Generate falsification-tested drug-repurposing hypotheses across corpora

Point the Hypothesis Crucible plugin at a disease and get back a short list of repurposing candidates that no single source proposes, each assembled from fragments spanning multiple bodies of evidence, screened by a falsification gauntlet, and delivered with the citations, the proof that the connection is novel, and the experiment that would kill it.


Problem class	Hypothesis generation
Subject areas	Drug Repurposing and Discovery, Neuroscience, Molecular and Cellular Biology
Evidence level	Proposed
Complexity	Autonomous system
Availability	Fully open
Compute	Laptop

Problem

The hard part of computational hypothesis generation is not producing ideas — an LLM will produce a plausible-sounding drug–disease link for almost any pair — but producing ideas that are simultaneously novel (not already in the literature), grounded (every mechanistic step traceable to real evidence), and not already falsified (no failed trial, no contradicting data, no pharmacological showstopper). Naïve prompting fails all three: it confabulates mechanisms, re-proposes approved indications as if new, and ignores the Phase III graveyard. What you want instead: hand over a disease, get back a handful of candidates where each one arrives with a mechanistic chain whose every link cites a real fragment (a PMID, a GEO series, a ChEMBL mechanism, an Open Targets association), a record of the searches that confirm no one has published the direct claim, the disconfirming evidence that was sought and did not sink it, and a discriminating experiment — plus a full log of everything that was rejected and why.

Recommended approach

Install the Hypothesis Crucible plugin and connect the bio-research MCP servers it drives:
```
/plugin marketplace add scripps-ai-enablement/sci-ai-enabler
/plugin install crucible
```
Connect pubmed, ot, chembl, c-trials, biorxiv, consensus, biomcp, and tooluniverse via /mcp; GEO is reached through the gget skill / tooluniverse.
Run the forge command with your goal:
```
/crucible:forge "repurposable approved drugs for Alzheimer's disease"
```
The skill frames the goal to an EFO/MONDO id, mines typed fragments across the three source tiers, discovers Swanson-style A–C bridges (no direct edge, ≥2 independent B-paths across ≥2 corpora), and runs each candidate through the falsification gauntlet.
Read the kill tally, not just the survivors. The run streams how many candidates each gate removed — G1 novelty (already published), G2 groundedness (a hallucinated mechanistic step), G3 contradiction (a failed trial or opposing signature), G4 plausibility (e.g. no blood–brain-barrier penetration). A healthy run kills most of what it considers; that aggressive rejection is the point, and the kill-log.jsonl is the audit trail.
Take delivery of the run bundle. The durable artifact is a directory (./crucible-runs/<goal>-<date>/) holding hypotheses.json (survivors, each with a fragment-cited chain + novelty evidence + falsification record + a discriminating test), kill-log.jsonl, fragments.jsonl, provenance.json, and an IEEE-2791 run.bco.json. This is what you commit and share — not the chat.
Hand a survivor downstream. Pass a candidate to the composer to design its wet-lab experiment, or export the top candidates as the input to an experimental validation loop. Crucible is a high-precision filter that sits upstream of validation.
For a time-sliced sanity check, run the evaluation. The worked example ships a deterministic scorer and gold fixtures (eval/score.py, eval/gold/) that measure rediscovery recall, planted-negative kill rate, per-gate attribution, and groundedness against a captured run — the protocol for deciding whether the gauntlet actually earns its place before trusting a run.

Why this assembly

Rung 4 of the simplicity ladder, and it earns it. Rung 1 (plain Claude) confabulates mechanisms and novelty; rung 2 (a single hypothesis-generation skill) produces ungrounded ideas with no rejection stage; rung 3 (a toolbelt of literature + database MCPs) gets you the fragments but not the bridge discovery, the aggressive falsification gauntlet, or the provenance-bound capture. The distinctive work here — Swanson-style cross-corpus bridging, four independent veto gates that log every kill, and an auditable run bundle — is an orchestrated multi-agent loop, which is what rung 4 is for. It is deliberately narrower than an end-to-end autonomous scientist: it generates and filters candidates rather than validating them experimentally.

Availability

Fully open. The plugin ships in this repository’s marketplace; the MCP servers it drives are the open bio-research set (PubMed, Open Targets, ChEMBL, ClinicalTrials.gov, bioRxiv, Consensus, BioMCP, ToolUniverse). No subscription. GEO access is via the OSS gget library / public NCBI endpoints. Some literature full text may sit behind publisher access, in which case the skill falls back to abstracts.

Compute requirements

Laptop. The workflow is API/MCP calls and model reasoning; no GPU, no large downloads. Wall-clock is dominated by the number of candidates run through the gauntlet and any tournament rounds, typically minutes to tens of minutes for a single disease goal.

Evidence

Proposed. This is a freshly composed assembly, screened but not yet validated by a surfaced hypothesis being confirmed in the lab. The design draws on published patterns: Swanson’s literature-based discovery / ABC model of undiscovered public knowledge; the generation–reflection–ranking–evolution tournament of Google DeepMind’s AI co-scientist (Gottweis et al., 2025, arXiv:2502.18864); provenance-bound literature-grounded generation (HypER, arXiv:2506.12937); groundedness/hallucination filtering (TruthHypo/KnowHD, arXiv:2505.14599); and novelty-optimized retrieval (SciMON, arXiv:2305.14259). Its differentiator is the falsification-first control structure — optimizing precision by aggressive rejection with logged kills — which the evaluation harness in the worked example is built to test.

Alternatives considered

Co-Scientist / Robin / OpenScientist directly. These are broader end-to-end systems (Co-Scientist ranks proposals; Robin closes a wet-lab loop; OpenScientist runs computational experiments). Reach for them when you want a full discovery loop rather than a high-precision candidate filter, or when you can execute experiments. Crucible is complementary — it can feed candidates into any of them.
Hypothesis Generation skill (K-Dense). A single skill for structured hypothesis formulation from observations. Lighter (rung 2) and good for ideation, but it has no cross-corpus fragment mining, no Swanson bridge discovery, and no falsification gauntlet — it will not reject aggressively or emit provenance.
Plain Claude + literature/database MCPs (rung 3 toolbelt). Gets you fragments and manual reasoning; use it for a one-off look. You lose the systematic bridge search, the four-gate rejection, and the reproducible run bundle.

Sources

Gottweis J. et al., “Towards an AI co-scientist,” arXiv:2502.18864, 2025 — the tournament/evolution pattern reused in Stage 4.
HypER: Literature-grounded Hypothesis Generation and Distillation with Provenance, arXiv:2506.12937, 2025 — provenance-bound generation.
Toward Reliable Scientific Hypothesis Generation (TruthHypo / KnowHD), arXiv:2505.14599, 2025 — groundedness filtering that G2 adapts.
Wang Q. et al., “SciMON: Scientific Inspiration Machines Optimized for Novelty,” arXiv:2305.14259, 2023 — novelty optimization.
Swanson D.R., “Undiscovered public knowledge,” Library Quarterly 56(2):103–118, 1986 — the ABC model behind Stage 2.

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.