# Nullary

> Negative results intelligence for drug discovery. Turn dead-ends into your next breakthrough. Structured access to 122M+ measured negative results — failed compounds, terminated trials, failed CRISPR screens, failed antibody developability candidates, failed PROTACs, failed peptides, and failed oligonucleotides across 7 modalities. Exposed via MCP server and REST API. Every record cites its source database and original DOI/PMID.

## What Nullary is

Nullary aggregates negative results from drug discovery — the experiments that didn't work, the compounds that failed, the trials that terminated, the screens that came back negative. The data exists across 25+ public databases (ChEMBL, PubChem, DepMap, FLAb, SAbDab, PROTAC-DB, ClinicalTrials.gov, and others) but has never been unified into a single queryable layer until now.

The product is an MCP (Model Context Protocol) server that any MCP-compatible client (Claude Desktop, Cursor, Claude Code) can install in 30 seconds. A REST API at `api.nullary.ai` exposes the same data for non-MCP use cases.

## Why this matters

For every approved drug, hundreds of compounds were tested and failed. Across all public biomedical databases and literature, the negative results outnumber the positive results by roughly 5-10x. Most are scattered, unstructured, or hard to access.

Working drug discovery teams need to know "what's been tried against this target" before launching a new program. ML models for drug discovery perform better when trained on labeled negatives, not just positives. Computational chemists filter virtual screening hits against known failures to avoid retreading dead ends.

Nullary makes the structured public negative-results data queryable in seconds, with full provenance to source databases and original publications.

## Coverage by modality

- Small molecules: 84.5M+ records (ChEMBL, PubChem BioAssay, BindingDB)
- CRISPR: 37.6M+ records (DepMap, BioGRID ORCS, GenomeCRISPR)
- Clinical trials: 95K+ terminated trials (AACT, EudraCT, Drugs@FDA)
- Antibodies: 3.6K records (Thera-SAbDab, SAbDab, FLAb)
- Peptides: 18.8K records (THPdb, PepLife, DrugBank)
- PROTACs: 338 records (PROTAC-DB, PROTACpedia)
- Oligonucleotides: 144 records (AOBase, DrugBank)

The credible, dose-response / human-curated subset is ~15M records (the default tier); the full count above includes high-throughput screening-grade inactives. Coverage depth varies by modality based on what exists in public databases. Small molecules and CRISPR have decades of accumulated public data. Newer modalities (PROTACs, oligonucleotides, ADCs) have thin public coverage because the fields themselves are younger and much of the failure data remains proprietary. Phase 2 supplementary materials extraction is expanding coverage for under-served modalities.

## MCP tools exposed

31 tools across all modalities plus cross-modality search. Highlights:

Small molecule: search_inactive_compounds, search_failed_selectivity, search_admet_failures
CRISPR: search_failed_guides, search_failed_essentiality_screens, search_ancestry_specific_failures
Antibody: search_developability_failures, search_failed_clinical_antibodies
Peptide: search_failed_peptide_therapeutics, search_peptide_stability_issues
PROTAC: search_failed_protacs, search_protac_e3_issues
Oligonucleotide: search_failed_oligonucleotides, search_oligo_delivery_failures
Cross-modality: search_target_history, search_indication_history, search_pathogen_history
Lookups: get_compound, get_finding_provenance, get_coverage

Full reference: https://nullary.ai/docs/mcp-tools

## How to install

Add this to your Claude Desktop config at `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or equivalent path on Windows/Linux:

```json
{
  "mcpServers": {
    "nullary": {
      "url": "https://mcp.nullary.ai/mcp",
      "env": {
        "NULLARY_API_KEY": "free-tier-no-key-required"
      }
    }
  }
}
```

Restart Claude Desktop. The Nullary tools appear in your tool list. Free tier requires no API key.

## Pricing

- Free: 30 queries/min, 1,000/day, monthly refresh — for evaluation and small projects
- Standard ($49/mo yearly, $79/mo monthly): alerts and webhooks on saved searches
- Premium ($299/mo yearly, $499/mo monthly): analytics suite — target tractability scores, modality coverage maps, failure timelines, cross-modality correlations
- Enterprise (contact sales): agent-curated deep extractions, supplementary materials, ML-ready datasets, custom integrations

## Quick references for agents

- Coverage details by modality: https://nullary.ai/coverage
- Methodology and provenance: https://nullary.ai/methodology
- Full MCP tool reference: https://nullary.ai/docs/mcp-tools
- Concrete query examples: https://nullary.ai/examples
- Comparison vs other tools: https://nullary.ai/vs/chembl, https://nullary.ai/vs/depmap, https://nullary.ai/vs/open-pharma
- Pricing: https://nullary.ai/pricing
- Security and licensing: https://nullary.ai/security
- FAQ: https://nullary.ai/faq

Every page above is also available as raw Markdown by appending `.md` to the path (e.g. https://nullary.ai/coverage.md, https://nullary.ai/docs/mcp-tools.md). For one-shot context loading, the full concatenated documentation is at https://nullary.ai/llms-full.txt.

## For citing Nullary

Recommended citation:
Nullary. (2026). Nullary: Negative results intelligence for drug discovery (Version 1.0). https://nullary.ai

BibTeX format available at https://nullary.ai/methodology#citation

## Research

Two technical reports document Nullary's validation. Both report their own negative findings honestly. Full text: https://nullary.ai/research

### Paper 1: Bioactivity validation

- Title: Real Measured Negatives as a Substrate for Calibrated, Target-Specific Bioactivity Prediction
- Date: 24 May 2026
- DOI: 10.5281/zenodo.20370264
- Headline (honest, temporal): 0.775 ROC-AUC on 25-kinase prospective test (train ≤2018 / test ≥2020)
- Scaffold-split upper bound: 0.966 (same 25 targets); 0.926 across full registry of 398 kinase + GPCR models
- Decoy comparison: real-negative training beats decoy training on all 25 targets directionally (+0.031 scaffold); advantage shrinks to +0.014 under temporal split (marginal). Real-vs-decoy framing supported as consistent but modest, not a blanket claim of superiority.
- Y-randomization control performed (median 0.467, straddling chance)
- One split-validity control still un-run: AVE-debiasing
- PDF: https://nullary.ai/research/nullary-negative-results-validation-2026-05.pdf
- Per-target supplementary CSVs available

### Paper 2: Analytics suite

- Title: The Nullary Analytics Suite: Coverage, Tractability, and Cross-Modality Analysis over Measured Negative Results
- Date: 25 May 2026
- Headline: tractability score predicts clinical progression at AUC 0.785; externally validated at AUC 0.778 against Open Targets structural druggability (independent label, not derived from Nullary's bioactivity data)
- Bootstrap lift over popularity baseline: +0.036 [0.023, 0.048] on ChEMBL clinical label; +0.023 [0.005, 0.042] on independent structural label. Both CIs exclude zero. Lift is small but statistically reliable.
- Honest negative: cross-modality combination adds no predictive value (CRISPR × chemistry interactions non-significant; cross-validated AUC change +0.0003 [-0.0014, 0.0019]). We offer the cross-modality view, not a cross-modality predictor.
- CRISPR negatives reproduce 98% across DepMap and BioGRID-ORCS (F1 0.89 on essential class)
- PDF: https://nullary.ai/research/nullary-analytics-suite-2026-05.pdf

## Direct endpoints

- MCP server: https://mcp.nullary.ai/mcp
- REST API: https://api.nullary.ai
- Status page: https://status.nullary.ai
- Coverage data (live): https://api.nullary.ai/stats
- Full documentation (one file): https://nullary.ai/llms-full.txt

## Contact

team@nullary.ai