🎯 Target-QA: The First QA Dataset Benchmarking Target Priorization Based on DepMap

---

📑 Dataset Summary

Target-QA is derived from the DepMap multi-omics and CRISPR screening cohorts, harmonized via BioMedGraphica.
It enables multi-modal reasoning by combining numeric evidence, topological knowledge and language context for CRISPR target prioritization.

This dataset supports the training and benchmarking of graph-augmented large language models (LLMs), such as GALAX, for reasoning across structured and unstructured biomedical information.

• 🔗 Source code: GitHub
• 🤗 Model parameters: Hugging Face

---

🛠️ Preprocessing Details

• Starting cohort: 985 DepMap cell lines
  • 649 annotated (cancerous)
  • 336 non-annotated or non-cancerous
• Target-QA subset: 363 overlapping with CRISPR gene effect data
• Omics modalities integrated into 834,809 entities:
  • Promoter: 86,238
  • Gene: 86,238
  • Transcript: 412,039
  • Protein: 121,419
• Knowledge graph integration:
  • Protein–protein interactions: 17,151,453 edges
  • Disease–target associations: 27,087,971 edges

---

📂 Data Splits

• Pretraining set: 336 samples
  • Train: 269
  • Test: 67
• Target-QA set: 363 samples
  • Train: 300
  • Test: 63

Test distribution includes: LUAD (7), BRCA (6), COAD/READ (5), PAAD (4), GBM (3), SARC (3), OV (3), SKCM (3), ESCA (3), SCLC (3), HNSC (2), LUSC (2), STAD (2), etc.

---

🧪 Supported Tasks & Benchmarks

• Graph-Augmented QA: Identify CRISPR targets from omics + KG context
• Knowledge Graph Reasoning: Subgraph extraction & signaling network prioritization
• Multi-Omic Target Prioritization: Integrated of epigenomic, genomic, transcriptomic and proteomic

---

📊 Dataset Structure

Example JSON

{
  "cell_line_name": "GAMG",
  "cell_line_id": "ACH-000098",
  "disease": "glioblastoma",
  "disease_bmgc_id": "BMGC_DS00965",
  "sample_dti_index": 123,
  "input": {
    "top_k_gene": {
      "hgnc_symbols": ["EGFR", "CDKN2A"],
      "protein_bmgc_ids": ["BMGC_PR01234"],
      "protein_llmname_ids": ["ENSP00000354587"]
    },
    "top_k_transcript": {...},
    "top_k_protein": {...},
    "knowledge_graph": {
      "disease_protein": {...},
      "ppi_neighbors": {...},
      "protein_relationships": ["BRCA1 → TP53"]
    }
  },
  "ground_truth_answer": {
    "hgnc_symbols": ["TP53", "EGFR"],
    "protein_bmgc_ids": ["BMGC_PR00987", "BMGC_PR04567"],
    "protein_llmname_ids": ["ENSP00000439978"]
  }
}

📝 Prompt Design

Initial Prompt (P_init_n)

Inputs:

• Top-10 ranked genes, transcripts, proteins
• Disease-associated proteins from KG
• Known PPI and disease–protein relationships

Output:

• 100 vulnerability genes r_init[n,1...100]

Refined Prompt (P_final_n)

Inputs:

• Same as above
• Subsignaling gene regulatory network (from graph generator)

Output:

• Refined 100 vulnerability genes r_hat[n,1...100]

---

⚖️ Licensing

This dataset is based on DepMap data and is subject to the DepMap Terms of Use:

• Free for research purposes only
• Commercial use prohibited without explicit Broad Institute license
• ML models may be trained for internal research use or shared for non-profit research
• Attribution to Broad Institute / DepMap is required

🔗 DepMap Terms of Use

---

📚 Citation

If you use Target-QA or GALAX, please cite:

@misc{zhang2025galax,
  title={GALAX: Graph-Augmented Language Model for Explainable Reinforcement-Guided Subgraph Reasoning in Precision Medicine},
  author={Heming Zhang and Di Huang and Wenyu Li and Michael Province and Yixin Chen and Philip Payne and Fuhai Li},
  year={2025},
  eprint={2509.20935},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2509.20935}
}