🎯 HippoTarget

A curated drug-target interaction dataset, teaching LLMs which molecules bind to which proteins.

---

💡 Overview

Welcome to HippoTarget, the fifth member of the ZemResearch Hippo Ecosystem. Before a drug can do anything useful in the body, it first has to bind to the right protein — 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.

🧬 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
🧬 HippoCrates	Molecular structures & SMILES	1.46M rows
⚗️ HippoSynth	Chemical reactions & synthesis	50K rows
🎯 HippoTarget (you are here)	Drug-target interaction	15.5K rows
🫀 HippoLv	ADMET & drug behavior in the body	~9.4K rows
☠️ HippoXic	Toxicology & clinical safety	~10.6K rows

The pipeline flows naturally: HippoCrates (what the molecule looks like) → HippoSynth (how it's made) → HippoTarget (what it binds to) → HippoLv (how it behaves in the body) → HippoXic (whether it's safe).

🧼 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.

📦 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.

📝 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 — compound SMILES and/or protein sequence
• output: The expected answer

Example Rows

Binding Interaction:

{
  "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:

{
  "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..."
}

💻 How to Use

from datasets import load_dataset

dataset = load_dataset("ZemResearch/HippoTarget")

print(f"Total drug-target pairs: {len(dataset['train'])}")
print(dataset['train'][0])

⚠️ Intended Use & Limitations

HippoTarget is designed for research and educational purposes — 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.

🤝 Citation & Collaboration

Created with ❤️ by 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.