عنوان

Evaluating and Improving Computational Models for Physical Property Predictions

TL2bd297fz

انگلیسی

Evaluating and Improving Computational Models for Physical Property Predictions

[Thesis]

Bannan, Caitlin Colleen

Mobley, David L

2019

Mobley, David L

2019

Simulations allow us to predict free energies and physical properties of molecules in advance of their synthesis saving time and resources. My research focuses on how to automatically evaluate and improve these predictions. I begin by describing my work to test the accuracy of free energy calculations by computing various partition coefficients. Then, I report on applications of the resulting procedure in the SAMPL5 blind challenge for 53 small drug-like molecules. Comparing computed and experimental values highlighted three areas which still need improvement: conformational sampling, protonation assignment, and force field accuracy. These results motivated my next project, designing a Gaussian process model for pKa prediction based on computed properties of small molecules. I tested this model in the SAMPL6 blind challenge on pKa prediction where it performed competitively with many established methods. My partition coefficient results also highlighted the limitations of current force fields -- used to calculate potential energy of a system based on atomic coordinates. To address these concerns, I joined the the Open Force Field Initiative, a collaboration working to automate force field parametrization. The culmination of my Ph.D. focuses on generating chemical perception -- the way a force field assigns parameters to a molecule -- without the historically required human intuition. Improved force fields will result in more accurate predictive models and a better understanding of a wide variety of fields including computer-aided drug design, biomaterials, and polymer chemistry.

Nataraj, Shakthi

Bannan, Caitlin Colleen

UC Irvine

p

[Thesis]

276903

a

Y