ACHIEVING MOLECULAR MODELING PRECISION AND AUTONOMY WITH MACHINE LEARNING
Saved in:
| Title: | ACHIEVING MOLECULAR MODELING PRECISION AND AUTONOMY WITH MACHINE LEARNING |
|---|---|
| Authors: | Keng, Mithony |
| Committee Members: | Merz, Kenneth Jr. KM; Hunt, Katharine KH; Wilson, Angela AW; Lee, Kin Sing KSL |
| Summary: | The accessibility of theoretical illustrations for chemical systems by computationally resolving their 3D structures provides quantitative and qualitative insight into structure-physiochemical property relationships that would otherwise go undetermined. From an experimentally derived observable like an ion-mobility mass spectrometry collision cross section, which by itself only yields information about size and approximate shape, an analyte precise gas phase conformation (i.e., equilibrium geometry) and charge state (i.e., protonation or deprotonation site) can be accurately assigned to within an experimental uncertainty (≤ 3% error). Generally, molecular modeling applications that are currently in circulation can capture relevant chemical empirical qualities; however, their performances can decay significantly with increasing molecular size and chemical group diversity. This may be due to such things as overfitting of training data, insufficient geometry optimization methods, or a desertion from empirical compliance during the software development process. In addition, there are no modeling packages that offer precision conformation sampling and/or that support preliminary results pertaining to conformer experimental viability. Thus, the absent of insight and preemptive control over a structure prediction job renders achieving optimal results fortuitous. Presented in this research literature are the validations of a standard workflow, which includes charge state modeling, conformation sampling, and a consistent DFT level of theory, that effectively handles a diverse class of molecules. The structures resolved using this workflow have been extensively challenged by experimental IM-MS reference to ensure validity. In subsequent works, the results from this standard workflow were then repurposed into training data for the development of two novel machine learning applications, CCSF (Collision Cross Section Focusing) and SEER (State Ensemble Energy Recognition), that enables precision sampling and accurate charge state prediction, respectively. Lastly, both CCSF and SEER were integrated as modules into a consolidated multi-model inference application called PEAS (Precise Ensemble Autonomous Sampling) that enables substantial molecular modeling throughput and autonomy through an exceptionally user-friendly platform. |
| URL: | https://doi.org/doi:10.25335/bv5c-wq97 |
| Database: | OpenDissertations |
| Abstract: | The accessibility of theoretical illustrations for chemical systems by computationally resolving their 3D structures provides quantitative and qualitative insight into structure-physiochemical property relationships that would otherwise go undetermined. From an experimentally derived observable like an ion-mobility mass spectrometry collision cross section, which by itself only yields information about size and approximate shape, an analyte precise gas phase conformation (i.e., equilibrium geometry) and charge state (i.e., protonation or deprotonation site) can be accurately assigned to within an experimental uncertainty (≤ 3% error). Generally, molecular modeling applications that are currently in circulation can capture relevant chemical empirical qualities; however, their performances can decay significantly with increasing molecular size and chemical group diversity. This may be due to such things as overfitting of training data, insufficient geometry optimization methods, or a desertion from empirical compliance during the software development process. In addition, there are no modeling packages that offer precision conformation sampling and/or that support preliminary results pertaining to conformer experimental viability. Thus, the absent of insight and preemptive control over a structure prediction job renders achieving optimal results fortuitous. Presented in this research literature are the validations of a standard workflow, which includes charge state modeling, conformation sampling, and a consistent DFT level of theory, that effectively handles a diverse class of molecules. The structures resolved using this workflow have been extensively challenged by experimental IM-MS reference to ensure validity. In subsequent works, the results from this standard workflow were then repurposed into training data for the development of two novel machine learning applications, CCSF (Collision Cross Section Focusing) and SEER (State Ensemble Energy Recognition), that enables precision sampling and accurate charge state prediction, respectively. Lastly, both CCSF and SEER were integrated as modules into a consolidated multi-model inference application called PEAS (Precise Ensemble Autonomous Sampling) that enables substantial molecular modeling throughput and autonomy through an exceptionally user-friendly platform. |
|---|
