Nullary MCP Server

Negative results intelligence for drug discovery — over the Model Context Protocol.

Nullary is a hosted MCP server that lets AI agents query measured negative results from drug discovery: inactive compounds, failed selectivity panels, terminated clinical trials, failed CRISPR screens, antibody developability failures, and more — each result carrying full provenance (source database, DOI/PMID, license).

This repository documents the public MCP interface. The server is hosted; the data pipeline and application code are maintained separately.

Connect

Nullary is a remote MCP server (streamable-HTTP) — nothing to install, no API key:

https://mcp.nullary.ai/mcp

Claude Code / generic MCP client

{
  "mcpServers": {
    "nullary": {
      "url": "https://mcp.nullary.ai/mcp"
    }
  }
}

Cursor

Settings → MCP → Add new MCP server → paste the URL above (transport: streamable-HTTP).

Claude Desktop

Settings → Connectors → Add custom connector → URL https://mcp.nullary.ai/mcp.

What you can ask

"What's failed against EGFR?" — failed compounds, trials, and screens for a target, across modalities
"Which kinase inhibitors were inactive in ChEMBL?"
"Show terminated Phase 2 oncology trials and why they stopped"
"Failed CRISPR knockouts for TP53"

Tools are organized by modality (small molecule, CRISPR, antibody, peptide, PROTAC, clinical trial, …); every response cites its source.

Tools

The server exposes 35 tools — served live via the MCP tools/list method. Full JSON Schemas (inputs per tool) are in tools.json. Tools span the seven modalities plus cross-modality history, compound/provenance lookups, and the Layer-1 model registry.

Tool	Description
`search_inactive_compounds`	Inactive small-molecule compound-target pairs.
`search_failed_selectivity`	Small molecules that failed selectivity.
`search_admet_failures`	Small-molecule ADMET failures.
`search_failed_guides`	Failed/ineffective CRISPR guides.
`search_failed_essentiality_screens`	Non-dependency / failed essentiality screens.
`search_ancestry_specific_failures`	Ancestry-specific CRISPR failures.
`search_developability_failures`	Antibody developability failures.
`search_failed_clinical_antibodies`	Discontinued/terminated clinical antibodies.
`search_failed_peptide_therapeutics`	Failed peptide therapeutics.
`search_peptide_stability_issues`	Peptide stability/half-life failures.
`search_failed_protacs`	PROTACs that failed degradation/ternary/permeability.
`search_protac_e3_issues`	PROTAC E3-ligase recruitment / ternary failures.
`search_failed_oligonucleotides`	ASOs/siRNAs that failed engagement/developability.
`search_oligo_delivery_failures`	Oligonucleotide delivery failures.
`search_failed_vaccines`	Failed/terminated vaccines (by pathogen/indication).
`search_vaccine_immunogenicity_failures`	Failed vaccine immunogen designs.
`search_failed_adcs`	ADCs that failed at any stage.
`search_adc_linker_failures`	ADC failures attributed to linker chemistry.
`search_failed_bispecifics`	Bispecifics that failed at any stage.
`search_bispecific_format_failures`	Bispecific format/engineering failures.
`search_admet_failures_all_modalities`	ADMET failures across ALL modalities.
`search_drug_drug_interaction_failures`	Drug-drug interaction failures.
`search_mechanism_failures`	Approaches that failed for a mechanism (by target).
`search_failed_replications`	Findings that failed to replicate.
`search_safety_failures`	Clinical/preclinical safety failures across modalities.
`search_target_history`	ALL failed approaches against a target across every modality.
`search_indication_history`	ALL failed approaches for an indication across every modality.
`search_pathogen_history`	Vaccine + antimicrobial + antibody failures for a pathogen.
`get_compound`	A compound + its full negative profile across modalities/sources.
`get_finding_provenance`	Full provenance + detail for a single finding by id.
`get_target_landscape`	Target "graveyard" / exhaustion index — how picked-over a target is, by modality and outcome.
`list_top_targets`	The most heavily-pursued targets, ranked by recorded negative findings.
`list_models`	Summary of the Layer-1 inactivity-scoring model registry.
`get_model_card`	Per-target Layer-1 model card: training counts + held-out scaffold-split metrics.
`get_coverage`	Per-modality and per-source coverage stats.

License

This documentation repository is MIT-licensed. The underlying data is provided under each source's respective license — see the coverage page.

negative-results-mcp