NegBioDB

A Negative-Results Database and Dual ML/LLM Benchmark for Biomedical Sciences

An estimated 90% of biomedical experiments produce null or inconclusive findings, yet the overwhelming majority remain unpublished. NegBioDB systematically aggregates experimentally confirmed negative results across five biomedical domains and pairs them with a dual-track benchmark — traditional ML prediction and modern LLM reasoning — that quantifies how publication bias propagates into AI systems.

This Hugging Face dataset mirrors the pre-built ML and LLM splits from the main repository.

---

Overview

Domain	Negative results	Key entities	Sources	ML runs	LLM runs
DTI — Drug–Target Interaction	30,459,583	919K compounds, 3.7K targets	ChEMBL, PubChem, BindingDB, DAVIS	24 / 24	81 / 81
CT — Clinical Trial Failure	132,925	177K interventions, 56K conditions	AACT, CTO, Open Targets, Shi & Du	108 / 108	80 / 80
PPI — Protein–Protein Interaction	2,229,670	18.4K proteins	IntAct, HuRI, hu.MAP 3.0, STRING	54 / 54	80 / 80
GE — Gene Essentiality (DepMap)	28,759,256	19,554 genes, 2,132 cell lines	DepMap CRISPR + RNAi	42 / 42	80 / 80
VP — Variant Pathogenicity	2,442,718	2.43M variants, 18.4K genes, 10K diseases	ClinVar, gnomAD, ClinGen, CADD/REVEL/AlphaMissense	72 / 72	20 / 20
Total	~64.0M	—	17 sources	300	341

PPI export rows after split filtering: 2,220,786. VP M1 balanced export: 1,255,150 rows.

---

Why NegBioDB?

Most biomedical ML benchmarks rely on synthetic negatives — random non-edges in a graph, decoy compounds, or unobserved pairs — which trivially leak through degree statistics or prior frequency. NegBioDB instead provides experimentally confirmed negatives: failed assays, failed clinical trials, validated non-interactions, non-essential genes in specific contexts, and benign variants. This lets you:

• Quantify the gap between random-split AUROC and real generalization (cold-entity / temporal / scaffold splits).
• Stress-test LLMs on the L4 task (tested vs. untested) — a discriminator that exposes memorization vs. reasoning.
• Compare your method on a publication-bias-corrected baseline, not a degree-matched proxy.

---

File structure

DTI — root level

File	Size	Rows	Description
negbiodb_dti_pairs.parquet	139 MB	~25M	All negative DTI pairs with 5 split columns + provenance
negbiodb_m1_balanced.parquet	270 MB	1,725,446	M1 balanced (1:1 active:inactive)
negbiodb_m1_realistic.parquet	753 MB	9,489,953	M1 realistic (1:10)
negbiodb_m1_balanced_ddb.parquet	1.0 GB	1,725,446	Degree-balanced split
negbiodb_m1_uniform_random.parquet	467 MB	1,767,380	Control: uniform random
negbiodb_m1_degree_matched.parquet	275 MB	1,767,380	Control: degree-matched
chembl_positives_pchembl6.parquet	—	863K	ChEMBL actives (pChEMBL ≥ 6)
compound_names.parquet	—	144K	Compound names for LLM tasks

CT — ct/

File	Rows	Description
ct/negbiodb_ct_pairs.parquet	102,850	All failure pairs, 6 splits
ct/negbiodb_ct_m1_balanced.parquet	11,222	Binary (success / failure), 1:1
ct/negbiodb_ct_m1_realistic.parquet	36,957	Binary, ~1:6
ct/negbiodb_ct_m1_smiles_only.parquet	3,878	SMILES-resolved subset
ct/negbiodb_ct_m2.parquet	112,298	7-way failure-mode classification

PPI — ppi/

File	Rows	Description
ppi/negbiodb_ppi_pairs.parquet	2,220,786	All negative pairs, 4 splits
ppi/ppi_m1_balanced.parquet	123,456	M1 (1:1)
ppi/ppi_m1_realistic.parquet	679,008	M1 (1:10)
ppi/ppi_m1_balanced_ddb.parquet	—	Degree-balanced split
ppi/ppi_m1_uniform_random.parquet	—	Control
ppi/ppi_m1_degree_matched.parquet	—	Control

GE — ge/

File	Description
ge/negbiodb_ge_pairs.parquet	22.5M gene–cell-line pairs, 5 split columns (~770 MB)
ge_gene_aggregates.parquet	Per-gene aggregated essentiality features

VP — vp_ml/

File	Rows	Description
vp_ml/vp_m1_balanced.parquet	1,255,150	M1 balanced (gold/silver positives, 1:1)
vp_ml/vp_m1_realistic.parquet	2,442,718	M1 realistic (full negative set)

LLM benchmarks — llm_benchmarks/ and per-domain *_llm/

LLM datasets cover four reasoning levels (L1–L4) per domain:

Level	Question	Evaluation
L1	Multiple-choice: which is not a known interaction / failure / ...	accuracy, MCC
L2	Structured extraction into a typed schema	field-level F1, schema compliance
L3	Open-ended scientific reasoning on a negative finding	LLM-judge rubric (1–5 across 4–6 axes)
L4	Discrimination: tested-as-negative vs. untested	MCC; contamination-flag analysis

Models evaluated under both zero-shot and 3-shot configurations: GPT-4o-mini, Claude Haiku 4.5, Gemini 2.5 Flash, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct.

---

Benchmark tasks

ML

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
GE-M1	GE	Binary (essential / non-essential)	random, cold_gene, cold_cell_line, cold_both, degree_balanced
VP-M1	VP	Binary (pathogenic / benign)	random, cold_gene, cold_disease, temporal

Metrics: LogAUC[0.001,0.1] (primary, early-enrichment), BEDROC (α=20), EF@1% / EF@5%, AUPRC, MCC, AUROC.

---

Headline findings

ML — the choice of negatives shapes every metric

DTI model	Random (NegBioDB)	Random (degree-matched)	Cold-target
DeepDTA	0.833	0.919	0.325
GraphDTA	0.843	0.967	0.241
DrugBAN	0.830	0.955	0.151

• CT. Confirmed-failure negatives are trivially separable for binary tasks (AUROC ≈ 1.0). The 7-way failure-mode classification (CT-M2) remains hard (best macro-F1 = 0.51).
• PPI. PIPR cold-both AUROC drops below random (0.409); MLP-on-features stays robust (0.950).
• GE. Cold-gene splits expose generalization gaps invisible under random splits.
• VP. Random splits saturate (AUROC 0.995 / MCC 0.932); cold-disease splits expose AUROC-vs-MCC calibration failures.

LLM — L4 (tested vs. untested) is where models actually differ

Domain	L4 MCC range	Memorization signal
DTI	≤ 0.18	Not detected
GE	≤ 0.22	Not detected
PPI	0.33–0.44	Yes — pre-2015 pairs identified at 59–79%; post-2020 at 7–25%
CT	0.48–0.56	Not detected
VP	n/a (single-class test)	n/a

Across PPI / GE / DC / CP / VP, L3 (open-ended reasoning, judge-graded) shows zero-shot ≫ few-shot for most models — providing exemplars degrades reasoning quality, a robust cross-domain pattern.

---

Quickstart

from huggingface_hub import hf_hub_download
import pandas as pd

# DTI: pull the M1 balanced split
path = hf_hub_download(
    repo_id="jang1563/NegBioDB",
    filename="negbiodb_m1_balanced.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)
print(df.head())

# CT: subdirectory addressing works the same way
ct_path = hf_hub_download(
    repo_id="jang1563/NegBioDB",
    filename="ct/negbiodb_ct_m1_balanced.parquet",
    repo_type="dataset",
)
ct_df = pd.read_parquet(ct_path)

For end-to-end ETL (raw download → SQLite → split export → ML / LLM evaluation), use the main repository which provides per-domain CLI entry points and SLURM scripts.

---

Data sources & licenses

Domain	Source	License	Contribution
DTI	ChEMBL v36	CC BY-SA 3.0	Curated bioactivity
	PubChem BioAssay	Public Domain	HTS screening
	BindingDB	CC BY 3.0	Binding measurements
	DAVIS	Public	Kinase selectivity
CT	AACT / ClinicalTrials.gov	Public Domain	Trial metadata
	CTO	MIT	Trial outcomes
	Open Targets	Apache 2.0	Drug–target mappings
	Shi & Du 2024	CC BY 4.0	Safety / efficacy
PPI	IntAct	CC BY 4.0	Curated non-interactions
	HuRI	CC BY 4.0	Y2H systematic negatives
	hu.MAP 3.0	MIT	Complex-derived
	STRING v12.0	CC BY 4.0	Zero-evidence pairs
GE	DepMap CRISPR (Chronos)	CC BY 4.0	Gene essentiality
	DepMap RNAi (DEMETER2)	CC BY 4.0	RNAi screens
VP	ClinVar	Public Domain	Clinical variants
	gnomAD v4.1	CC0	Population variants
	ClinGen	CC0	Gene–disease validity
	CADD	Free non-commercial	Functional scores
	REVEL	Free	Missense pathogenicity
	AlphaMissense	CC BY-NC-SA 4.0	Missense pathogenicity

Per-source attribution (versions, download dates, normalization steps) is in docs/methodology_notes.md.

---

Citation

@misc{negbiodb2026,
  title  = {NegBioDB: A Negative-Results Database and Dual ML/LLM Benchmark
            for Biomedical Sciences},
  author = {Kim, JangKeun},
  year   = {2026},
  url    = {https://github.com/jang1563/NegBioDB}
}

---

License

This dataset is released under CC BY-SA 4.0 — required by the viral clause of ChEMBL's CC BY-SA 3.0. All redistributed source data retain their original licenses (table above). AlphaMissense scores are non-commercial only; commercial users should remove the AlphaMissense columns before downstream use.