Agentic Target Evidence

Agentic Target Evidence

Testing
Containers
Tooling
Stack
Meta

A multi-agent system that gathers and interprets evidence on whether a gene is a viable drug target for a disease. Given a (gene, disease, direction) triple — e.g. BRCA1, breast cancer, inhibit — it retrieves evidence from ~two dozen biomedical sources, screens and interprets it through six independent lenses (genetics, biology, safety, clinical, commercial, regulatory), and produces a provenanced dossier: a consensus verdict, a single 0–100 suitability score, per-lens narratives, and a categorized, link-rich evidence list.

Every source connector — DepMap, gnomAD, ClinicalTrials.gov, OpenTargets, PubMed, FAERS, and ~20 more — lives once under src/mcp_servers/ and is consumed two ways: in-process by the pipeline's agents (fast, typed, no protocol tax), or through the MCP gateway, which composes the same connectors into one MCP server exposing ~40 read-only tools to any MCP host — Claude Desktop, Claude Code, your own agent — for ad hoc lookups outside a full run. A bundled chat assistant offers the same lookups from a browser. See § MCP gateway & servers below.

Built on LangGraph (orchestration) + MCP (the data layer), with full tracing (Langfuse + OpenTelemetry), Postgres-backed checkpointing, and configurable local/cloud LLM routing.

Every verdict is LLM-generated over retrieved evidence — a preliminary research aid, not ground truth. It is built to accelerate the evidence-gathering phase of target validation, not to replace expert review. See NOTICE.md for the full disclaimer, licenses, and data notices.

📄 See it in action: Example dossier — TRPC6 in Focal Segmental Glomerulosclerosis. A real end-to-end run: consensus verdict, 0–100 suitability score, six per-lens narratives, and a link-rich evidence list over 135 kept sources.

---

Quickstart

uv sync                  # Python ≥ 3.12, via uv: https://docs.astral.sh/uv/
cp .env.example .env      # fill in any keys you want; most sources are keyless

make up                   # infra + Langfuse + OTEL + the app, as containers
make run GENE=BRCA1 DISEASE="breast cancer"

Output lands under results/report/{gene}/{disease}/{direction}/report.md. The Langfuse trace UI is at http://localhost:3000. Windows: use make.bat instead of make — see docs/tutorial.md.

Don't want a full run? Ask one-off questions against the same connectors (e.g. "What's TRPC6's DepMap dependency score?") via the bundled chat UI or Claude Desktop/Code — see docs/mcp_tutorial.md.

Pre-built images

make up builds all service images locally. Every tagged release also publishes the same images to GHCR, so you can pull instead of building:

docker pull ghcr.io/athril/agentic-target-evidence/mcp-servers:latest
docker pull ghcr.io/athril/agentic-target-evidence/mcp-gateway:latest
docker pull ghcr.io/athril/agentic-target-evidence/agents-knowledge:latest
docker pull ghcr.io/athril/agentic-target-evidence/agents-reasoning:latest
docker pull ghcr.io/athril/agentic-target-evidence/report-agent:latest
docker pull ghcr.io/athril/agentic-target-evidence/planner:latest
docker pull ghcr.io/athril/agentic-target-evidence/chat:latest

latest tracks the most recent release; pin a version instead (e.g. :v0.1.2) for reproducibility. To use these instead of a local build, replace a service's build: block in docker-compose.yml with image: ghcr.io/athril/agentic-target-evidence/<target>:<tag>.

---

MCP gateway & servers

Every biomedical source connector lives under src/mcp_servers/ as a self-contained tools.py + MCP server.py pair — 27 source connectors, ~46 read-only tools, spanning 30+ named public sources (some connector folders bundle more than one upstream API — see docs/data_sources.md) plus your own internal data:

ChEMBL · ClinGen · ClinicalTrials.gov · ClinVar · DepMap · DGIdb · ENCODE · Expression Atlas · GBD (IHME) · GenCC · gnomAD · Google Patents · GTEx · GWAS Catalog · HGNC · HPA · IMPC · Monarch Initiative · MONDO · OMIM · OpenAlex · OpenFDA · OpenTargets · Orphanet · Project Score · PubMed · SCImago (SJR) · SPOKE · TTD · UniProt · USPTO · internal data (your org's private tables)

Full per-source details (what each provides, licensing/gating status) in docs/data_sources.md. The MCP gateway (src/mcp_gateway/server.py) dynamically discovers and composes all of them into one MCP server, with no hand-maintained registry — drop a new src/mcp_servers/<name>/server.py in and it's mounted automatically (subject to feature gates; internal_data is never mounted).

Three ways to reach it, without running the full pipeline:

Client	What it is
Chat assistant	A Gradio chat UI backed by a local Ollama model — make chat or see docs/mcp_tutorial.md.
Claude Desktop / Claude Code	Connect over stdio alongside your other MCP servers.
Any other MCP client	Call the tools programmatically over HTTP (bearer-token auth via MCP_GATEWAY_TOKEN).

Self-hosting the gateway alone (no pipeline, no other services) is a single container: docker run -p 8765:8765 ghcr.io/athril/agentic-target-evidence/mcp-gateway:latest — it defaults to HTTP on 0.0.0.0:8765. Point any MCP client at http://<host>:8765/mcp.

To use it from Claude Desktop or Claude Code without cloning the repo, have the client launch the same image over stdio — add this to claude_desktop_config.json (or .mcp.json):

{
  "mcpServers": {
    "agentic-target-evidence": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT=stdio",
               "-e", "NCBI_API_KEY", "-e", "USPTO_API_KEY",
               "ghcr.io/athril/agentic-target-evidence/mcp-gateway:latest"]
    }
  }
}

Both keys are optional (-e NAME with no value forwards it from your environment). The gateway is also listed in the official MCP Registry as io.github.athril/agentic-target-evidence.

For all bulk retrieval the pipeline never talks to the gateway — it imports each tools.py directly, keeping the hot path free of protocol overhead. The gateway is a second, additive surface onto the same connectors, for ad hoc use outside a full run. (The one in-pipeline gateway client is the synthesis-phase Investigator agent, which calls retrieval tools over MCP to close evidence gaps before the report; it degrades gracefully if the gateway is down.) For the design — exposure model, security, transports, discovery internals — see docs/mcp_gateway.md; for a step-by-step walkthrough, see docs/mcp_tutorial.md.

---

Documentation

Full documentation lives in docs/. Start there — it has reading paths for "I just want to run it," "I want to understand the design," and "I want to contribute." A few entry points:

Doc	What it covers
docs/README.md	Start here. Doc set index, full table of documents, and reading paths by goal (run it, understand the design, contribute).
docs/tutorial.md	Run an analysis and read the resulting dossier.
docs/mcp_tutorial.md	Ad hoc tool access via the MCP gateway and chat assistant, step by step.
docs/mcp_gateway.md	MCP gateway reference: exposure model, security, transports, discovery internals.
docs/data_sources.md	Every source connector, what it provides, and its licensing/gating status.
docs/restricted.md	Step-by-step setup for gated sources (OMIM, SCImago, GBD, TTD): API keys, data downloads, verification.
docs/architecture.md	The full design: pipeline graph, HITL, capabilities, observability.
docs/developers.md	Extension points and conventions for contributing.
docs/faq.md	Nuances and easy-to-get-wrong points, as Q&A.

---

Contributing

Contributions are welcome — see CONTRIBUTING.md for setup, conventions, and how to submit a change, and docs/developers.md for extension points. Please also read our Code of Conduct.

---

License

Apache-2.0. Some data sources carry narrower terms and are gated off by default — see docs/data_sources.md for the reference table, docs/restricted.md for step-by-step setup, and NOTICE.md for the full disclaimer and per-source licenses.

---

Contact

Patryk Orzechowski, Ph.D.