(209b) Structure-Activity Relationship Predicted Toxicity Metric for Metal Oxide Nanoparticles

Zhang, H., University of California, Los Angeles
Ji, Z., University of California, Los Angeles
Cohen, Y., University of California, Los Angeles
Nel, A. E., UCLA

The rapidly growing use of Engineered Nano-Materials (ENMs) in modern industrial products and processes has increased public concern regarding their potential of adverse environmental and health impacts. As a result, toxicity screening, which is critical for characterization of the potential hazard of ENMs for subsequent risk assessment and development of environmental and health regulatory policies, is confronted by a formidable task of (in vitro and in vivo) testing the large number and diversity of current and expected future ENMs. In order to tackle this challenge, there is a need for in silico methods (computational approaches) that enable virtual but rapid toxicity assessment. Structure-Activity Relationships (SARs) is such an approach that intends to generate predicted toxicity metric for studied materials from their physicochemical (structural-dependent) properties. In the present work, classification SAR were developed using a biological profile that depicts the cytotoxicity of twenty-four metal oxide nanoparticles. The cytotoxicity of these nanoparticles was assessed using human bronchial epithelial (BEAS-2B) as well as murine myeloid (RAW 264.7) cell lines with both single parameter assays (including MTS, ATP and LDH) and multi-parameter assays (comprising ROS production (MitoSox red fluorescence), intracellular calcium flux (Fluo4 fluorescence), mitochondrial membrane potential (JC1 fluorescence), and surface membrane permeability (PI uptake)) over a range of concentrations of 0.375-200 mg∙L-1 and exposure times of up to 24 h. The cytotoxicity profile was statistically processed via dose-response analysis and rank analysis, which labeled seven nanoparticles as toxic and seventeen as nontoxic. The labeling result was then used as an endpoint for developing SARs with an initial pool of thirty nanoparticle descriptors. In order to identify the most suitable SAR models, five different models (including Naïve Bayesian Classifier (NBC), Linear Regression (LIR), Linear Discriminate Analysis (LDA), Logistic Regression (LGR), and Support Vector Machine (SVM)) were evaluated. For each model, the most suitable nanoparticle descriptors were identified via an exhaustive feature (descriptor) selection. In order to ensure the simplicity and generalization capability of SARs, simple models (including NBC, LIR, LDA, and LGR) were restricted to having up to three descriptors, while having no more than two descriptors were enforced for more complex models (i.e., second order LGR and SVM). The models were assessed and compared using a validation approach which provided a weighted average of a re-substitution classification accuracy and a 200-round bootstrapping/re-sampling classification accuracy to estimate model performance. The second order LGR and SVM were identified as the best performing models, both demonstrating a good classification accuracy above 90% as well as Y-randomization validated robustness. More importantly, both SAR models selected conduction band energy and hydration energy as the most suitable descriptors, which agreed well with recent studies suggesting that the electron transfer from biological system to nanoparticles and the metal ion shedding from nanoparticle surface are two important mechanisms of cellular oxidative stress generation. Relying on the posterior class probabilities (i.e., the likelihoods of a nanoparticle being toxic or nontoxic) inferred in the model development, the SARs can assist in environmental and health regulatory decision making for nanoparticles under different acceptance levels of false negative and false positive. In addition, the developed SARs can also be used to guide the design of safer nanomaterials as well as prioritizing nanomaterials for confirmatory studies via expanded toxicity testing.