data/README.md

---
license: cc-by-sa-4.0
language:
- en
tags:
- drug-target-interaction
- clinical-trials
- protein-protein-interaction
- negative-results
- benchmark
- bioinformatics
task_categories:
- text-classification
- question-answering
- text-generation
size_categories:
- 10M<n<100M
---

# NegBioDB: A Negative Results Database for Biomedical Sciences

NegBioDB is a large-scale database of experimentally confirmed negative results across three biomedical domains, paired with dual ML/LLM benchmarks for evaluating how well computational methods handle negative evidence.

## Overview

| Domain | Negative Results | Entities | Benchmark Tasks |
|--------|-----------------|----------|-----------------|
| DTI (Drug-Target Interaction) | 24,965,618 pairs | 919K compounds, 3.7K targets | ML (M1) + LLM (L1-L4) |
| CT (Clinical Trial Failure) | 102,850 results | 216K trials, 176K interventions | ML (M1-M2) + LLM (L1-L4) |
| PPI (Protein-Protein Interaction) | 2,220,786 pairs | 18.4K proteins | ML (M1) |

## Data Sources

### DTI Domain
| Source | License | Contribution |
|--------|---------|-------------|
| ChEMBL 34 | CC BY-SA 3.0 | Curated bioactivity data |
| PubChem BioAssay | Public domain | HTS screening results |
| BindingDB | CC BY-SA 3.0 | Binding measurements |
| DAVIS | CC BY 4.0 | Kinase selectivity panel |

### CT Domain
| Source | License | Contribution |
|--------|---------|-------------|
| AACT (ClinicalTrials.gov) | Public domain | Trial metadata |
| Open Targets | Apache 2.0 | Drug-target mappings |
| CTO | MIT | Clinical trial outcomes |
| Shi & Du 2024 | CC BY 4.0 | Safety/efficacy data |

### PPI Domain
| Source | License | Contribution |
|--------|---------|-------------|
| IntAct | CC BY 4.0 | Curated non-interactions |
| HuRI | CC BY 4.0 | Y2H systematic negatives |
| hu.MAP 3.0 | CC BY 4.0 | Complex-derived negatives |
| STRING v12.0 | CC BY 4.0 | Zero-score protein pairs |

## File Structure

### DTI Files
| File | Size | Rows | Description |
|------|------|------|-------------|
| `negbiodb_dti_pairs.parquet` | 139.4 MB | 24,965,618 | All negative DTI pairs with metadata |
| `negbiodb_m1_balanced.parquet` | 270.2 MB | 1,725,446 | M1 balanced dataset (1:1 pos:neg) |
| `negbiodb_m1_realistic.parquet` | 753.3 MB | 9,489,953 | M1 realistic dataset (1:10 pos:neg) |
| `negbiodb_m1_uniform_random.parquet` | 467.3 MB | 1,767,380 | Control: uniform random negatives |
| `negbiodb_m1_degree_matched.parquet` | 274.9 MB | 1,767,380 | Control: degree-matched negatives |
| `negbiodb_m1_balanced_ddb.parquet` | 1.0 GB | 1,725,446 | M1 balanced with degree-balanced split |

### CT Files
| File | Size | Rows | Description |
|------|------|------|-------------|
| `ct/negbiodb_ct_pairs.parquet` | 1.7 MB | 102,850 | All CT failure pairs |
| `ct/negbiodb_ct_m1_balanced.parquet` | 308.7 KB | 11,222 | CT-M1 balanced (success vs failure) |
| `ct/negbiodb_ct_m1_realistic.parquet` | 896.4 KB | 36,957 | CT-M1 realistic ratio |
| `ct/negbiodb_ct_m2.parquet` | 2.2 MB | 112,298 | CT-M2 multiclass failure category |

### PPI Files
| File | Size | Rows | Description |
|------|------|------|-------------|
| `ppi/negbiodb_ppi_pairs.parquet` | 685.2 MB | 2,220,786 | All negative PPI pairs |
| `ppi/ppi_m1_balanced.parquet` | 120.5 MB | 123,456 | PPI-M1 balanced (1:1 pos:neg) |
| `ppi/ppi_m1_realistic.parquet` | 744.8 MB | 679,008 | PPI-M1 realistic (1:10 ratio) |

### LLM Benchmark Files (DTI)
| File | Items | Description |
|------|-------|-------------|
| `llm_benchmarks/l1_mcq.jsonl` | 1,942 | L1: 4-class activity MCQ |
| `llm_benchmarks/l3_reasoning_pilot.jsonl` | 50 | L3: Scientific reasoning (pilot) |
| `llm_benchmarks/l4_tested_untested.jsonl` | 500 | L4: Tested vs untested discrimination |

## Benchmark Tasks

### ML Benchmarks
| Task | Domain | Type | Splits |
|------|--------|------|--------|
| M1 | DTI | Binary (active/inactive) | random, cold_compound, cold_target, degree_balanced |
| CT-M1 | CT | Binary (success/failure) | random, cold_drug, cold_condition, temporal, scaffold, cold_both |
| CT-M2 | CT | 7-way failure category | Same as CT-M1 |
| PPI-M1 | PPI | Binary (interact/non-interact) | random, cold_protein, cold_both, degree_balanced |

### LLM Benchmarks (DTI)
| Task | Type | Size | Description |
|------|------|------|-------------|
| L1 | MCQ classification | 1,600 | 4-class activity level |
| L2 | Structured extraction | ~100 | Extract results from abstracts |
| L3 | Reasoning | 50 | Explain compound-target inactivity |
| L4 | Discrimination | 400 | Tested vs untested pair |

## License

This dataset is released under **CC BY-SA 4.0**, due to the viral clause of ChEMBL's CC BY-SA 3.0 license. See the LICENSE file for details.

## Citation

```bibtex
@dataset{negbiodb2026,
  title={NegBioDB: A Negative Results Database for Biomedical Sciences},
  author={Jang, Jungwon},
  year={2026},
  url={https://github.com/jang1563/NegBioDB}
}
```