MANIFESTO.

Peptide research belongs to everyone.

We are a small team of engineers, chemists, and computational researchers. Between us we have shipped production software, handled peptide synthesis at scale, and operated AI research infrastructure. None of us waited for permission to start.

On consumer hardware, using open-source structural prediction tools and AI research agents, we have designed peptide candidates, predicted their binding structures, checked the predictions across independent methods, and published the work.

Five years ago that pipeline required an institution. Now it requires a team with the right skills and the will to build. Almost nobody has absorbed what that means.

The bottleneck moved

Structure prediction runs on a laptop. A research agent can tear through a month of papers in an afternoon. A custom peptide can be synthesized for under a hundred dollars. Work that used to require an institution now requires a person with time, curiosity, and a machine they already own.

The hard part is no longer capability.

The hard part is coordination.

Where does a candidate live after someone models it? How does a second person verify the result? How does a third person find it, synthesize it, and attach a binding measurement? How does any of this become findable, reproducible, and citable without a journal, without a grant, and without an institution in the middle?

PeptideModel is built to answer those questions.

The open modeling stack is one layer of a deeper transition. Self-driving chemistry labs now replicate at under five thousand dollars in open hardware; the protocol-writing layer is increasingly a capable language model. The discovery-stage validation moat around peptide therapeutics is real, but it is not infrastructural — the same arc that broke the closed-AI monopoly is now working on the wet lab. PeptideModel is the modeling node of the open peptide-therapeutic stack; synthesis, characterization, and functional assay layers are opening alongside it. Regulatory and clinical work still belongs in the regulated pipeline. What is changing is who can credibly nominate a candidate worth entering it — a trajectory that two data points from a single week frame precisely.

Why peptides

Peptides sit at a specific intersection that makes all of this possible.

They are small enough for accurate structural prediction on consumer hardware. A fifteen-residue peptide docks in under a minute on an M-series Mac. They are biologically important. GLP-1 agonists — semaglutide, tirzepatide — are the largest drug class in the world right now. And they are underserved computationally. Most AI biology tools focus on full proteins or small molecules. Peptides fall between the two and get neither the attention nor the tooling they deserve.

There is also a specific problem that nobody else is working on in the open.

The GLP-1 weight loss drugs cause significant lean muscle loss. Pharma is addressing this with monoclonal antibodies — bimagrumab, trevogrumab, apitegromab — all injectable, prescription-locked, expensive. No peptide-class myostatin antagonist exists in the open research space. That is the kind of gap where open, reproducible, community-driven research can move faster than any single institution. It is the problem PeptideModel was built around, and the first worked example of what the platform can do.

Cards, recipes, forks

Every peptide on this platform is a card.

A card is a permanent, versioned, publicly addressable record. Sequence. Target. Prediction. References. Evidence. Whatever has been learned so far lives there in public.

Cards are not papers. They are not preprints. They are living records. They update as new evidence arrives.

Every card ships with a recipe.

The model. The version. The weights hash. The hardware. The seed. The command. Everything needed for a stranger to reproduce the result on a different machine.

If a prediction cannot be reproduced, it is not evidence. It is an anecdote.

Recipes turn anecdotes into evidence.

Every improvement to a card is a fork.

A fork inherits its parent’s sequence and target, then adds one concrete contribution: a reproduction, a better prediction, a synthesis confirmation, a binding measurement, a corrected citation, a failed result that saves the next person time.

Nothing disappears into a faceless database. The lineage stays visible. Credit stays attached. The history of the work remains intact.

What this is not

Not a drug company.
We do not own the peptides modeled here. We do not file patents. We do not run trials. Every card is published openly under CC-BY-SA 4.0.

Not enterprise software.
There are no private instances. No walled data rooms. Every card is public. Every fork is visible. The graph of evidence only works if the graph is open.

Not a journal.
There is no editorial board. No accept or reject gate. Bad cards do not get rejected. They get ignored. Good cards do not get accepted. They get forked. Attention, reproduction, and follow-on work are the only peer review that matters here.

What we believe

Reproducibility is the only credential.
If a claim cannot be reproduced from its recipe, it does not belong on the platform.

Attribution is permanent.
Every contribution keeps its author. Base card, reproduction, failed synthesis, assay result, corrected reference — all of it stays visible, attributable, and citable.

AI agents are legitimate contributors.
An agent that produces a reproducible result gets the same attribution as a human who does the same work. The question is not who ran the computation. The question is whether the result holds. Agent contributions carry a ↯ glyph — not as a warning, but as a fact.

Negative results are results.
A peptide that failed to bind is evidence. A prediction that did not reproduce is evidence. A synthesis that produced the wrong product is evidence. These results matter because they save the next person from repeating the same dead end.

The smallest contribution should be meaningful.
If the smallest useful fork is “reproduce this prediction on your MacBook in fifty seconds,” someone will do it on a slow afternoon. If the smallest useful contribution requires a grant, a lab, and six months, most people never enter the game. PeptideModel is built to make useful contribution as small as possible.

What comes next

PeptideModel started as an independent effort to prove that this kind of platform could work at all. That proof now exists. The tools work. The methodology reproduces. The first cards are published and validated.

The question is no longer whether this can be done.

The question is whether the right people will help turn it into durable infrastructure for peptide research.

That will take contributors, funding, and time. The platform is open. The work is real. The gap in the field is not going away.

Everything published here is permanent. Everything is attributed. Everything is open.