PhD Research Projects


During my PhD, I performed classical molecular dynamics (MD) simulations and developed machine learning algorithms to predict the effect of parameter selection on system properties for two industrially relevant areas, guiding the selection of gold nanoparticle parameters for biomedical applications and solvent composition for new biomass conversion processes. If interested, check out my thesis here: [Link]

Application 1: Screening Gold Nanoparticle Properties for Biomedical Applications

A Schematic showing the use of SAM-protected GNPs for selectively binding to molecules (e.g. proteins). Simulation snapshot shows a GNP with R = OH ligands in water. B GNP surface properties (e.g. hydrophilicity) emerging from collective ligand interactions could inform on GNP behavior with biomolecules. C Principal component analysis using molecular descriptors can predict experimental immune response data [Moyano, D. F. et al. J. Am. Chem. Soc. 2012, 134, 3965–3967]. 

Monolayer-protected gold nanoparticles (GNP) are promising synthetic materials applicable in many fields (e.g. drug delivery, pollutant detection, etc.) because of their ease of fabrication, detectability, and tunable parameters. GNPs are often coated by self-assembled monolayers (SAMs), a set of ligands that consist of a sulfur head group, carbon backbone, and terminal end groups. SAM-protected GNPs have a vast design space (e.g. gold core shape, size, and ligand selection), making it challenging to predict how these parameters may affect GNP interactions with other molecules, such as proteins. I hypothesize that MD simulations could help guide parameter selection by computing MD-descriptors that account for variations in GNP parameters could predict GNP interactions with biomolecules.

I developed a generalized system preparation workflow that could model a wide-range of SAM-protected GNPs. Then, I developed MD-descriptors that account for collective interactions (e.g. formation of bundles) and encode GNP surface properties (e.g. hydrophilicity). These MD-descriptors are advantageous because they account for ligand-ligand and ligand-water interactions that are missed in conventional descriptors, such as single-ligand octanol-water partition coefficients (logP). Using data-centric tools (e.g. principal component analysis), MD-descriptors could accurately predict GNP behavior in the body, such as immune response, which outperform prediction models using single-ligand descriptors (logP). I am also collaborating with experimentalists (Prof. Pedersen group, UW-Madison) to identify new mechanisms for GNP-lipid membrane interactions and developing MD-descriptors that could predict GNP behaviors more broadly. My computational models enable cost-efficient screening of GNP properties that could be tuned for selective biomolecular interactions, useful in a wide-range of applications.

Relevant publications:

Application 2: Solvent Screening for Biomass Conversion Reactions

A Schematic showing biomass converted into fuel via acid-catalyzed reactions. An example molecular dynamics snapshot is shown for xylitol (XYL) in 90 wt% DIO. B Molecular dynamics (MD) trajectories analyzed by 3D convolutional neural networks can predict the conversion of biomass-derived model compounds (ethyl tert-butyl ether (ETBE), tert-butanol (TBA), levoglucosan (LGA), 1,2-propandiol (PDO), cellobiose (CEL), fructose, and XYL).

Lignocellulosic biomass (organic material from plants) is a promising renewable, sustainable resource that can be converted into transportation fuels or commodity chemicals. Biomass conversion is performed through acid-catalyzed reactions in aqueous solution, but the process is hindered by low reactivity and poor selectivity. One way to improve reactivity is to modify the solvent composition by mixing water with organic cosolvents, which have been shown to improve the reactivity of biomass conversion reactions by 100-fold compared to the same reactions in pure water. However, solvent selection is challenging, so I hypothesized that computational tools can help guide solvent selection by gaining physical insight into the effects of solvent composition on reaction rates and selectivities.

I developed classical MD simulations that modeled biomass-derived reactants in mixtures of water with polar aprotic cosolvent in collaboration with experimentalists (Profs. Huber and Dumesic groups, UW-Madison). I found that inclusion of polar aprotic cosolvents leads to the formation of water-enriched local domains around hydrophilic reactants, which draws the acid catalyst to these regions due to preferred catalyst-water interactions and results in improved reaction rates. By quantifying the extent of water-enrichment around the reactant, I found that MD-descriptors can accurately predict experimental reaction rates for dioxane-water mixtures, showing that classical simulation techniques can inform reaction rates without modeling the reaction mechanism or the catalyst. I also used MD to gain physical insight into how solvents influence selectivities in biomass conversion reactions and combined these tools into a generalized workflow for screening solvents in biomass-related reactions.

Relevant publications: