Evaluating the Transferability and Distinguishability of Atomic Moments Across Atom in Molecule (AIM) Partitioning Schemes
TL;DR
By identifying AIM methods that result in conformationally stable and transferable atomic properties, this study aims to improve IDP modelling and guide the development of more accurate force fields for biologically relevant systems.
Abstract
Molecular dynamics (MD) simulations are widely used to study biomolecular systems, and their accuracy depends on the chosen force field. More accurate models enable better MD simulations, improve machine learning force fields, and have a stronger predictive power for real-world applications, including drug-protein binding affinity, property prediction, and chemical reactivity. Standard force fields work well for folded globular proteins, but they often fail to accurately describe intrinsically disordered proteins (IDPs), producing overly compact or artificially structured ensembles due to imbalanced non-bonded interactions, ultimately reducing the accuracy of simulation results and motivating the development of improved modelling approaches. This project evaluates atom in molecule (
Aim
partitioning schemes as candidates for improved electrostatic descriptions in force field development. Individual atom’s properties (charges, dipole moments, and quadrupole magnitudes) were computed for a backbone dataset of 400 overlaid samples of N-methylacetamide across multiple fuzzy methods (H, HI, MBIS, AVH-B, and AVH-M). The transferability and distinguishability of these measurements were quantified using geometric (e.g., convex-hull area) and statistical (e.g., covariance spectra) methods for 2D and 3D feature vectors containing combinations of the three atom properties used. The standard data was compared to Z-score and min-max normalized versions of the data as well. By identifying
Aim
methods that result in conformationally stable and transferable atomic properties, this study aims to improve IDP modelling and guide the development of more accurate force fields for biologically relevant systems. Future work will extend the analysis to include side chain, organic, and silica datasets, enabling comparison of
Aim
methods across organic and inorganic environments. Graph-theoretic measures will be incorporated to quantify cluster structure and apply distinguishability metrics to capture separation between atom types.Faculty Supervisor: Dr. Farnaz Heidar-Zadeh
