Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
License:
| language: | |
| - en | |
| license: apache-2.0 | |
| task_categories: | |
| - text-generation | |
| tags: | |
| - chemistry | |
| - smiles | |
| - drug-target-interaction | |
| - protein | |
| - binding-affinity | |
| - pharmacology | |
| - drug-discovery | |
| - alpaca | |
| pretty_name: HippoTarget 15K Drug-Target Interaction Dataset | |
| size_categories: | |
| - 10K<n<100K | |
| dataset_info: | |
| features: | |
| - name: instruction | |
| dtype: string | |
| - name: input | |
| dtype: string | |
| - name: output | |
| dtype: string | |
| splits: | |
| - name: train | |
| num_bytes: 1896666 | |
| num_examples: 15520 | |
| <div align="center"> | |
| <h1>π― HippoTarget</h1> | |
| <p><i>A curated drug-target interaction dataset, teaching LLMs which molecules bind to which proteins.</i></p> | |
| </div> | |
| <p align="center"> | |
| <img src="https://img.shields.io/badge/Dataset_Size-15.5K_Rows-blue?style=for-the-badge" alt="Size"> | |
| <img src="https://img.shields.io/badge/Format-Apache_Parquet-green?style=for-the-badge" alt="Format"> | |
| <img src="https://img.shields.io/badge/Language-Global_English-purple?style=for-the-badge" alt="Language"> | |
| <img src="https://img.shields.io/badge/License-Apache 2.0-orange?style=for-the-badge" alt="License"> | |
| </p> | |
| <hr> | |
| ## π‘ Overview | |
| Welcome to **HippoTarget**, the fifth member of the [ZemResearch](https://huggingface.co/ZemResearch) Hippo Ecosystem. Before a drug can do anything useful in the body, it first has to bind to the right protein β like a key fitting into a lock. HippoTarget teaches LLMs exactly that: given a small molecule, which protein does it interact with? | |
| This dataset combines real binding interaction data with a curated list of FDA-approved drugs and their known protein targets, giving models exposure to both experimental binding relationships and clinically validated drug-target pairs. | |
| ## 𧬠Part of the Hippo Ecosystem | |
| HippoTarget is designed to work alongside the other Hippo datasets, together forming an end-to-end drug discovery pipeline: | |
| | Dataset | Focus | Size | | |
| |---|---|---| | |
| | 𧬠[HippoCrates](https://huggingface.co/datasets/ZemResearch/HippoCrates) | Molecular structures & SMILES | 1.46M rows | | |
| | βοΈ [HippoSynth](https://huggingface.co/datasets/ZemResearch/HippoSynth) | Chemical reactions & synthesis | 50K rows | | |
| | π― **HippoTarget** (you are here) | Drug-target interaction | 15.5K rows | | |
| | π« [HippoLv](https://huggingface.co/datasets/ZemResearch/HippoLv) | ADMET & drug behavior in the body | ~9.4K rows | | |
| | β οΈ [HippoXic](https://huggingface.co/datasets/ZemResearch/HippoXic) | Toxicology & clinical safety | ~10.6K rows | | |
| The pipeline flows naturally: **HippoCrates** (what the molecule looks like) β **HippoSynth** (how it's made) β **HippoTarget** (what it binds to) β **HippoLv** (how it behaves in the body) β **HippoXic** (whether it's safe). | |
| ## π§Ό Curation Process | |
| 1. **Multi-source aggregation:** We combined real experimental binding interaction data with a clinically curated list of FDA-approved drugs and their protein targets. | |
| 2. **Text parsing:** Binding interaction sentences were parsed to extract SMILES structures and protein sequences using pattern matching, handling multiple sentence format variations found in the source data. | |
| 3. **RDKit validation:** Every SMILES string was validated using RDKit to confirm it represents a chemically valid molecule. | |
| 4. **Protein sequence validation:** Every protein sequence was checked to contain only standard amino acid characters within a reasonable length range. | |
| 5. **Deduplication:** We removed 199 duplicate rows based on input/output content. | |
| 6. **Result:** 15,520 clean, validated drug-target interaction rows. | |
| ## π¦ Data Sources | |
| HippoTarget was built by merging and curating data from the following public sources: | |
| | Source | Type | Contribution | | |
| |---|---|---| | |
| | `jablonkagroup/chempile-lift` (uniprot_binding_single) | Experimental binding interactions | ~6.4K rows | | |
| | `alimotahharynia/approved_drug_target` | FDA-approved drugs & their protein targets | ~9.4K rows | | |
| All sources are publicly available. The approved drug target data was curated with pharmacist and physician review, sourced from DrugBank, ChEMBL, ZINC20, and UniProt. | |
| ## π Data Structure | |
| HippoTarget uses the standard **Alpaca instruction-tuning format**. Each row has three columns: | |
| - `instruction`: What the model is being asked to do (determine binding, identify target, etc.) | |
| - `input`: The context β compound SMILES and/or protein sequence | |
| - `output`: The expected answer | |
| ### Example Rows | |
| **Binding Interaction:** | |
| ```json | |
| { | |
| "instruction": "Determine whether this molecule interacts with the specified protein sequence.", | |
| "input": "Compound SMILES: COc1ccccc1N1CCN(CCN2C=CC3(CCCC3)CC2=O)CC1\nProtein sequence: MSLEQKKGADIISKILQIQNSIGKTTSPSTLKTKLSEISRKEQENARIQSKLSDLQKKKIDIDNKLLKEKQNLIKEEILERKKL...", | |
| "output": "Yes, this compound binds to the given protein target. The SMILES structure COc1ccccc1N1CCN(CCN2C=CC3(CCCC3)CC2=O)CC1 has documented binding affinity with the specified peptide sequence." | |
| } | |
| ``` | |
| **Approved Drug-Target Identification:** | |
| ```json | |
| { | |
| "instruction": "Identify the protein target for the following approved drug compound.", | |
| "input": "Drug SMILES: CC(=O)Oc1ccccc1C(=O)O", | |
| "output": "This approved drug compound targets the protein with sequence: MDHQDPYSVQATAAIAAAITFLILFTIFGNALVILAV..." | |
| } | |
| ``` | |
| ## π» How to Use | |
| ```python | |
| from datasets import load_dataset | |
| dataset = load_dataset("ZemResearch/HippoTarget") | |
| print(f"Total drug-target pairs: {len(dataset['train'])}") | |
| print(dataset['train'][0]) | |
| ``` | |
| ## β οΈ Intended Use & Limitations | |
| HippoTarget is designed for **research and educational purposes** β fine-tuning chemistry and bioinformatics LLMs, exploring drug-target interaction prediction, and building cheminformatics tools. | |
| A few things to keep in mind: | |
| - Protein sequences in this dataset are truncated to 200 characters for practical instruction-tuning purposes. For full-length sequences, refer to the original UniProt source. | |
| - Binding interactions represent documented relationships from the source data, not experimentally verified binding affinity values (Kd, IC50, Ki are not included in this version). | |
| - This dataset should not be used as a substitute for proper computational docking or wet-lab validation in real drug discovery pipelines. | |
| - Always validate model predictions with proper bioinformatics tools before applying them to real research. | |
| ## π€ Citation & Collaboration | |
| Created with β€οΈ by [ZemResearch](https://huggingface.co/ZemResearch). If you use HippoTarget in your research or projects, we'd love to hear about it! Feel free to open a discussion in the community tab. | |