Understanding environmental processes, risks and toxicities of persistent organic pollutants (POPs) are needed to protect human and ecosystem health. Usually the toxic effects of POPs on human health are assessed using a variety of time- and cost-intensive in vivo and in vitro experiments; in vivo evaluations utilizing animals and further complicated by ethical concerns. Computational models provide an alternative way to laboratory based experiments. Indeed, computational models have recently become widely used to study reaction mechanisms, make predictions of chemical toxicity, and for risk assessment. In this dissertation, I study two computational methods that could potentially be used to advance remediation of dioxin and assess chemical toxicity: 1) molecular dynamics simulation of dioxin adsorption in activated carbon pores and 2) toxicity prediction with deep learning models, with a special focus on geometric scattering methods. Polychlorinated dibenzo-p-dioxins/furans (PCDD/Fs) are ubiquitous environmental contaminants that resist chemical, biological and physical routes of dissipation. They are well known for their toxicity, including adverse effects on reproductive health, impairment of mammalian immunity, and carcinogenicity. Adding activated carbons (ACs) to soils or sediments has been suggested as a means to promote the sequestration of polychlorinated dibenzo-p-dioxins/furans (PCDD/Fs) in forms with reduced bioavailability and hence toxicity to humans and other mammals. However, the mechanisms and adsorption processes of dioxin by ACs are not well understood. Thus, molecular dynamics simulations were used to study the mechanism of dioxin adsorption in AC pores, and to evaluate the effects of pore size on dioxin adsorption. The results showed that smaller pores created a comparatively more hydrophobic sub-aqueous environment that promoted the adsorption of dioxins. Deep learning has achieved great successes in image recognition, natural language processing and many other tasks. Recently, the application of deep learning methods for toxicity predictions of organic molecules has gained increasing interest. Molecules can be treated as graphs, where atoms are nodes and bonds are edges. However regular deep learning methods cannot be directly applied to data in a non-Euclidean space such as graphs. Therefore, geometric scattering methods that generalize regular scattering transforms to non-Euclidean spaces are proposed herein. Scattering transforms was used to provide mathematical understanding of convolutional networks. The results in this dissertation showed that geometric scattering methods achieved near state-of-art results on multiple standard graph classification tasks, and can be used for various explorations of biochemical data. Finally, geometric scattering was applied for toxicity prediction with real-world toxicity datasets. The results demonstrate that it has excellent potential as an alternative approach for toxicity predictions.
Read
