MANBO: An ab initio metadynamics database of RNA-Au interface for machine learning potentials

To develop accurate machine-learning interatomic potentials (MLIPs), we present MANBO (Metal-adsorbed Au Nanocluster Binding Oligonucleotide), a comprehensive metadynamics-enhanced AIMD database for a sulfur-modified RNA dinucleotide (GpGp) on Au(111) in explicit water. Systematic exploration of RNA at the RNA-Au interface generated 180,000 candidate configurations at the PBE-D3 level. These were distilled into a compact dataset of 72,880 configurations using Smooth Overlap of Atomic Positions (SOAP) descriptors and farthest-point sampling. Based on this dataset, two state-of-the-art MLIP architectures, Message-passing Atomic Cluster Expansion (MACE) and DPA4, were benchmarked for energy and force prediction at chemical level. The resulting MLIPs remained stable throughout subsequent molecular dynamics simulations and reproduced the key features of the reference AIMD trajectories across a broad temperature range.