Related Experiment Video
Updated: Oct 28, 2026
Synthesizing Amino Acids Modified with Reactive Carbonyls in Silico to Assess Structural Effects Using Molecular Dynamics Simulations
Published on: April 26, 2024
RapidMetalFF-AI: An AI-Assisted Web Server for Minute-Scale Bonded Force-Field Parameterization and Structural
1Department of Chemistry, Faculty of Science and Literature, Muş Alparslan University, 49250 Muş, Türkiye.
Abstract:
Automated bonded-force-field generation for metal-containing molecular systems remains a practical bottleneck because it is expertise-intensive and computationally demanding. RapidMetalFF-AI converts this fragmented process into a minute-scale automated web pipeline integrating an Advanced Numbering System (ANS), machine learning-based geometry optimization and Hessian generation, rapid charge assignment, bonded-parameter generation, minimization, and structural validation. The platform produces bonded parameters, assembled AMBER topology/coordinates, converted GROMACS/CHARMM representations, structural-retention metrics, and ANS-SMILES identifiers. In a deliberately heterogeneous stress test of 90 unedited CSD-derived MOL2 structures, 82 completed the full workflow and 72 contained recognized metal-site subsets. Median native/Hybrid optimized-reference global RMSDs were 0.430/0.415 Å, and metal-site medians were 0.079/0.107 Å. Relative to experimental CSD geometries, final native RapidMetalFF global/metal-site medians were 0.626/0.176 Å; corresponding Hybrid values were 0.623/0.166 Å. Nine representative systems with completed independent HF/SDD optimizations gave median native heavy-atom and metal-site RMSDs of 0.333 and 0.101 Å, a median metal-ligand bond-length MAE of 0.074 Å, and a median coordination-angle MAE of 2.865°. Ten explicit-water, periodic boundary MD tests (5 ns production after heating and equilibration; 50 ns total production) gave median time-averaged global and metal-site RMSDs of 0.571 and 0.112 Å relative to their production-start structures. These additional tests assess local geometry and nanosecond-timescale dynamical stability; they do not establish free-energy, spectroscopic, or general production-MD accuracy. For the 82 completed benchmark structures, the mean and median end-to-end runtimes were 65.5 and 48.9 s per molecule on an NVIDIA GeForce RTX 4060 Laptop GPU (NVIDIA Corporation, Santa Clara, CA, USA).
