Improved Molecular Dynamics of Alanine Dipeptide, Alanine Tetrapeptide, Alanine Hexapeptide, and ELIL Tetrapeptide in Implicit Solvent

This repository contains the datasets used in the publications "Learning Boltzmann Generators via Constrained Mass Transport" (CMT, presented at ICLR 2026, https://arxiv.org/abs/2510.18460) and "Efficient Training of Boltzmann Generators Using Off-Policy Log-Dispersion Regularization" (LDR, presented at ICML 2026, https://arxiv.org/abs/2602.03729). We provide training, validation, and test datasets subsampled from molecular dynamics trajectories performed in OpenMM with Amber force fields and implicit solvation. We cover the molecular systems alanine dipeptide, alanine tetrapeptide, alanine hexapeptide, and ELIL tetrapeptide, with details on force fields and data generation described in our publications (see above). Coordinates are given in Cartesian coordinates (Å). We only report data at 300K. Note: We observed small deviations from equilibrium in the energy histograms, so we slightly adjusted the MD procedure compared to previous publications (FAB, TA-BG) to generate conformations closer to the equilibrium distribution. Overview of the folder structure and files contained in data.zip: data/aldp – Files for the alanine dipeptide system (Amber ff96 force field) 300K_test.npy – Test dataset at 300K, $10^7$ samples 300K_train.npy – Training dataset at 300K, $10^6$ samples 300K_val.npy – Validation dataset at 300K, $10^6$ samples data/tetra – Files for the alanine tetrapeptide system (Amber ff99SB-ILDN force field) 300K_test.npy – Test dataset at 300K, $10^7$ samples 300K_train.npy – Training dataset at 300K, $10^6$ samples 300K_val.npy – Validation dataset at 300K, $10^6$ samples data/hexa – Files for the alanine hexapeptide system (Amber ff99SB-ILDN force field) 300K_test.npy – Test dataset at 300K, $10^7$ samples 300K_train.npy – Training dataset at 300K, $10^6$ samples 300K_val.npy – Validation dataset at 300K, $10^6$ samples data/ELIL – Files for the ELIL tetrapeptide system (Amber ff99SB-ILDN force field) 300K_test.npy – Test dataset at 300K, $10^7$ samples 300K_train.npy – Training dataset at 300K, $10^6$ samples 300K_val.npy – Validation dataset at 300K, $10^6$ samples data_ldr.zip contains additional data used only in the LDR publication: data_ldr/aldp – Additional files for the alanine dipeptide system (Amber ff96 force field) 300K_train_large.npy – Training dataset at 300K, $9 \cdot 10^6$ samples (the first $10^6$ samples correspond to 300K_train.npy in data.zip) 300K_biased_train.npy – Biased training dataset at 300K, $10^6$ samples 300K_biased_val.npy – Biased validation dataset at 300K, $10^6$ samples data_ldr/hexa – Additional files for the alanine hexapeptide system (Amber ff99SB-ILDN force field) 300K_train_large.npy – Training dataset at 300K, $9 \cdot 10^6$ samples (the first $10^6$ samples correspond to 300K_train.npy in data.zip) data_ldr/aldp_cart – Files for the alanine dipeptide system used for training in Cartesian coordinates (Amber ff99SB-ILDN force field) 300K_test.npy – Test dataset at 300K, $10^7$ samples 300K_train_large.npy – Training dataset at 300K, $9 \cdot 10^6$ samples 300K_val.npy – Validation dataset at 300K, $10^6$ samples