tox-antitargets-mcp-server

An MCP server that reproduces the results of Nikitin et al., "Towards Explainable Computational Toxicology: Linking Antitargets to Rodent Acute Toxicity" (Pharmaceutics 2025, 17, 1573) as callable tools. Every figure, statistic and conclusion is computed deterministically from the public ld50-antitargets dataset (12 654 ligands × 44 antitarget docking scores + mouse-intravenous pLD50). The dataset is bundled, so the server runs offline on a CPU — no GPU and no docking step required.

Run

docker compose up -d --build                      # -> http://localhost:7335/mcp
# or, without Docker:
uv sync && uv run python -m server.tox_server     # -> http://localhost:7331/mcp
uv run pytest tests                               # 11/11 reproduction checks

No configuration needed — it works out of the box.

What it does

16 MCP tools covering the whole paper:

• Reproduce everything — reproduce_all (headline numbers vs the paper) and reproduce_claims (the 11 conclusions, each restated with the recomputed numbers).
• Per-figure analyses — antitarget→LD50 ranking, binder/non-binder Mann–Whitney test, NIH/Brenk filtering, Spearman correlations, Butina clustering, physicochemical profiles, t-SNE, the logP-confounder check.
• Interactive — inverse_docking_profile (antitarget profile / target fishing for a molecule by SMILES or name) and protein_panel (the 44 Bowes-panel targets).

Validated against the paper: identical dataset shape and pLD50 range; top-5 antitargets KCNH2, AVPR1A, CACNA1C, KCNQ1, EDNRA (exact); binders significantly more toxic than non-binders (p ≈ 5·10⁻¹³², median gap 0.38 → 0.70 after filtering). Full table and the few version-related deviations are in docs/ARCHITECTURE.md.

Use with CoScientist

CoScientist discovers MCP tools via RAG. With its RAG stack running, register once — then the agent finds and calls the tools for any toxicity / LD50 / mechanism-of-action query:

python scripts/rag_tools/cli.py load rag_registration.json

Docs

• docs/QUESTIONS.md — what the paper answers, and the exact questions to ask the agent.
• docs/ARCHITECTURE.md — how it is built: modules, tools, parameters, fidelity.

reproduce_paper.py runs the numbers → LLM → conclusions loop with an OpenRouter key, without the full CoScientist stack. Optional TOX_* env vars (port, thresholds, S3 figure storage) are listed in the architecture doc.

Cite

@article{Nikitin2025,
  author  = {Ilia Nikitin and Igor Morgunov and Victor Safronov and Anna Kalyuzhnaya and Maxim Fedorov},
  title   = {Towards Explainable Computational Toxicology: Linking Antitargets to Rodent Acute Toxicity},
  journal = {Pharmaceutics},
  year    = {2025},
  volume  = {17},
  pages   = {1573},
  doi     = {10.3390/pharmaceutics17121573}
}

License

MIT (code; see LICENSE). Data and methods belong to Nikitin et al. 2025; dataset from chemagents/ld50-antitargets.