LinkD: An Agentic Platform for Drug Repurposing Unified across Molecular, Phenotypic, and Clinical Scales

LinkD unifies molecular, phenotypic, and clinical evidence for cancer drug discovery — binding affinity and selectivity, CRISPR drug-response, and EHR associations — behind an AI agent (LinkD-Agent) that plans and executes multi-step analyses from natural language.

License: MIT
Code: https://github.com/Huang-lab/LinkD
Submission release: v1.0-submission
Interactive: https://linkd-agent.net/
Data: Zenodo concept DOI 10.5281/zenodo.19241151

Live demo

URL: https://linkd-agent.net/

Manuscript figure reproduction (reviewers): For_Reviewer/ — see docs/FOR_REVIEWER.md. All panel inputs download as one checksum-verified Zenodo archive.

Modules

Module	Description	Key data
LinkD-Bind	Drug–target binding affinity explorer	1,068 targets with pKd, 20K+ binding pairs
LinkD-Select	Drug selectivity profiling (UMAP)	14,981 drugs, selectivity scores
LinkD-Pheno	Phenotype–drug associations from EHR	Mount Sinai + UK Biobank cancer associations
LinkD-Agent	Multi-step NL analysis agent	OpenAI, Google Gemini, or Anthropic APIs

Quick start

# 1. Environment
conda create -n ttdrug python=3.12
conda activate ttdrug
conda install nodejs

# 2. Dependencies
pip install -r requirements.txt
cd interactive_web_server/frontend && npm install && npm run build && cd ../..

# 3. Optional LLM keys for LinkD-Agent
cp .env.example .env
# Edit .env — GEMINI_FREE_KEY and/or OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEY

# 4. Launch
cd interactive_web_server && ./start.sh
# http://localhost:8000

Development mode: ./start.sh dev (FastAPI and Vite with hot reload).

Data

# Auto-download (also used on Render)
python scripts/download_data.py
python scripts/download_data.py --verify-only

# Or extract Zenodo archives into the project root:
# Database/, DrugTargetMetrics/, EHR_Results/, DrugResponse/, Target_Disease_Association/

Directory	Highlights
Database/	Oncogene annotations
Target_Disease_Association/	ChEMBL drug–target–disease; Open Targets causal links
DrugTargetMetrics/	Selectivity scores; per-target pKd parquet
EHR_Results/	Mount Sinai + UK Biobank aggregate associations
DrugResponse/	CRISPR drug-response correlations (PRISM + GDSC)

Versions: ChEMBL 34 · Mount Sinai / UKB EHR 2024-11 · PRISM/GDSC 2024-Q4 · Open Targets 24.09.

LinkD-Agent

LinkD-Agent decomposes biomedical questions into tool-using plans over the LinkD database (binding, selectivity, CRISPR, EHR, clinical-phase evidence) and returns structured multi-evidence summaries. Use it in the web UI (/agent) or via the Python package under agent/.

Weighted multi-evidence scoring lives in agent/evidence_scoring.py (weights in config/evidence_weights.yaml).

A JSON CLI over the same layers is available at .claude/skills/linkd/scripts/linkd:

.claude/skills/linkd/scripts/linkd target-info EGFR
.claude/skills/linkd/scripts/linkd evidence CHEMBL553 EGFR --disease "lung cancer" --icd C34 --drug-name Erlotinib

Manuscript figures & supplementary agent benchmark

Audience	Path
Reviewers regenerating figures	For_Reviewer/
Supplementary agent-eval harness (T1–T7; not a submitted Figure 6c panel)	benchmark/

Deployment (Render)

Frontend dist/ is committed; Render installs Python deps and serves the prebuilt bundle.

1. Connect the GitHub repo as a web service
2. Build: pip install -r requirements.txt
3. Start: python scripts/download_data.py && cd interactive_web_server/backend && python -m uvicorn main:app --host 0.0.0.0 --port $PORT
4. Env: DATABASE_DIR=/opt/render/project/src/data/Database, GEMINI_FREE_KEY, optional other LLM keys
5. Persistent disk (40 GB, allowing atomic dataset upgrades) at /opt/render/project/src/data

After frontend source changes, rebuild and commit interactive_web_server/frontend/dist/.

Data deposit staging: bash scripts/prepare_zenodo.sh and bash scripts/prepare_for_reviewer_zenodo.sh → upload a new version under concept DOI 10.5281/zenodo.19241151. Staging folder zenodo_upload/ is gitignored and never read by the web server.

Architecture

React frontend  →  FastAPI (/api/*)  →  agent/ (DB query + LLM planner)
                                      →  CSV/Parquet (Zenodo-hosted)

Project structure

LinkD/
├── agent/                     # Query module, evidence scoring, LLM planner
├── interactive_web_server/    # FastAPI + React (LinkD-Bind/Select/Pheno/Agent)
├── For_Reviewer/              # Manuscript figure reproduction package
├── benchmark/                 # Supplementary LinkD-Agent evaluation
├── scripts/download_data.py   # Zenodo download
├── config/                    # Evidence weights, etc.
├── requirements.txt
└── render.yaml

API (selected)

Endpoint	Description
GET /api/health	Health + loaded datasets
GET /api/overview	Database statistics
POST /api/binding/search	Binding landscape
POST /api/selectivity/search	Selectivity detail
GET /api/ehr/preload	EHR associations
POST /api/agent/plan	Stateless plan request: provider, model, optional transient key, query
POST /api/agent/execute	Stateless execution request: provider configuration + validated plan

Swagger UI: http://localhost:8000/docs

Configuration

Variable	Description
PORT	Server port (default 8000)
DATABASE_DIR	Path to Database/ (default ./Database)
GEMINI_FREE_KEY	Free-tier Gemini for LinkD-Agent
OPENAI_API_KEY / GOOGLE_API_KEY / ANTHROPIC_API_KEY	Optional LLM providers
LINKD_ALLOWED_ORIGINS	Comma-separated development CORS origins

License

Repository software is licensed under the MIT License. The figure-reproduction bundle is CC BY 4.0. Upstream application datasets retain their respective source terms; downloading them does not relicense them.

Research use only. LinkD associations are observational, predictions require independent validation, and LLM-generated plans or summaries may be incomplete or incorrect. LinkD does not provide clinical advice.

Contact

• Institution: Icahn School of Medicine at Mount Sinai
• Email: chengwangosu@gmail.com
• GitHub: github.com/Huang-lab/LinkD