This document explains how OpenAMP Foundry can grow from a dry-lab antimicrobial peptide pipeline into a reusable open infrastructure pattern for biotech.
The ambition is large, but it must be earned through narrow evidence.
The next important open-source layer in biotech is not merely a model that generates biological designs.
It is a trustworthy stack for deciding which AI-generated biological hypotheses deserve real experiments, recording what happened, and learning without rewriting history.
OpenAMP Foundry is an attempt to build that stack first for antimicrobial peptides.
AI can produce abundant biological suggestions.
Biology cannot test abundant suggestions cheaply.
This creates a new bottleneck:
old bottleneck: generating hypotheses
new bottleneck: selecting experiments worth running
A system that generates 100,000 plausible peptides is not automatically useful.
A system that helps a qualified lab choose the 8–12 most informative candidates for the next assay may be useful.
A system that learns from the result of that assay and improves the next batch may become infrastructure.
Linux became powerful because it became a dependable neutral layer that many competing groups could build on.
The OpenAMP version of that ambition is not a kernel for all biology.
It is a neutral experimental-decision layer for a constrained biological design problem.
The analogy only holds if OpenAMP becomes:
- open enough to inspect;
- stable enough to build on;
- boring enough to trust;
- modular enough to extend;
- safe enough to release responsibly;
- rigorous enough to survive hostile review;
- useful enough that independent groups adopt its artifacts.
A slogan does not make infrastructure.
Repeated external use does.
OpenAMP should become a layered stack.
Layer 0: Safety and governance
Layer 1: Data provenance and licensing
Layer 2: Candidate validation and normalization
Layer 3: Feature extraction and baseline scoring
Layer 4: Novelty, safety-risk, synthesis, and diversity checks
Layer 5: Benchmarking and leakage resistance
Layer 6: Evidence certificates and candidate manifests
Layer 7: External predictor and virtual-assay adapters
Layer 8: Human review and lab-handoff packages
Layer 9: Result intake and negative-result archive
Layer 10: Calibration and active learning
Layer 11: Cross-lab reproducibility and governance memory
The upper layers are only trustworthy if the lower layers are boring and stable.
Safety is the base layer.
The project must maintain clear boundaries around:
- harmful biological optimization;
- pathogen handling;
- wet-lab protocols;
- unscreened candidate release;
- model-weight release;
- clinical language;
- human-review requirements.
Without safety, openness becomes fragile.
Biology projects often fail silently through data confusion.
OpenAMP should make these distinctions explicit:
- generated candidates;
- public reference data;
- benchmark positives;
- benchmark negatives;
- known controls;
- lab-tested candidates;
- failed candidates;
- restricted or non-redistributable data.
Every dataset should have provenance, license status, date, known bias, and intended use.
Before scoring, the system should reject or flag invalid candidates.
This includes:
- invalid amino-acid symbols;
- malformed rows;
- duplicate sequences;
- length outliers;
- obvious synthesis problems;
- missing metadata;
- unsafe release context.
A clean pipeline starts by refusing bad inputs.
Baseline scoring is not a temporary embarrassment.
It is the permanent adversary of model hype.
The stack should preserve transparent baselines such as:
- charge;
- length;
- hydrophobicity;
- amphipathicity;
- similarity;
- simple selectivity proxies;
- random valid selection.
Advanced models earn authority only by beating these under fair tests.
A candidate is not useful merely because it scores high for predicted activity.
The project must score and report:
- novelty against known references;
- similarity to known families;
- predicted hemolysis/cytotoxicity risk;
- synthesis feasibility;
- redundancy inside a panel;
- structural-class diversity;
- known model blind spots.
A panel should be designed as an experiment, not scraped from the top of a ranking table.
The stack must assume benchmarks are easy to fool.
Important benchmark defenses include:
- train/reference/test separation;
- near-duplicate detection;
- charge-matched controls;
- composition-matched controls;
- family-stratified performance;
- cross-dataset evaluation;
- cheap-baseline comparisons;
- regression gates in CI;
- explicit benchmark cards.
A benchmark result is useful only if the easiest cheating explanation has been attacked.
A selected candidate should carry a machine-readable evidence package.
A useful certificate should answer:
- What is the sequence?
- Where did it come from?
- What config scored it?
- What features were computed?
- What baselines did it beat?
- What novelty checks were run?
- What safety risks were flagged?
- What failure modes remain?
- What exact code version produced this output?
This is how a dry-lab decision becomes reviewable.
The stack should welcome better models without trusting them by default.
Adapters should be:
- versioned;
- explicit about scope;
- optional;
- fail-closed;
- provenance-preserving;
- compared against baselines;
- forbidden from silently downloading weights or sending sequences to external services without consent.
A model is a plugin.
Evidence discipline is the platform.
The project should help qualified reviewers make decisions.
A good handoff package should include:
- candidate manifest;
- evidence certificates;
- selection rule;
- controls;
- novelty report;
- safety-risk report;
- synthesis notes;
- benchmark caveats;
- explicit claims that are not supported;
- pre-registered success/failure criteria.
The handoff should not include unsafe wet-lab instructions.
A discovery system that hides failures cannot learn.
OpenAMP should treat negative results as structured evidence when safe to publish.
Result intake should preserve:
- candidate identity;
- assay context at a safe abstraction level;
- endpoints;
- controls;
- uncertainty;
- interpretation boundaries;
- whether the result is strong enough to affect recalibration.
Negative results prevent repeated waste.
The real infrastructure leap is controlled learning from experiments.
The loop should be:
select batch
-> freeze evidence
-> test through qualified partner
-> ingest result
-> evaluate calibration gate
-> update only if allowed
-> choose next batch
The gate matters because otherwise the system can rewrite itself to fit a story.
The final layer is shared memory across groups.
The system should eventually support:
- independent reruns;
- external validation packages;
- public benchmark audits;
- safe negative-result archives;
- governance decisions;
- deprecation of misleading modules;
- compatibility promises for core schemas.
This is where a project becomes infrastructure.
OpenAMP should try to standardize artifacts before it tries to standardize conclusions.
High-value standards include:
- candidate manifest format;
- evidence certificate format;
- result intake schema;
- benchmark card format;
- negative-result entry format;
- model adapter contract;
- simulation-result interface;
- recalibration report;
- safety review checklist;
- lab handoff packet.
If external groups reuse these artifacts, OpenAMP can matter even beyond its own scoring model.
Do not prematurely standardize:
- one universal AMP scoring formula;
- one universal assay endpoint;
- one universal safety threshold;
- one closed candidate-generation model;
- one interpretation of noisy biology;
- one domain expansion path.
The project should standardize the evidence pipeline before standardizing scientific conclusions.
Agents need constraints.
OpenAMP gives them useful constraints:
- exact claim boundaries;
- safety rules;
- benchmark gates;
- schemas;
- deterministic commands;
- small-loop execution style;
- known failure modes;
- documented next bottlenecks.
That makes the repo an excellent environment for agents to do safe, cumulative work.
The agent does not need to be trusted as a biologist.
The agent needs to produce artifacts that qualified humans can inspect.
Humans need leverage.
OpenAMP gives serious contributors leverage by turning their work into reusable infrastructure:
- a benchmark becomes a guardrail;
- a safety review becomes a policy;
- a failed candidate becomes a negative-result entry;
- a lab result becomes calibration evidence;
- an adapter becomes an ecosystem bridge;
- a doc improvement becomes onboarding for future contributors.
The project should make expert judgment compound.
OpenAMP may expand beyond AMPs only after it demonstrates the loop in AMPs.
Expansion should require:
- a constrained problem representation;
- safe release boundaries;
- available reference data;
- meaningful benchmarks;
- qualified domain reviewers;
- realistic experimental feedback;
- clear baselines;
- evidence that the artifact standards still apply.
Do not expand because the story sounds bigger.
Expand because the evidence pattern transfers.
A project is infrastructure when other people can build on it without believing the founders' story.
OpenAMP should pass that test.
A serious outsider should be able to inspect the code, rerun the pipeline, reject weak claims, contribute a better module, add a benchmark, package a candidate panel, ingest results, and understand what changed.
That is the open biotech stack worth building.