INVESTIGATING BIOLOGICAL SYSTEMS: MOLECULAR DYNAMICS SIMULATION AND MACHINE LEARNING APPROACHES TO PEPTIDE-LIPID, IONIZABLE LIPID-RNA INTERACTIONS, AND IRON-SULFUR PROTEINS

Abstract

In my dissertation, I present an investigation of molecular interactions and iron-sulfur protein through four interconnected research works. In Chapter 2, I characterized the thermodynamics of arginine-phosphate binding, focusing on D-myo-inositol-1,4,5-triphosphate (IP3) and the arginine-glycine-arginine tripeptide (RGR). Additionally, I investigated guanidinium-phosphate interactions in both monoester (such as glycerol-3-phosphate and glucose-6-phosphate) and diester (dimethyl-phosphate (DMP)) configurations using methyl guanidine (MGUA) as a model compound. Through comparison with experimental isothermal titration calorimetry data, I validated the computational binding energies with mono-ester phosphate and refined a specific nitrogen-oxygen interaction parameter in the CHARMM force field for MGUA-DMP interactions to improve the accuracy of molecular simulations. In Chapter 3, I investigated supramolecular interactions between novel ionizable lipids (AMG1041 and AMG1541) and RNA. Through molecular dynamics simulations, I demonstrated that β-hydroxyls significantly enhance RNA binding, with AMG1541 forming more hydrogen bonds than AMG1041 (11.2 vs 9.8) and stronger interaction energies (−136.04 ± 10.70 vs −113.68 ± 24.81 kcal/mol). Removal of β-hydroxyls nearly halved the hydrogen bonds (to 5.17 and 4.4) and substantially reduced interaction energies, providing insights for optimizing mRNA delivery systems. In Chapter 4, I developed a novel machine learning framework for identifying iron-sulfur (Fe-S) cluster-containing proteins. My sequence-based approach, utilizing my curated FeSseqdb database and random forest classification, achieved high precision (0.94 ± 0.01) and F1-score (0.88 ± 0.02). Through this framework, I uncovered key determinants of Fe-S protein identification, particularly the role of cysteine spacing and proline content. Building on this work, Chapter 5 focused on predicting Fe-S cluster redox potentials. I created a regression model that achieved strong correlation with experimental data (R² = 0.82) using only two features: total cluster charge and average Fe atom valence. Applying this model across 9,821 Fe-S clusters from the Protein Data Bank, I achieved 88% agreement with experimental values. Through these four chapters, I have advanced our understanding of fundamental molecular interactions and demonstrated the power of integrating experimental and computational approaches. My refined force field parameters, machine learning frameworks, and interaction analyses contribute valuable tools to the fields of molecular modeling, protein engineering, and drug delivery system design.