GapForge

Open-source translational gap hunter for stalled or failed-but-promising drug programs.

Agents propose. Humans dispose. Not a molecule generator. Not clinical decision support.

Context of Use: GapForge assembles literature- and graph-backed gap hypotheses for scientific discussion. It is not for clinical care, prescribing, regulatory submission evidence, or synthesis planning.

---

Why GapForge

Most clinical candidates fail (~85–90%). AI has accelerated molecule discovery but has not fixed translation. GapForge focuses on the harder question:

This program looked promising and stalled — what are the plausible scientific gaps, and what should we check next?

We do	We do not
Evidence assembly across trials, literature, and target–disease graphs	Autonomous molecule design
Gap taxonomy (efficacy, biomarker, endpoint, PK, safety, …)	Patient treatment recommendations
Hypothesis cards with citations + adversarial critic	Auto-approving high-stakes claims
Mandatory human review (HITL) before “team conclusions”	Claiming regulatory-grade credibility without validation

Design details: docs/GAPFORGE.md · HITL: docs/HUMAN_IN_THE_LOOP.md

---

Risk tiers

Tier	Capability	Policy
L0	Read-only graph / program explore	Auto-allowed
L1	Summaries with citations	Auto + spot check
L2	Gap hypotheses / next experiments	HITL required
L3	Chemistry generation, dosing, patient advice	Blocked (v1)

---

Quick start

Prerequisites: Docker Desktop, Python 3.11+, Node.js 22+.

git clone https://github.com/LordKay-sudo/gapforge.git
cd gapforge
cp .env.example .env

docker compose up --build

Service	URL
Web UI	http://localhost:8080
GapForge programs	http://localhost:8080/programs
HITL review queue	http://localhost:8080/gaps/review
API docs	http://localhost:8000/docs
Neo4j Browser	http://localhost:7474 (neo4j / changeme)

Demo case studies

1. Flurizan (tarenflurbil) — Alzheimer Phase 3 efficacy miss (educational historical framing)
2. Astegolimab — COPD Phase 2b signal / Phase 3 miss (biomarker vs all-comers)

Workflow:

1. Open GapForge → pick a program
2. Inspect trials, taxonomy, and L2 hypothesis cards
3. Open Review → run Discern / OntoHarness → approve / reject / request more
4. Export review bundle — only approved cards become team conclusions

Seed data: data/gapforge/ (all *.json cases load via scripts/seed_gapforge.py)
Discern: docs/DISCERN.md

OntoHarness semantic gate (optional)

GapForge projects each L2 hypothesis (Neo4j records → Turtle) and validates via OntoHarness before propose/approve. Failures block HITL approval and show repair hints in the review UI (same panel pattern as Discern).

# Stack + OntoHarness sidecar (strict: fail closed if sidecar unreachable)
docker compose -f docker-compose.yml -f docker-compose.ontoharness.yml up --build

Local API without Docker: set ONTOHARNESS_ENABLED=true in .env and run OntoHarness on :8010.

---

Architecture

flowchart TB
  subgraph humans [Human ownership]
    Reviewer[Scientist reviewer]
    Gate[HITL approval gate]
  end
  subgraph agents [Propose only]
    Planner[plan_gap_investigation]
    Critic[run_critic]
  end
  subgraph core [GapForge core]
    API[FastAPI /api/v1]
    UI[React explorer]
    N4j[(Neo4j)]
  end
  Reviewer --> Gate
  Planner --> API
  Critic --> API
  Critic --> Gate
  UI --> API --> N4j

Loading

Built on a disease–target knowledge graph (Open Targets–style associations) with GapForge program/trial/hypothesis nodes and PROV-style SUPPORTED_BY / CONTRADICTED_BY edges.

---

Ecosystem

GapForge is the product name for this repo. Related open components:

Repository	Role
gapforge (this repo)	Graph + GapForge API/UI + Flurizan case study
kg-rag-demo	Citation-grounded literature / ClinicalTrials RAG
ontoharness	SHACL + vocab gate sidecar (L2 semantic validation)
embabel-mcp	MCP tools + research_program_gaps agent
peerlens	Paper quality signals (retraction / concern filter)
bioinsight-graph	Upstream disease–target graph lineage

Docs: docs/ECOSYSTEM.md · MCP tools: plan_gap_investigation, build_program_dossier, propose_gap_hypotheses, run_critic, export_review_bundle.

---

API (GapForge)

Method	Path	Notes
GET	/api/v1/programs	List stalled programs
GET	/api/v1/programs/{id}/dossier	L1 dossier
GET	/api/v1/gaps	Hypothesis cards
POST	/api/v1/gaps/{id}/critic	Adversarial critic
POST	/api/v1/discern	Universal I/O weigh (compliance / reliability / …)
GET	/api/v1/discern/policy	Active thresholds by risk tier
POST	/api/v1/gaps/{id}/ontology-validate	Re-run OntoHarness SHACL + vocab gate
GET	/api/v1/reviews/queue	HITL queue
POST	/api/v1/reviews/{gap_id}	approve / reject / request_more
GET	/api/v1/export/review-bundle	Provenance export

---

Safety posture

Aligned with FDA/EMA Guiding Principles of Good AI Practice in Drug Development: human-centric design, clear context of use, risk-based controls, data provenance, and reviewable outputs.

• Associations are correlative, not causal
• Predictions (if added later) are labeled model_estimate and never auto-written as facts
• Literature informs RAG; only curated ETL writes the main Neo4j fact graph

See PROVENANCE.md.

---

Development

# API tests
cd api && py -3 -m venv .venv
.\.venv\Scripts\pip install -r requirements.txt pytest
.\.venv\Scripts\python -m pytest -q

# Web
cd web && npm install && npm run build

Docs: GETTING_STARTED.md · ARCHITECTURE.md · ROADMAP.md

---

License

MIT © 2026 LordKay-sudo

Upstream Open Targets–style data: see licence notes in PROVENANCE.md (typically CC0 for platform data — always verify for your release).