APScore Dataset: An interatomic distance probability model enables physically realistic and interpretable protein–ligand interaction prediction

APScore Dataset This repository contains the processed datasets and source data accompanying the manuscript: An interatomic distance probability model enables physically realistic and interpretable protein–ligand interaction prediction Repository Structure APScore_Data/ ├── Processed_data/ │ ├── pdbbind_dataset/ │ └── pdbbind_cleansplit/ ├── Source_data/ │ ├── Source_data_pdbbind_original.xlsx │ ├── Source_data_pdbbind_cleansplit.xlsx │ └── Source_data_clash.xlsx ├── README.md └── LICENSE Processed Data The Processed_data directory contains binary feature files generated during the data preprocessing stage. These files are directly loaded by the APScore training pipeline and are provided to facilitate reproduction of the reported experiments. pdbbind_dataset Processed feature files generated from the original PDBbind dataset without structural cleaning or redundancy removal. pdbbind_cleansplit Processed feature files generated from the structurally cleaned and redundancy-reduced PDBbind dataset. This set strictly ensures no scaffold overlap between the training and test sets to evaluate out-of-distribution generalization and test for molecular scaffold memorization. Feature File Descriptions Each directory contains the following binary files representing the extracted physical and topological features: File Description angles.bin Bond angle features lig_atom_feats.bin Ligand atom features lig_neighbor.bin Ligand neighborhood information prot_atom_feats.bin Protein atom features prot_neighbor.bin Protein neighborhood information ll_atom1_feats.bin Features of the first ligand atom in ligand–ligand atom pairs ll_atom2_feats.bin Features of the second ligand atom in ligand–ligand atom pairs ll_neighbor1.bin Neighbor features associated with the first ligand atom ll_neighbor2.bin Neighbor features associated with the second ligand atom ll_dist.bin Ligand–ligand atomic distance features pl_dist.bin Protein–ligand atomic distance features metadata.bin Metadata describing each processed sample These processed files are intended for direct use during model training and evaluation. Source Data The Source_data directory contains the numerical values underlying the figures and tables presented in the manuscript and Supplementary Information, as well as the exact data split configurations. File Description Source_data_pdbbind_original.xlsx Source data for benchmarking experiments using the original PDBbind dataset Source_data_cleansplit.xlsx Source data for generalization capability tests using the CleanSplit dataset Source_data_clash.xlsx Source data for structural rationality and steric clash analyses PDBbind_data_split_cleansplit.json Explicit lists of PDB IDs assigned to the training, validation, and test sets for the CleanSplit benchmark, ensuring exact reproducibility of the strict scaffold separation Code Availability The complete source code, training scripts, and instructions for reproducing the experiments and feature extraction are available from the project's GitHub repository. Notes The processed binary files were generated from the PDBbind dataset using the physically-guided preprocessing pipeline described in the manuscript. Users requiring the original 3D molecular structure files (SDF/PDB) should obtain the PDBbind dataset from its official distribution (http://www.pdbbind.org.cn/). Citation If you use these datasets, feature engineering pipelines, or the APScore model in your research, please cite the accompanying publication.