Computed surface and chemical potentials, expansion coefficients, structures, models and results for the PMFPredictor Toolkit

The PMFPredictor toolkit enables the prediction of the potentials of mean force describing the interaction between a surface and a small molecule in aqueous solution, which would otherwise be obtained from lengthy metadynamics simulations. This repository contains files to enable the operation of the toolkit, with source code available at https://github.com/ijrouse/PMFPredictor-Toolkit and corresponding to release v0.5-alpha. In PMFPredictor-Repository.zip we provide supplementary data necessary for the operation of the PMFPredictor Toolkit including: Structures of surfaces ("Structures/Surfaces") and chemicals ("Structures/Chemicals") in a united tabulated (.csv) format, listing x/y/z co-ordinates, atom IDs, mass (in amu), charge (in elementary units), Lennard Jones 6-12 parameters: sigma (in nm) and epsilon (in kJ/mol). Interaction potentials of surfaces ("SurfacePotentials") and chemicals ("ChemicalPotentials") with probe atoms and molecules in tabulated format with distances relative to reference points in nm and energies in kJ/mol. Also included in these folders are the potentials with the molecular probes in individual files. Hypergeometric expansion coefficients of the interaction potentials ("Datasets/SurfacePotentialCoefficientsNoise-1-oct12.csv" and "Datasets/ChemicalPotentialCoefficients-oct10.csv") in tabulated form, corresponding to potentials with units of nm for distance and kJ/mol for energy. Descriptions of the headers are provided in DatasetHeaderDescription.txt, included in the archive. Trained TensorFlow models for the prediction of potentials of mean force from HG interaction coefficients, suitable for loading via the Keras backend. PMFs generated for a range of surfaces and chemicals as output from the trained model, in both text format and figures showing comparisons to training PMFs where available. PMFs are supplied as tabulated data with comma separated values of distance in nm and interaction energies in kJ/mol. Adsorption energies in kJ/mol evaluated at T=300K extracted from all PMFs and compared to the values obtained from known PMFs where available. The surface_pmfpredictor.zip archive contains PMFs selected for the operation of the UnitedAtom software package for the calculation of protein-nanoparticle interactions. This data is included in the main repository file and provided separately to avoid the download of unnecessary data if only the final PMFs are required. As with the main set, these are provided in tabulated form with distance [nm], energy [kJ/mol] pairs. This repository also contains the sets of figures illustrating these PMFs for each surface. Both archives contain further information on the contents, including descriptions of the surfaces and chemicals for which PMFs are computed. We also supply the training data used to build the model in a separate archive, PMFPredictor-TrainingData.zip, along with a text file containing descriptions of all headers in this file. This training data is quite large when uncompressed, c.a. 7 Gb, hence its exclusion from the main archive. If you use results from this repository please cite the following paper in addition to the repository itself: I. Rouse, V. Lobaskin, Machine-learning based prediction of small molecule -- surface interaction potentials, arXiv:2211.07999<br> https://arxiv.org/abs/2211.07999

PMFPredictor工具包（PMFPredictor toolkit）可用于预测平均力势（potentials of mean force），该势描述水溶液中表面与小分子间的相互作用，此类相互作用势原本需通过耗时的元动力学（metadynamics）模拟获取。
本仓库包含可运行该工具包的相关文件，其源代码托管于https://github.com/ijrouse/PMFPredictor-Toolkit，对应版本为v0.5-alpha测试版。
在PMFPredictor-Repository.zip中，我们提供了PMFPredictor工具包运行所需的补充数据，具体包括：以统一制表格式（.csv）存储的表面结构（"Structures/Surfaces"）与化学品结构（"Structures/Chemicals"）文件，其中列出了x/y/z坐标、原子ID、质量（单位：原子质量单位amu）、电荷（单位：基本电荷元）、伦纳德-琼斯6-12势参数：σ（单位：纳米nm）与ε（单位：千焦每摩尔kJ/mol）；表面与化学品分别与探针原子、探针分子的相互作用势文件（"SurfacePotentials"与"ChemicalPotentials"），采用制表符分隔格式存储，其中距离以相对于参考点的纳米数表示，能量单位为千焦每摩尔，此类文件夹中还包含了各分子探针对应的相互作用势文件；相互作用势的超几何展开系数文件（"Datasets/SurfacePotentialCoefficientsNoise-1-oct12.csv"与"Datasets/ChemicalPotentialCoefficients-oct10.csv"），同样采用制表符分隔格式，其中距离单位为纳米，能量单位为千焦每摩尔。该压缩包中还附带了DatasetHeaderDescription.txt文件，用于说明各数据表头的含义；经训练的TensorFlow（TensorFlow）模型，可通过Keras（Keras）后端加载，用于基于超几何相互作用系数预测平均力势；由训练好的模型生成的一系列表面与化学品的平均力势（PMFs），包含文本格式数据，以及可与训练所得平均力势（若有）进行对比的可视化图表。此类平均力势以制表符分隔的数值形式提供，包含距离（单位：纳米）与相互作用能量（单位：千焦每摩尔）的逗号分隔值；从所有平均力势中提取的、在T=300K下计算得到的吸附能（单位：千焦每摩尔），并与已知平均力势（若有）对应的吸附能进行对比。
surface_pmfpredictor.zip压缩包包含了为UnitedAtom（UnitedAtom）软件包（用于计算蛋白质-纳米颗粒相互作用）运行所需的平均力势数据。此类数据已包含在主仓库文件中，同时单独提供，以便仅需最终平均力势的用户避免下载不必要的文件。与主数据集一致，此类数据同样以距离[nm]、能量[kJ/mol]的键值对制表符分隔格式提供。本仓库还包含了各表面对应的平均力势可视化图表集。
两个压缩包均包含了更多关于数据内容的说明，包括用于计算平均力势的表面与化学品的相关描述。我们还单独提供了用于构建模型的训练数据压缩包PMFPredictor-TrainingData.zip，以及一份用于说明该训练数据中所有表头含义的文本文件。该训练数据解压后体积约为7吉字节（Gb），因此未纳入主压缩包中。
若您使用本仓库的研究结果，除引用本仓库外，还请引用以下论文：I. Rouse, V. Lobaskin, 基于机器学习的小分子-表面相互作用势预测，arXiv:2211.07999<br> https://arxiv.org/abs/2211.07999