Skip to content

GAPFORGE.md

Latest commit

 

History

History
199 lines (144 loc) · 7.58 KB

GAPFORGE.md

File metadata and controls

199 lines (144 loc) · 7.58 KB
Raw

BioInsight GapForge

Translational gap hunter for stalled or failed-but-promising drug programs. Agents propose; humans dispose. Not a molecule generator. Not clinical decision support.

Field Value
Status MVP (educational case studies)
COU Generate literature- and graph-backed gap hypotheses for scientific discussion
Non-use Clinical care, prescribing, regulatory submission evidence, synthesis planning

Context of Use (COU)

GapForge helps translational scientists assemble evidence-backed hypotheses about why a public development program may have stalled (efficacy, safety, PK, formulation, biomarker, endpoint, competitive), with citations and a mandatory human review step before any card is treated as a team conclusion.

Every agent run and API write for L2 outputs must carry this COU (see cou on hypothesis records and MCP tool responses).

Risk tiers

Tier Capability Policy
L0 Read-only graph / program explore Auto-allowed
L1 Summaries with citations Auto + post-hoc spot check
L2 Gap hypotheses / ranked next experiments HITL required — status stays needs_review until approve/reject
L3 Chemistry generation, dose advice, patient-level recommendations Blocked in v1

Gap taxonomy

Code Class Typical signals
efficacy Lack of clinical efficacy Phase 2/3 primary endpoint miss
safety Safety / toxicity AE rates, FAERS clusters, discontinuations
pk_exposure PK / bioavailability / exposure Dose–exposure inadequacy
formulation Formulation / delivery Stability, route, adherence
biomarker Stratification / predictive biomarker All-comers vs enriched; Phase 2→3 divergence
endpoint Endpoint / surrogate mismatch Mechanistic biomarker ≠ clinical outcome
target_validity Wrong or weak biological hypothesis Poor genetic support, species translation
competitive SoC / portfolio (non-scientific) Marked as extrinsic; still documented

Dual-channel evidence rule

An L2 hypothesis must include either:

  1. At least one structured fact (graph edge, trial registry record, curated program field), and at least one citable literature/trial passage; or
  2. Explicit insufficient_evidence: true with empty or weak bundles (honest failure mode).

Silent promotion of association scores to causal claims is forbidden.

Discern (universal I/O weighing)

POST /api/v1/discern and MCP discern_artifact score artifacts on compliance, reliability, provenance, and safety_language against tier thresholds (gapforge-discern-v1).

  • hard_fail → block (e.g. dosing language, L3)
  • soft_fail / L2 pass → require_hitl (never auto-approve team conclusions)

Details: DISCERN.md.

Neo4j model (GapForge)

(:Drug {id, name, synonyms, chembl_id?})
(:Program {id, name, status, indication_name, moa, cou_note})
(:Trial {id, nct_id, phase, status, primary_endpoint, outcome_summary})
(:GapHypothesis {
  id, gap_class, claim, confidence,
  suggested_experiment, status,  // draft|needs_review|approved|rejected
  insufficient_evidence, provenance_hash, risk_tier, cou
})
(:Review {id, decision, reviewer, notes, decided_at})

(:Program)-[:INVESTIGATES]->(:Drug)
(:Program)-[:FOR_INDICATION]->(:Disease)
(:Program)-[:TARGETS]->(:Gene)
(:Program)-[:TESTED_IN]->(:Trial)
(:Trial)-[:FOR_INDICATION]->(:Disease)
(:GapHypothesis)-[:ABOUT]->(:Program)
(:GapHypothesis)-[:SUPPORTED_BY]->(:Trial|:Gene|:Disease|:DocumentRef)
(:GapHypothesis)-[:CONTRADICTED_BY]->(:Trial|:Gene|:Disease|:DocumentRef)
(:GapHypothesis)-[:DERIVED_FROM]->(:Trial|:Program)
(:Review)-[:REVIEWS]->(:GapHypothesis)

Main target–disease associations remain curated Open Targets ingest only. Literature informs RAG / citations; it does not silently write new association edges.

API surface (MVP)

Method Path Tier
GET /api/v1/programs L0
GET /api/v1/programs/{id} L0
GET /api/v1/programs/{id}/dossier L1
GET /api/v1/programs/{id}/taxonomy L1
GET /api/v1/gaps L0
GET /api/v1/gaps/{id} L0
POST /api/v1/gaps/propose L2 create → needs_review
POST /api/v1/gaps/{id}/critic L2 critic notes
GET /api/v1/reviews/queue HITL
POST /api/v1/reviews/{gap_id} approve / reject / request_more
GET /api/v1/export/review-bundle?gap_id= Provenance export

MCP tools (embabel-mcp)

Tool Purpose
plan_gap_investigation Intent + tool sequence + stop rules + COU
build_program_dossier Structured program + trials + linked genes/diseases
propose_gap_hypotheses Create/list L2 cards (HITL pending)
run_critic Adversarial pass: counter-evidence + confidence clamp
export_review_bundle Audit JSON: meta, dossier, cards, reviews, hashes

Default MCP mode: propose-only. UI review queue is ground truth (see HUMAN_IN_THE_LOOP.md).

Seeded case studies (MVP)

Flurizan (tarenflurbil) — Alzheimer's disease

Field Value
Program id prog-flurizan-ad
Public framing Historical educational case study — not a recommendation to resurrect the asset
Disease Alzheimer disease (MONDO_0004975)
Linked gene (graph) APOE (ENSG00000130203) — stratification / AD genetics context
Exemplar trial NCT00105547-style Phase 3 efficacy miss (curated summary)
Gap themes endpoint, biomarker, efficacy / target engagement

Seed: data/gapforge/flurizan_case.json

Astegolimab — COPD (Phase 2b signal / Phase 3 miss)

Field Value
Program id prog-astegolimab-copd
Public framing Educational case on biomarker-agnostic enrollment and Phase 2→3 divergence
Indication Chronic obstructive pulmonary disease (EFO_0000341 when present in graph)
Exemplar trials ALIENTO (Phase 2b met) / ARNASA (Phase 3 miss) — curated summaries
Gap themes biomarker, endpoint, efficacy

Seed: data/gapforge/astegolimab_case.json

All data/gapforge/*.json files load via scripts/seed_gapforge.py. See HUMAN_IN_THE_LOOP.md.

Explicit non-goals (v1)

  • De novo molecule design or synthesis planning (ChemCrow-class tools deferred)
  • Clinical decision support or dosing advice
  • Regulatory-grade credibility claims without wet-lab / clinical validation
  • Replacing Open Targets, ChEMBL, ClinicalTrials.gov, or FDA systems
  • Auto-approving L2 hypotheses

Compliance posture

Aligned with FDA/EMA Guiding Principles of Good AI Practice in Drug Development:

  1. Human-centric by design (HITL for L2)
  2. Risk-based approach (tiers L0–L3)
  3. Clear context of use
  4. Data governance and provenance (PROVENANCE.md, /meta, review bundles)
  5. Life-cycle documentation (this file + schema migration notes)

Implementation map

Layer Location
Design (this doc) docs/GAPFORGE.md
Constraints scripts/neo4j/init.cypher
Seed scripts/seed_gapforge.py + data/gapforge/
API api/app/routers/programs.py, gaps.py, reviews.py, discern.py
UI /programs, /program/:id, /gaps/review
MCP BioInsightGapMcpTools.java (run_gap_discern, …)
Literature kg-rag ClinicalTrials + Europe PMC corpora; PeerLens pre-filter

Update when adding case studies or changing risk-tier policy.