Datasets and ML/MM models for simulating uracil, N-methylacetamide and alanine dipeptide

The same types of files are provided for the three molecules studied: uracil, N-methylacetamide, and alanine dipeptide. Datasets These archives contain the training and test sets for each molecule and each level of theory. The first artificial dataset was generated using normal-mode displacements combined with random charges placed on external points around the molecule. The validation test sets were extracted from ML/MM simulations in solvent obtained with the trained models. The two reference levels of theory are ωB97XD/6-31G(d) and B2PLYP-D3/cc-pVTZ. For each molecule we provide datasets_*.tar.gz, which includes: Artificial training and test sets training-set_wB97XD_6-31G(d).npz - 1000 samples; training-set_B2PLYPD3_cc-pVTZ.npz - 400 samples; test-set_wB97XD_6-31G(d).npz - 200 samples; test-set_B2PLYPD3_cc-pVTZ.npz - 200 samples; Test sets extracted from ML/MM simulations test-set_dynamics_wB97XD_6-31G(d).npz - 500 samples; test-set_dynamics_B2PLYPD3_cc-pVTZ.npz - 250 samples. Each dataset contains QM and MM coordinates, MM charges, vacuum energies and forces, vacuum dipole moments, QM/MM energies and forces, and QM/MM dipole moments. Dipoles from dynamics These archives contain the dipole moments used for IR spectra, extracted from ML/MM simulations. Uracil - dipoles_from_dynamics_uracil.tar.gz: gas-phase and water ML/MM simulations of 1 ns using the base-models (Mbase_vacuum.npy, Mbase_water.npy); gas-phase and water ML/MM simulations of 1 ns using the Δ-learning correction only for the vacuum (Mdelta_vacuum.npy, Mdelta_water.npy); water ML/MM simulations of 1 ns using the Δ-learning correction for both vacuum and environment (Mdelta_deltaEnv_water.npy); short simulation of 10 ps of aqueous uracil using ML/MM and QM/MM (short_Mbase_MLMM.npy, short_QMMM.npy); N-methylacetamide - dipoles_from_dynamics_nmethylacetamide.tar.gz: gas-phase, chloroform, and water ML/MM simulations of 1 ns using the base-models (Mbase_vacuum.npy, Mbase_chloroform.npy, Mbase_water.npy); gas-phase, chloroform, and water ML/MM simulations of 1 ns using the Δ-learning correction only for the vacuum (Mdelta_vacuum.npy, Mdelta_chloroform.npy, Mdelta_water.npy); Alanine dipeptide - dipoles_from_dynamics_alanine_dipeptide.tar.gz: water and DMSO ML/MM simulations of 100 ps starting from αR and PII conformations. For αR we report only the replicas that remained in this conformation, whereas for PII we considered all the replicas that stayed in the PII or β conformation. These simulations were run with both the base-models (Mbase_water_*.npy, Mbase_DMSO_*.npy) and correcting with Δ-learning the vacuum part (Mdelta_water_*.npy, Mdelta_DMSO_*.npy). Models These archives contain all parameters for the machine learning models trained with GPX (permut_symm branch). For the base-models the reference level of theory is ωB97XD/6-31G(d) and for the Δ-learning ones is B2PLYP-D3/cc-pVTZ. models_uracil.tar.gz: vaccum and environment base-models (modelvacgs.npz, modelenvgs.npz), and vacuum and environment Δ-learning models (modelvacgsdelta.npz, modelenvgsdelta.npz); models_nmethylacetamide.tar.gz: vaccum and environment base-models (modelvacgs.npz, modelenvgs.npz), and vacuum Δ-learning model (modelvacgsdelta.npz); models_alanine_dipeptide.tar.gz: vaccum and environment base-models (modelvacgs.npz, modelenvgs.npz), and vacuum Δ-learning model (modelvacgsdelta.npz). These models can be directly used in ML-server to run ML/MM simulations. Scripts for dataset generation The archive scripts_for_datagen.tar.gzincludes two python scripts for generating geometries of isolated molecules and artificial environment configurations: nm_displacement.py: reads the high-precision normal modes computed by Gaussian and the equilibrium geometry to compute normal-mode displacements; gen_charges.py: reads the isolated molecule geometry from Gaussian input files and the Mulliken charges from Gaussian output files, calculates a layered grid of points around the molecule, and selects a subset on which random external charges are placed.为所研究的三种分子——尿嘧啶（uracil）、N-甲基乙酰胺（N-methylacetamide）以及丙氨酸二肽（alanine dipeptide）——均提供了相同类型的文件。 ## 数据集 这些压缩包包含了每种分子、每种理论级别的训练集与测试集。首个人工数据集通过简正模式位移结合置于分子周围外点的随机电荷生成。验证测试集源自经训练模型得到的溶剂中ML/MM（机器学习分子力学，ML/MM）模拟结果。两类参考理论级别分别为ωB97XD/6-31G(d)与B2PLYP-D3/cc-pVTZ。针对每种分子，我们提供了`datasets_*.tar.gz`，其包含以下内容： ### 人工训练集与测试集 - `training-set_wB97XD_6-31G(d).npz`：1000个样本； - `training-set_B2PLYPD3_cc-pVTZ.npz`：400个样本； - `test-set_wB97XD_6-31G(d).npz`：200个样本； - `test-set_B2PLYPD3_cc-pVTZ.npz`：200个样本。 ### 从ML/MM模拟中提取的测试集 - `test-set_dynamics_wB97XD_6-31G(d).npz`：500个样本； - `test-set_dynamics_B2PLYPD3_cc-pVTZ.npz`：250个样本。 每个数据集均包含QM/MM（量子力学/分子力学，QM/MM）坐标、MM（分子力学，MM）电荷、真空能量与力、真空偶极矩、QM/MM能量与力以及QM/MM偶极矩。 ## 动力学偶极矩 此类压缩包包含用于红外光谱的偶极矩数据，均源自ML/MM模拟结果。 ### 尿嘧啶：`dipoles_from_dynamics_uracil.tar.gz` - 采用基础模型（`Mbase_vacuum.npy`、`Mbase_water.npy`）完成的1 ns气相与水相ML/MM模拟数据； - 仅针对真空部分采用Δ学习校正的1 ns气相与水相ML/MM模拟数据（`Mdelta_vacuum.npy`、`Mdelta_water.npy`）； - 针对真空与环境均采用Δ学习校正的1 ns水相ML/MM模拟数据（`Mdelta_deltaEnv_water.npy`）； - 采用ML/MM与QM/MM方法完成的10 ps水溶液尿嘧啶短时间模拟数据（`short_Mbase_MLMM.npy`、`short_QMMM.npy`）。 ### N-甲基乙酰胺：`dipoles_from_dynamics_nmethylacetamide.tar.gz` - 采用基础模型（`Mbase_vacuum.npy`、`Mbase_chloroform.npy`、`Mbase_water.npy`）完成的1 ns气相、氯仿相与水相ML/MM模拟数据； - 仅针对真空部分采用Δ学习校正的1 ns气相、氯仿相与水相ML/MM模拟数据（`Mdelta_vacuum.npy`、`Mdelta_chloroform.npy`、`Mdelta_water.npy`）。 ### 丙氨酸二肽：`dipoles_from_dynamics_alanine_dipeptide.tar.gz` - 从αR与PII构象起始的100 ps水相与二甲基亚砜（DMSO）相ML/MM模拟数据。其中针对αR构象，仅统计保持该构象的复现结果；针对PII构象，纳入所有保持PII或β构象的复现结果。此类模拟分别采用基础模型（`Mbase_water_*.npy`、`Mbase_DMSO_*.npy`）与针对真空部分进行Δ学习校正的模型（`Mdelta_water_*.npy`、`Mdelta_DMSO_*.npy`）完成。 ## 模型 此类压缩包包含采用GPX（排列对称分支，permut_symm branch）训练得到的所有机器学习模型参数。 其中基础模型的参考理论级别为ωB97XD/6-31G(d)，Δ学习模型的参考理论级别为B2PLYP-D3/cc-pVTZ。 - `models_uracil.tar.gz`：气相与环境基础模型（`modelvacgs.npz`、`modelenvgs.npz`）以及气相与环境Δ学习模型（`modelvacgsdelta.npz`、`modelenvgsdelta.npz`）； - `models_nmethylacetamide.tar.gz`：气相与环境基础模型（`modelvacgs.npz`、`modelenvgs.npz`）以及气相Δ学习模型（`modelvacgsdelta.npz`）； - `models_alanine_dipeptide.tar.gz`：气相与环境基础模型（`modelvacgs.npz`、`modelenvgs.npz`）以及气相Δ学习模型（`modelvacgsdelta.npz`）。 上述模型可直接在ML-server中运行，用于开展ML/MM模拟。 ## 数据集生成脚本 压缩包`scripts_for_datagen.tar.gz`包含两份用于生成孤立分子几何构型与人工环境配置的Python脚本： 1. `nm_displacement.py`：读取Gaussian计算得到的高精度简正模与平衡几何构型，以计算简正模式位移； 2. `gen_charges.py`：从Gaussian输入文件读取孤立分子几何构型，从Gaussian输出文件读取马利肯电荷，计算分子周围的分层网格点，并选取部分点以放置随机外电荷。