the dataset by removing values artificially.…”

“…They found that the model captured nonlinear dependence between descriptors and cytotoxicity as well as possible interactions. RF is an ML recursive ensemble algorithm based on a combination of independently grown binary decision trees constructed with various samples of a bootstrap [64]. By aggregating the predictions of each tree, the RF algorithm makes forecasts depending significantly on two model parameters.…”

Application of Bayesian networks in determining nanoparticle-induced cellular outcomes using transcriptomics

“…material safety data sheets) which resulted in a second dataset (PD). Density values derived from manufactures information combined with particle size distribution data (assuming spherical shape and smooth surface) can be used to calculate the Specific Surface Area (SSA) (Ha et al 2018). Furthermore, as well as data gap filling, discretization was also performed during preprocessing.…”

“…Most instances in the training dataset correspond to no effect outcomes for the majority of the endpoints. Imbalanced datasets can limit the performance of most classification algorithms, making the prediction biased to the dominant class value (Ha et al 2018). To avoid this, we adjusted the relative frequency of triggered/ no effect instances by resampling the second dataset by applying SMOTE (Synthetic Minority Oversampling Technique), a supervised instance algorithm that oversamples the minority instances using the k-nearest neighbors algorithm (Chawla et al 2002).…”

Toxicity modelling of nanomaterials by origin evaluation of their physicochemical descriptors using a combination of principal component analysis and support vector machine methods

“…The structure–activity relationship (Liu et al, ), perturbation (Kleandrova et al, ; Kleandrova et al, ; Luan et al, ; Speck‐Planche, Kleandrova, Luan, & Cordeiro, ), quasi‐SMILES‐ (Trinh et al, ), and theoretical descriptors‐based (Boukhvalov & Yoon, ) models have been established in the prediction of toxicity of nanoparticles. Some QSARs models relating physicochemical descriptors to cellular responses of nanomaterials based on the multivariate analysis (Le, Epa, Burden, & Winkler, ), including principal component (PC) analysis (PCA; Sayes & Ivanov, ; Lynch, Weiss, & Valsami‐Jonesa, ), hierarchical clustering (Shaw et al, ), linear discriminant analysis (Sayes & Ivanov, ), artificial neural network (Winkler et al, ), support vector machine (SVM; Fourches et al, ), naïve Bayes, k‐nearest neighbour (Chau & Yap, ), linear and nonlinear regression analysis (Can, ; Chau & Yap, ), PChem score‐based screening and data imputation approaches (Ha et al, ), and artificial neural network, random forest, SVM, and generalized linear models (Choi, Ha, Trinh, Yoon, & Byun, ), so forth were discussed in the previous studies. A short summary of some latest development in toxicity modelling of nanomaterials is presented in Table .…”