# The interlocking process in molecular machines explained by a combined approach: the nudged elastic band method and machine learning potential

## TL;DR

This paper introduces a new method combining machine learning and simulations to study how molecular machines assemble and move.

## Contribution

A novel framework combining nudged elastic band and machine learning potential for analyzing molecular machine assembly.

## Key findings

- The threading process in rotaxane-like complexes involves hydrogen bond stabilization and macrocycle expansion.
- ANI-1ccx accurately models supramolecular assembly despite not being trained on charged systems.
- Calculated energy barriers align well with experimental data.

## Abstract

Engineering molecular machines requires a precise knowledge of the mechanisms involved in programmed motions. Among artificial molecular machines, rotaxanes have emerged as a prominent model due to their ability to perform diverse and controlled motions, such as threading, shuttling, and pirouetting. In this work, we present a reliable theoretical framework to describe the threading motion during the assembly of rotaxane-like complexes. Our approach combines the climbing image nudged elastic band method with the ANI-1ccx neural network potential, trained with gold-standard data. Energetic and structural variations along a normalized displacement coordinate allowed an accurate atomistic description of the threading process of different dumbbell-shaped molecules (axles) through the cavity of two different macrocyclic hosts (tori). Using the methodology proposed herein, two key steps are identified: stabilization through hydrogen bonds, which we call the claw mechanism, and the expansion of the macrocycle. An energy decomposition analysis, performed by single-point calculations on selected structures, allows an analysis of the role of steric and electrostatic effects in the structural stabilization of supramolecular assemblies. We find that, although ANI-1ccx was not explicitly trained for charged systems, this neural network potential effectively discriminates among various charged states. Furthermore, calculated potential energy barriers are in good agreement with reported experimental free energy barriers. The featured methodology has the potential to become a fundamental artificial-intelligence-based tool for the study of diverse motions observed in supramolecular systems.

Engineering molecular machines requires a precise knowledge of the mechanisms involved in programmed motions.