Syzseisus/S3-Prompt

S³-Prompt — Preprocessed MoleculeNet Data

Preprocessed data used by S³-Prompt: Resolving Activity Cliffs via Structure-Disentangled Molecular Graph-Text Alignment (CIKM '26). Derived from MoleculeNet; each dataset is cleaned, embedded, split, and equipped with scaffold-clustering / retrieval caches so the experiments run without recomputation.

• 💻 Code: https://github.com/Syzseisus/S3-Prompt
• 📦 Results & checkpoints: https://huggingface.co/Syzseisus/S3-Prompt

Datasets

Name	Task	Metric
BACE	classification	ROC-AUC
BBBP	classification	ROC-AUC
ESOL	regression	RMSE
LIPOPHILICITY	regression	RMSE

Per-dataset contents

{DATASET}/
├── raw.csv / trimmed.csv              # original & cleaned SMILES + labels
├── train.csv / valid.csv / test.csv   # 8:1:1 split (stratified for classification)
├── embed_fp_r=2_b=1024.pk             # Morgan fingerprint embeddings (r=2, 1024 bits)
├── embed_molt5_small_d512.pt          # MolT5-small encoder embeddings
├── topk_fp/ , topk_molt5-small/       # top-k neighbour caches (retrieval used by all paper runs)
├── mining/ (classification) or mining_reg/ (regression)      # per-query mining info for valid/test
└── {fp,molt5-small}_eps={eps}_{random,similarity}-sampling/  # classification only:
    ├── clusters_idx.pk / clusters_scaf.pk / meta.json        #   scaffold clustering
    ├── query_to_shot_{valid,test}_k={k}.json                 #   cached few-shot pools
    └── plots/                                                #   clustering visualizations

The scaffold-clustering folders are only needed for --method pool_clustering; the paper's runs use the topk caches.

Retrieval modalities: FP (Morgan fingerprint, eps=0.6) for binding/partition tasks, MolT5 (eps=0.8) for macroscopic-property tasks.

Usage

from huggingface_hub import snapshot_download
snapshot_download("Syzseisus/S3-Prompt", repo_type="dataset", local_dir="MoleculeNetICL_S3-Prompt")

Point the code at it via DATA_ROOT in utils/constants.py (default ../MoleculeNetICL_S3-Prompt).

Citation

@inproceedings{s3prompt2026,
  title     = {S$^3$-Prompt: Resolving Activity Cliffs via Structure-Disentangled Molecular Graph-Text Alignment},
  author    = {Cho, Wooseong Syz and Park, Ha-Myung},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM '26)},
  year      = {2026},
  address   = {Rome, Italy},
  publisher = {ACM},
  doi       = {10.1145/3799682.3841121},
}