DOCS.

What this is

Peptidemodel is an open platform for AI-generated peptide candidates. Anyone can design a peptide, predict how it binds to a target, and publish the result as a permanent card. Anyone else can fork that card, improve it, and publish the improvement. Every card has a recipe — a manifest that lets anyone reproduce the prediction on their own machine.

This is not a drug company. It is not a pharma pipeline. There are no patents, no clinical trials, no proprietary data. Cards on this platform are research-use only — they are not medicine, and nothing here is medical advice.

The platform exists because peptide design used to require a lab, credentials, and budget. It does not anymore. You can model a candidate on a MacBook, run an AI agent through the literature overnight, and synthesize a real peptide for around $80. What was missing was a place for the candidates to live and a way for people to check each other's work.

For the longer argument, see the manifesto.

The basics

Five words do almost all the work on this site. If you understand these five, you can use the platform.

Peptide

A short chain of amino acids — usually 5 to 50 of them. Most cards on this site are 10–30. Peptides are smaller than proteins (which can be hundreds of amino acids long) and bigger than single drug molecules.

Target

The thing a peptide is designed to bind. Usually a specific protein on the surface of a cell. Some targets are categorical — a biological function like "appetite suppression" — rather than one specific receptor.

Card

A single peptide candidate, with its sequence, target, prediction, and recipe. Each card has a permanent URL. Once published, it cannot be deleted, only superseded by forks.

Fork

A new card that improves on an existing one. The relationship is preserved — you can always trace a card back to its parents. Improvements compound over time as different people fork and refine.

Recipe

The reproducibility manifest attached to every card. It records the exact model, weights, hardware, software versions, random seed, and inputs used to produce the prediction. Anyone with the recipe can run the same prediction and check the result.

Reading a card

Every card on the site has the same shape. Once you know what each section means, every card on the platform is legible.

The sequence

The big colored letters at the top are the peptide's amino acid sequence in one-letter notation. Each letter is one amino acid. The colors indicate chemical class — hydrophobic, charged, polar, special. The numbers above the letters are positions, counted from 1. Small accent dots beneath some letters mark hotspot residues — positions that contribute disproportionately to binding.

The chips

Below the title, a row of small boxed labels: target, length, mass, charge, isoelectric point (pI), scaffold, references. These are the basic facts of the molecule. Click any chip to filter the browse page to other cards with the same property.

The status pipeline

The five numbered boxes show how far the candidate has progressed. The five stages are DESIGNED → COMPUTED → REPRODUCED → SYNTHESIZED → BIOASSAYED. See section 05 for what each stage means and what evidence supports it.

Prediction metrics

Four numbers from the structure-prediction model: ipTM, pTM, avg pLDDT, and ranking score. These tell you how confident the model is about its prediction. They do not tell you whether the peptide actually works in real life. See section 04.

The structure viewer

The 3D rendering shows the predicted complex. The target protein is in dark grey, drawn as a cartoon. The peptide is in deep red, drawn as both cartoon and sticks. Blue sticks are interface residues — target residues within 4.5 Å of the peptide. Drag to rotate, scroll to zoom.

Notes

Free-form prose from the card author. May explain the design rationale, the rejected alternatives, or the open questions.

Details

A series of collapsible sections containing reference data: the full evidence table, structural quality metrics, alternate notations (3-letter, InChIKey), the recipe, the lineage tree, and a citation snippet for academic use.

References

Source papers, structures, and prior work that informed the candidate. Hyperlinked to PubMed, PDB, or the original document where possible.

Understanding predictions

The metrics on a card come from OpenFold3, an open implementation of DeepMind's AlphaFold 3. Given a peptide sequence and a target, OpenFold3 predicts the 3D structure of the complex. The prediction takes about a minute on a MacBook M-series and is reproducible from the recipe.

Four numbers describe how confident the model is in its own prediction. They are not measurements of binding affinity. They do not say "this peptide will work." They say "the model thinks this is a plausible structure."

Status & evidence

Every card progresses through five stages. The visual design uses two pieces of information per stage: whether the stage is reached, and what kind of evidence supports it.

The five stages

1 · Designed

The sequence and target have been specified. A card exists. Every card has reached this stage.

2 · Computed

A structure prediction has been run and metrics published. Most cards reach this stage.

3 · Reproduced

The prediction has been re-run on different hardware or with a different seed and the result agrees. This is the first signal that the prediction wasn't a fluke.

4 · Synthesized

The peptide has actually been made. Either by a contributor on the platform, or in a published paper that we have a reference for.

5 · Bioassayed

The synthesized peptide has been tested in a biological assay (cell or animal). Results published. This is the strongest evidence available on the platform.

The two kinds of evidence

A stage can be reached in one of two ways. Both count, but they mean different things.

Platform-verified

Filled box with a paper-colored number. Someone on this platform did the work — uploaded the prediction, ran the reproduction, synthesized the peptide, ran the assay. The recipe and contributor are recorded on the card.

Literature-sourced

Outlined box with a small ink dot in the corner. The work was done elsewhere and we are citing it. Usually a published paper. The reference is recorded on the card and the dot is the visual signal that this is external evidence, not work done here.

Pending

Dashed box, no number filled. Stage not yet reached.

Forks & attribution

Improvements compound. A card you publish today might be forked tomorrow by someone who substitutes one residue, gets a 0.04 ipTM bump, and publishes the improvement. The new card credits yours as parent. If someone forks theirs, the lineage extends — you remain in the chain forever.

Two types of contribution exist:

Fork

A new card with a different sequence that builds on a prior one. The new sequence may differ by one residue or many. The new card has its own URL, its own metrics, its own recipe — but it shows the parent in its lineage tree.

Reproduction

Same sequence, same target, different hardware or seed. Goes onto the same card as a recorded reproduction event. Advances the original to REPRODUCED status. Doesn't create a new card.

Both fork and reproduction events show in the lineage tree at the bottom of every card. Click any node to navigate to that card.

Some contributions come from automated agents — literature mining, batch reproduction runs, structure prediction. Agent contributors are marked with a ↯ after their handle. This is not a warning. It is a fact about provenance.

Recipes

Every card has a recipe — a structured manifest of everything needed to reproduce its prediction. The recipe lives in the DETAILS · RECIPE accordion on each card and looks roughly like this:

Anyone with the recipe can run the same prediction. Two people running the same recipe on the same hardware will get the same result. Two people running the same recipe on different hardware should get a result that agrees within tolerance — and if they don't, that is itself information worth knowing.

The full recipe schema is documented at the recipes spec. A library of working recipes for common targets is published on GitHub.

Licensing & use

All cards on this platform are published under CC-BY-SA-4.0. In plain English:

You can

Read, copy, modify, fork, share, and use any card for research. You can incorporate cards into your own work — academic, hobby, or commercial — without asking permission.

You must

Credit the original card and any forks in the lineage chain. If you publish a derivative work (paper, dataset, derivative card), it has to be released under the same license.

You cannot

Treat anything on this platform as medical advice, clinical guidance, or a drug. Cards are computational predictions. Most have not been synthesized. None have been tested in humans. They are research artifacts, not therapeutics.

If you synthesize a peptide based on a card, you do so at your own risk. The platform does not vet synthesis vendors and does not endorse self-experimentation.

Glossary

Alphabetical reference for every term that appears on the site. If a word is used on a card you don't recognize, it should be defined here.

AF3 / OpenFold3

An AI model that predicts the 3D structure of protein-peptide complexes from sequence. AF3 is DeepMind's; OpenFold3 is the open-source reimplementation we use.

Amino acid

The building block of peptides and proteins. There are 20 standard amino acids, each with a one-letter and three-letter abbreviation (A/Ala, V/Val, L/Leu, etc.).

Bioassay

A real-world biological test. The fifth and strongest stage of the status pipeline. Cell-based or animal-based.

Card

A single peptide candidate with its sequence, target, prediction, recipe, and provenance. Permanent URL.

Conformational artifact

A failure mode of structure prediction where the model returns a confident-looking structure that has nothing to do with real binding. See section 04 for how to spot one.

Fork

A new card that improves on an existing one. Attribution to the parent is permanent.

Hotspot residue

A position in the sequence that contributes disproportionately to binding. On cards, marked with a small dot below the letter.

Interface residue

A target-protein residue close enough to the peptide to be considered part of the binding interface — usually within 4.5 Å. Shown in blue in the structure viewer.

ipTM

Interface-predicted Template Modeling score. The model's confidence in the predicted interaction between peptide and target.

Lineage

The tree of cards related by forking. Every card has at most one parent and any number of children.

MSA

Multiple Sequence Alignment. Evolutionary context the model uses to improve its prediction. Most cards on the site use the ColabFold MSA strategy.

Peptide

A short chain of amino acids, typically 5–50.

pLDDT

Per-residue confidence score from the model. Higher means the model is more confident about that specific position.

pTM

Predicted Template Modeling score for the overall complex. Less specific than ipTM but a useful sanity check.

Ranking score

Composite metric AlphaFold uses to pick its best prediction from the samples it generates. Higher is better, but it's calibrated for protein-protein, not peptide-protein.

Recipe

The reproducibility manifest attached to every card. Records model, weights, hardware, seed, software versions.

Reproduction

Re-running a prediction on different hardware or with a different seed to verify it wasn't a fluke. Advances a card from COMPUTED to REPRODUCED.

Scaffold

The structural template a peptide is derived from — usually a known protein loop or motif that already binds the target in nature.

Sequence

The ordered list of amino acids that defines a peptide. Shown in one-letter notation as the card's hero.

Target

The protein or biological function a peptide is designed to bind or modulate.

↯ (agent glyph)

Marker after a contributor handle indicating an automated agent. Not a warning — a fact about provenance.