Machine Learning for Medicine: Accelerating Metadynamics of Supramolecular Host-Guest Complexes for Therapy and Imaging (MiAMI)

Project Description

In the context of 3-dimensional porous supramolecular materials (metallacages), computational methods hold promise to accelerate the challenging process of designing and optimizing their host-guest chemistry and encapsulation properties. This project establishes a comprehensive computational workflow to investigate the structural dynamics of lantern-shaped cationic [Pd₂L₄]⁴⁺ metallacages in explicit solvents with relevance to guest encapsulation and drug delivery. By integrating quantum-chemical calculations, classical force-field parametrization, and Machine Learning Interatomic Potentials (MLIPs), the study benchmarks the conformational behavior of two cage variants—endo-C and endo-N—across different charge models and environments. The workflow evaluates how partial charge assignment, solvent effects, and modeling methodology influence cage flexibility, stability, and cavity integrity. Comparative molecular dynamics simulations using both force fields and NequIP-based MLIPs reveal solvent-dependent conformational switching and highlight the advantages and trade-offs between classical and machine-learning-assisted approaches. Overall, the project provides a transferable platform for accurately modeling supramolecular assemblies in solution and lays the groundwork for predictive studies on host–guest chemistry in biomedical applications.

Results

• Established a two-branch workflow combining ab initio calculations, classical molecular dynamics (MD), and MLIP-based MD to model [Pd₂L₄]⁴⁺ cages in explicit solvents.
• Benchmarked four partial charge models (natural bond orbital analysis, Mulliken, electrostatic potential, restrained electrostatic potential) and demonstrated their strong impact on conformational behavior and energetic accuracy.
• Identified solvent-dependent dynamics: water promotes cage flexibility and opening, while DMSO stabilizes closed conformations through dipole–dipole interactions.
• Demonstrated that third-generation force fields capture solvent effects more reliably than earlier parameter sets.
• Trained and validated two NequIP MLIP versions, with v2 almost achieving density functional theory (DFT) accuracy and smoother conformational transitions than classical MD.
• Revealed that classical force fields overestimate energies of open conformations, whereas MLIPs reproduce DFT trends with higher fidelity.
• Provided a versatile methodology for future host–guest simulations relevant to drug delivery design.

Follow-up

This workflow enables systematic extension toward predictive modeling of guest binding thermodynamics, drug‑cage residence times, and solvent‑dependent loading efficiencies. Future efforts may focus on simulating clinically relevant small molecules, refining MLIP training sets for broader chemical coverage, and integrating enhanced sampling methods to capture rare binding and release events. Such developments could support rational design of metal–organic cages with optimized specificity, stability, and delivery performance in complex biological environments.