Demo

▶ Watch the agent run a full DMTA campaign

DMTA Agent

An autonomous drug discovery agent that closes the Design → Make → Test → Analyze loop using agentic AI and physical lab instrument integration.

Claude reasons over experimental data to propose compounds, a Hamilton STAR liquid handler dispenses them, an assay station measures potency and ADMET properties, and the agent analyzes results to decide next steps — pausing between iterations for scientist review.

Built on FastAPI, Pydantic, and the Anthropic SDK. Docker and Kubernetes ready.

Architecture

┌─────────────────────────────────────────────┐
│                 FastAPI Service              │
│  POST /campaigns   GET /campaigns/{id}/run  │
└──────────────────────┬──────────────────────┘
                       │ SSE stream
              ┌────────▼────────┐
              │   DMTAAgent     │  ← orchestrates the loop
              └──┬──────────┬───┘
                 │          │
        ┌────────▼──┐  ┌────▼──────────┐
        │  Claude   │  │  Instruments  │
        │  (LLM)    │  │  Layer        │
        └───────────┘  └──┬────────┬───┘
           Design +       │        │
           Analyze     Hamilton  AssayStation
                        STAR     (TR-FRET +
                       (Make)     ADMET)
                                  (Test)

Key design decisions

Decision	Rationale
SSE streaming	Client sees each phase event in real-time; no polling needed
Instrument abstraction layer	Swap mock → real SDK without touching agent logic
Pydantic throughout	Every LLM response and instrument payload is validated
Parallel dispensing	asyncio.gather() fires all Hamilton jobs concurrently
Stateless agent	Campaign state lives in Redis (or in-memory for dev); agent is pure function

---

Prerequisites

• Python 3.12+
• Anthropic API key — get one at console.anthropic.com or ask me for a live demo.

Quickstart

# 1. Install
pip install -r requirements.txt

# 2. Set API key
export ANTHROPIC_API_KEY=sk-ant-...

# 3. Run
uvicorn main:app --reload --port 8000

# Start a campaign
curl -X POST http://localhost:8000/campaigns \
  -H "Content-Type: application/json" \
  -d '{"goal": "JAK2-selective inhibitor: IC50 < 1 nM, >100x selectivity vs JAK1"}'

# Stream agent events (replace {id} with campaign_id from above)
curl -N http://localhost:8000/campaigns/{id}/run

---

Docker

docker build -t dmta-agent .
docker run -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY -p 8000:8000 dmta-agent

---

Connecting real instruments

In agent/instruments.py, replace the _dispense_real and _real_panel methods:

# Hamilton VENUS SDK
async def _dispense_real(self, job: DispenseJob) -> DispenseResult:
    import venus
    result = await venus.execute_method("Dispense", job.model_dump())
    return DispenseResult(**result)

# LIMS (e.g. Benchling, Dotmatics)
async def _real_panel(self, compounds) -> list[AssayResult]:
    async with httpx.AsyncClient() as c:
        r = await c.post(f"{self.lims_url}/assay/jak2", json=[...])
    return [AssayResult(**x) for x in r.json()]

The agent loop in agent/dmta_agent.py does not change.

---

Production checklist

• Replace in-memory campaigns dict with Redis
• Add JWT auth middleware
• Wire real instrument SDKs in instruments.py
• Set mock_instruments=False in main.py
• Deploy as Kubernetes Deployment + Service
• Add Prometheus metrics on /metrics