orb-antibiotic-scientist

An autonomous AI scientist that designs novel antibiotic candidates for drug-resistant bacteria. Runs 24/7 on Orb Cloud with Claude Opus via the Agent SDK. Sleeps between LLM calls. Wakes when results arrive.

Antibiotic resistance kills ~1.27 million people every year. Big pharma left the space in the 2010s — the economics don't work at their scale. We need R&D at 100× lower cost.

What it does

The agent runs a continuous design-and-validate loop, 24 hours a day:

1. Designs a novel candidate — small molecule (RDKit scaffold hopping) or antimicrobial peptide (ESM-2).
2. Validates the candidate through 12 orthogonal computational filters (see Verification).
3. Scores rigour + binding + ADMET + novelty into a composite leaderboard rank.
4. Documents the candidate with SMILES / FASTA, 3D pose, mechanism hypothesis, and literature citations.
5. Critiques itself — a separate adversarial Claude call tries to find reasons the candidate will fail.
6. Sleeps until the next docking job or LLM call returns.

Initial target: MRSA DNA gyrase B (GyrB), PDB 4URL. Target #2 (planned): M. tuberculosis InhA, PDB 4TZK.

What this is NOT

These are computational candidates, not validated therapies. They have not been synthesized, tested in any biological assay, or reviewed by any regulatory body. See DISCLAIMER.md before drawing any conclusion from the output.

Verification

Every candidate passes through a 12-layer filter chain before it appears on the leaderboard:

#	Layer	Tool	Gate
1	SMILES validity	RDKit parse	valid atoms + valence
2	Structural hygiene	PAINS + REOS	no pan-assay-interference scaffolds
3	Drug-likeness	Lipinski + Veber + Ghose	all three pass
4	Synthesizability	SA Score (RDKit)	SA ≤ 6
5	Primary docking	AutoDock Vina	ΔG < −8.0 kcal/mol
6	Secondary docking	DiffDock / GNINA	pose consensus with Vina
7	Mechanism sanity	PLIP interaction fingerprint	contacts ≥ 2 catalytic residues
8	ADMET	SwissADME + ProTox-II + pkCSM	no critical red flags
9	Novelty	ChEMBL Tanimoto (Morgan r=2)	max similarity < 0.4
10	Resistance-proof	target hotspot lookup	not hitting known resistance site alone
11	Retrosynthesis	AiZynthFinder	≥ 1 feasible route in ≤ 4 steps
12	Red team	adversarial Claude call	no substantive flaw found

Beyond the per-candidate chain, continuous pipeline-health loops run in the background:

• Positive-control loop (hourly) — re-docks known GyrB inhibitors (novobiocin, GSK299423, clorobiocin); their rank must stay in the top 5%.
• Negative-control loop (hourly) — docks DUD-E property-matched decoys; the pipeline must maintain EF1% ≥ 5 and ROC-AUC ≥ 0.75 on the actives/decoys split.
• Cross-method consensus (weekly) — top 20 re-docked with Vina + DiffDock + GNINA; disagreements published.
• Novelty-drift loop (weekly) — newly published antibiotics pulled from ChEMBL; existing candidates re-scored for novelty.
• Self-consistency loop — top candidates re-scored 3× with scrambled prompts; high variance demotes.
• Reproducibility loop — CI re-runs scoring on every candidate; same SMILES + target must produce same scores.

All loop metrics are surfaced on the public dashboard in real time. The whole point is that you can watch the pipeline verify itself.

Architecture

orb-antibiotic-scientist/
├── src/
│   ├── agent.py               # Claude Agent SDK runner (Opus, auto-compact, resume)
│   ├── agent-prompt.md        # System prompt: mission, rubric, output contract
│   ├── watchdog.sh            # Process supervisor (restart on crash / 5h stall)
│   ├── control.sh             # start / stop / status / logs / findings
│   └── tools/
│       ├── validate_smiles.py
│       ├── dock_vina.py
│       ├── dock_diffdock.py
│       ├── admet.py
│       ├── novelty.py
│       ├── mechanism.py
│       ├── retrosynthesis.py
│       └── redteam.py
├── loops/
│   ├── positive_control.py
│   ├── negative_control.py
│   ├── consensus.py
│   └── novelty_drift.py
├── targets/
│   ├── mrsa-gyrb.yaml
│   └── tb-inha.yaml
├── findings/
│   ├── candidates/            # SMILES + metadata
│   ├── docking/               # Vina poses + scores
│   ├── admet/                 # SwissADME / ProTox-II / pkCSM reports
│   ├── novelty/               # ChEMBL similarity reports
│   ├── mechanism/             # Binding-pocket interaction analysis
│   ├── red-team/              # Adversarial critiques
│   ├── leaderboard.json       # Top candidates, updated hourly
│   └── weekly-reports/        # Agent-written weekly summaries
├── dashboard/                 # Next.js public site
├── orb.toml                   # Orb Cloud deploy config
└── DISCLAIMER.md

Running the agent

# Start the agent (runs in background, survives terminal close)
./src/control.sh start

# Check if it's running and see stats
./src/control.sh status

# Tail logs
./src/control.sh logs

# See what it's found
./src/control.sh findings

# Stop the agent
./src/control.sh stop

The agent picks up where it left off by inspecting findings/ on restart.

Deployment

Runs on Orb Cloud. orb.toml configures checkpoint-to-NVMe during LLM waits and docking jobs — the agent hibernates between bursts of activity, so continuous operation costs fractions of a cent per hour.

End-to-end deploy via the Orb Cloud REST API:

# Default provider: Z.AI GLM Coding Plan (~$10/mo flat, routes through
# https://api.z.ai/api/anthropic using the Anthropic-compatible surface
# that Claude Code / Claude Agent SDK speak natively).
ORB_API_KEY=orb_... ANTHROPIC_AUTH_TOKEN=<zai_key> ./scripts/orb_deploy.sh deploy

# Switch to native Anthropic API:
ORB_LLM_PROVIDER=anthropic ORB_API_KEY=orb_... ANTHROPIC_API_KEY=sk-... \
  ./scripts/orb_deploy.sh deploy

# Lifecycle:
./scripts/orb_deploy.sh status    # computer + agents JSON
./scripts/orb_deploy.sh promote   # manual wake
./scripts/orb_deploy.sh demote    # manual sleep
./scripts/orb_deploy.sh usage     # cost so far

The deploy script caches computer/agent IDs under .orb-state/ so follow-up commands don't re-register. The Z.AI key can also be dropped into .orb-state/zai-key for local reuse (gitignored).

Dashboard (live leaderboard + pipeline-health charts):

python dashboard/server.py            # http://localhost:8000

Tests

pip install -r requirements.txt
python -m pytest tests/ -v

129 tests cover every per-layer gate, the composite scoring function, leaderboard upsert semantics, and the pipeline-health loop logic. CI runs the suite on Python 3.10 and 3.11 with an additional reproducibility check that scores an identical dossier twice and asserts byte-identical output.

License

MIT — see LICENSE.

Contributing

Academic and industry wet labs interested in testing top-ranked candidates: please open an issue or email the maintainer. The top 5 candidates (by composite rigor score) each week are prepared as SMILES + mechanism brief + suggested assay for wet-lab validation.

---

Built as part of the Orb Cloud open-source repo strategy. The Orb-specific moat here is that an autonomous scientist with continuous pipeline-health loops is economically viable only when compute sleeps during the ~95% of time the agent waits on docking jobs and LLM responses.