ChaseCheng/SAR-DRG

Dataset Card for SAR-DRG

Dataset Details

SAR-DRG is a scaffold-pocket dataset for realistic R-chain generation in lead optimization. It provides affinity-labeled R-chain samples organized by shared scaffold-pocket contexts, supporting SAR-informed molecular generation and evaluation. The dataset contains 96,158 samples across 29,387 scaffold-pocket groups.

Dataset Description

• —Curated by: Cheng et al.
• —Funded by: This work was supported by the Key Project of Technology Innovation and Application Development of Chongqing under Grant Nos. CSTB2024TIAD-STX0032 and 2025TIAD-STX0124, and the National Natural Science Foundation of China (22573012). The experimental and computational work in this research runs on the Huawei Cloud AI Compute Service. The authors appreciate the stable computing supply from this platform.
• —Shared by: Cheng Cheng

Dataset Architecture

• —`data_parquet/`: Contains the processed Parquet files organized according to the train/test split. These files are intended for direct loading through the Hugging Face datasets API.

• —`data_pt/`: Contains the processed PyTorch .pt files organized according to the train/test split. The files are stored as split-specific shards for efficient downloading and model training.

• —`data_pt.tar.gz`: Contains the original processed PyTorch data files before sharding. Each example corresponds to one .pt file, and the file name matches the data_id field in metadata.csv.

• —`data.tar.gz`: Contains the original SDF structure and PDB files used to construct the dataset. The corresponding file paths and identifiers are recorded in metadata.csv.

• —`dataset_split.json`: Stores the dataset splitting information, including train_group, test_group, train_data, and test_data. The *_group fields define the split at the group_id level, while the *_data fields define the split at the data_id level. Both identifiers can be mapped to entries in metadata.csv.

• —`metadata.csv`: Contains the original metadata table for all examples, including molecular SMILES, file paths, data_id, group_id, affinity values, and other dataset-related information.

Dataset Sources

• —Paper: SAR-DRG: A Scaffold-Pocket Dataset with Structure-Activity Relationships for Realistic R-Chain Generation in Lead Optimization

Citation

BibTeX:

Not available yet

APA:

Not available yet