OpenAMP Foundry is a verification-first, safety-constrained dry-lab foundry for AI-assisted antimicrobial peptide (AMP) discovery.
It is designed around a strict principle:
Computers can triage, falsify, rank, and document candidates. They do not prove biological efficacy. Wet-lab assays are still required before any scientific claim of activity.
The current repository is a rigorous dry-lab foundry.
The larger mission is more ambitious:
Build an open wet-lab compression engine for AMP discovery: a system that helps qualified scientists decide which small number of experiments are most worth running next, then learns from those outcomes.
The long-term infrastructure ambition is described in VISION.md, GOAL.md, docs/research/OPEN_BIOTECH_STACK.md, docs/research/OPEN_INFRASTRUCTURE_MOAT.md, and docs/trust/TRUST_CENTER.md.
AI generation is cheap.
Trusted candidate selection is scarce.
OpenAMP Foundry exists to build the open evidence layer between AI-generated biological hypotheses and qualified experimental testing. The project is not trying to make biology look solved by software. It is trying to make experiment selection more reproducible, auditable, baseline-aware, and safe.
This repo gives you a safe starting point for:
- building AMP candidate datasets;
- scoring candidates with transparent baseline heuristics;
- checking novelty against known references;
- penalizing likely safety/synthesis risks;
- selecting diverse candidates;
- generating auditable JSON evidence certificates;
- bundling expert-review packs with provenance and identity hashes;
- maintaining run manifests, schema registry, and release-status records;
- running a demo pipeline without downloading external biological datasets;
- expanding later with real predictors and qualified external validation.
It also establishes the architecture and governance needed for a future virtual assay layer that can improve experiment selection without pretending to replace biology.
A starter implementation of a computer-first AMP candidate foundry:
candidate records
-> validity checks
-> physicochemical features
-> activity-likeness score
-> safety-risk score
-> feasibility score
-> novelty score against references
-> ensemble rank
-> evidence certificate
-> run manifest
-> expert-review package, if human review approves
The present repo answers:
Can we build a reproducible, leakage-aware, safety-first ranking pipeline that earns the right to guide real experiments?
The next-horizon repo should answer:
Can we compress wet-lab cost by learning which peptide experiments are worth running, better than cheap predictors alone?
This repo is not:
- a medical product;
- a drug-discovery guarantee;
- a wet-lab protocol collection;
- an unsafe biological-design tool;
- a generator for harmful biological capabilities;
- a replacement for qualified microbiologists, toxicologists, or regulatory experts.
The default repo contains only toy/demo data and transparent baseline scorers. It deliberately avoids:
- operational biological instructions;
- unsafe optimization objectives;
- release of unscreened high-risk candidate lists;
- trained generator weights;
- clinical or medical advice.
See SAFETY.md, RESPONSIBLE_USE.md, and MODEL_RELEASE_POLICY.md.
OpenAMP now operates on two connected horizons:
- Current horizon — trustworthy dry lab Build deterministic ranking, evidence certificates, leakage-resistant benchmarks, novelty auditing, feasibility checks, and reviewable shortlist generation.
- Next horizon — wet-lab compression Add higher-fidelity membrane, selectivity, stability, and learned-surrogate layers that improve which small number of experiments a qualified lab should run next.
The second horizon only matters if the first one stays honest. Better simulation without calibration is not a breakthrough.
Requires Python 3.11+.
python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
make demoValidate one generated evidence certificate:
python -m openamp_foundry.cli validate \
--certificate outputs/evidence/AMPF-000001.json \
--schema schemas/candidate.schema.jsonFor first-run interpretation, read docs/getting-started/FIRST_RUN_WALKTHROUGH.md. For command interpretation, read docs/getting-started/COMMAND_SURFACE.md. Commands produce artifacts; artifacts support claims only when the proof ladder allows them.
openamp-foundry/
README.md # primary entrypoint
VISION.md # long-term infrastructure vision
GOAL.md # milestones, kill rules, metrics
MISSION.md # project mission and claim boundaries
GOVERNANCE.md # decision governance
AGENTS.md # agent operating contract
CLAUDE.md # concise collaborator guidance
CONTRIBUTING.md # contributor workflow and PR checklist
CODE_OF_CONDUCT.md # community and scientific integrity standard
SAFETY.md # safety policy
SECURITY.md # security and safety-sensitive reporting
RESPONSIBLE_USE.md # allowed/disallowed use
MODEL_RELEASE_POLICY.md # model and artifact release policy
DATA_LICENSE_NOTICE.md # data license and redistribution policy
CITATION.cff # citation metadata
.github/ # PR template, issue templates, CODEOWNERS
configs/ # scoring and recalibration policy
data/README.md # data directory rules
models/README.md # model directory rules
docs/README.md # task-based documentation front door
docs/PROJECT_INDEX.md # exhaustive document catalog
docs/trust/TRUST_CENTER.md # safety/evidence/governance trust front door
docs/research/OPEN_INFRASTRUCTURE_MOAT.md # durable infrastructure thesis
docs/research/NUMBER_ONE_REPO_STANDARD.md # category-leader standard
docs/getting-started/FIRST_RUN_WALKTHROUGH.md # first-run path
docs/getting-started/COMMAND_SURFACE.md # command workflows and claim boundaries
docs/engineering/SCHEMA_REGISTRY.md # schema and artifact registry
docs/engineering/RUN_MANIFEST_STANDARD.md # provenance standard
docs/engineering/ADAPTER_AUTHOR_GUIDE.md # safe adapter authoring
docs/trust/RISK_REGISTER.md # major risks and mitigations
docs/operations/SUSTAINABILITY_AND_BUS_FACTOR.md # sustainability and bus-factor plan
docs/trust/PUBLICATION_POLICY.md # public claims policy
docs/research/NEXT_100_PR_MAP.md # PR-sized roadmap
docs/engineering/CI_AND_QUALITY_GATES.md # CI and quality gates
docs/operations/HUMAN_AGENT_COLLABORATION.md # human-agent collaboration model
docs/getting-started/REVIEWER_ONBOARDING.md # reviewer guide
docs/research/ADOPTION_METRICS.md # adoption metrics focused on trust
docs/operations/DECISION_RECORD_TEMPLATE.md # decision record template
docs/getting-started/HUMAN_ONBOARDING.md # human contributor onboarding
docs/getting-started/AGENT_ONBOARDING.md # agent task protocol
docs/evidence/PROOF_LADDER.md # evidence levels and claim ladder
docs/evidence/CLAIM_REVIEW_CHECKLIST.md # claim review checklist
docs/trust/DATA_GOVERNANCE.md # data governance standard
docs/trust/MODEL_CARD_TEMPLATE.md # model/adapter card template
docs/engineering/ARTIFACT_VERSIONING.md # artifact compatibility policy
docs/trust/RELEASE_CHECKLIST.md # release checklist
docs/evidence/BENCHMARKING.md # benchmark suite
docs/evidence/BENCHMARK_GOVERNANCE.md # benchmark lifecycle and governance
docs/evidence/METRICS_CURRENT.md # current benchmark summary
docs/evidence/CALIBRATION_POLICY.md # recalibration gate policy
docs/evidence/EVIDENCE_CERTIFICATE.md # candidate certificate spec
docs/evidence/VIRTUAL_ASSAY_SCOPE.md # virtual-assay scope and gates
docs/review/WET_LAB_HANDOFF.md # safe expert-review handoff guide
examples/ # toy datasets only
outputs/.gitkeep # generated files ignored by git
schemas/ # JSON schemas
scripts/ # helper entrypoints and compatibility shims
scripts/benchmarks/ # canonical benchmark and baseline entrypoints
scripts/calibration/ # canonical calibration workflow entrypoints
scripts/external/ # canonical external predictor and handoff entrypoints
scripts/lab/ # canonical lab handoff entrypoints
scripts/novelty/ # canonical novelty DB and audit entrypoints
scripts/release/ # canonical demo, evidence, and reproducibility entrypoints
scripts/research/ # canonical exploratory generation and screening scripts
scripts/waves/ # canonical wave-program generation and panel scripts
src/openamp_foundry/ # Python package
tests/ # repository tests
tests/benchmarks/ # benchmark and regression-gate tests
tests/calibration/ # calibration workflow tests
tests/external/ # external workflow and report tests
tests/lab/ # lab handoff and return-validation tests
tests/novelty/ # novelty scoring and novelty-pressure tests
tests/release/ # release artifact and reproducibility tests
tests/waves/ # wave-program gate tests
The project optimizes for honest candidate selection, not impressive claims.
A candidate is only worth lab money if it survives independent attacks:
- basic validity;
- novelty check;
- feasibility review;
- predicted activity;
- predicted safety;
- diversity selection;
- reproducible evidence bundle;
- human review.
The first serious milestone is not “AI discovered an antibiotic.”
The first serious milestone is:
A reproducible pipeline can recover known AMP positives, reject weak controls, avoid leakage, and generate a small shortlist of candidates that survives qualified external review.
The longer-range milestone is:
A calibrated virtual assay layer helps the project choose fewer, smarter experiments and improves hit-rate or safety-adjusted yield relative to cheap predictors alone.
- Core code: Apache-2.0.
- Documentation: intended for CC BY 4.0 reuse where marked.
- Third-party data: not bundled unless redistribution is allowed.
- Generator weights and unscreened candidate lists: not released by default.
- Project name and logo: trademark retained.