Performance of universal machine-learned potentials with explicit long-range interactions in biomolecular simulations: Datasets and models

Description of datasets

This repository provides train/validation/test splits of atomistic datasets used to train and evaluate ICTP (Irreducible Cartesian Tensor Potential) models for molecular and biomolecular simulations. Each structure includes:

Atomic positions in Å

Total energies in eV

Atomic forces in eV/Å

Total charges in units of e

These quantities were used directly for training and evaluating ICTP models.

These datasets are largely based on SPICE-v2 and are derived from first-principles reference calculations.

A detailed description of dataset curation, reference level of theory, and evaluation is provided in the accompanying paper:

Performance of universal machine-learned potentials with explicit long-range interactions in biomolecular simulations

Description of models

All models used in the experiments are ICTP models, including:

ICTP-LR(S), ICTP-LR(M), ICTP-LR(L) (with explicit long-range electrostatics and dispersion)

ICTP-SR(M) (short-range model)

Examples for training ICTP models with the curated datasets are provided in the official ICTP repository:

https://github.com/nec-research/ictp/blob/main/examples/run_training_SPICE.py

Examples for running molecular dynamics simulations with trained ICTP models (including input geometries) are available at:

https://github.com/nec-research/ictp/tree/main/examples/dimos

Please cite the preprint in any work that uses these datasets or ICTP models with explicit long-range electrostatics and dispersion if you find them useful.