A peptide AI, and the data it still has to make

The promise arrives on schedule. Point an AI at a disease, and it designs the molecule that fixes it. The peptide conversation has started to borrow that confidence, the way it borrows most confident things. So it is worth noticing a new project that opens from the other direction, by admitting what the field doesn't yet have.

What they announced

On October 16, the University of Toronto's Leslie Dan Faculty of Pharmacy described a project called OpenPep: an open-access database and an AI model meant to predict which peptides will bind to which proteins. It is led by Joseph Brown, an assistant professor there, with co-investigators across the university and a co-founder of the biotech ProteinQure. The funding is US$2.3 million from Google.org's AI for Science initiative. A peptide, for scale, is a short chain of amino acids, bigger than the small molecules most pills are built from, smaller than a full working protein.

The goal is to screen peptides against as many as 500 disease-relevant proteins, what the team calls a snapshot of the human proteome. The targets will be crowd-sourced from hospitals, companies, and research institutes, with a lean toward pediatric and rare diseases, cancer, neurodegeneration, and heart disease.

Why peptides are hard for an AI

The quietly honest line in the announcement is Brown's: "we don't have a lot of existing data for peptide discovery." That is the whole problem in a sentence. A model is only as good as what it learned from, and the models that design antibodies or small molecules were trained on decades of accumulated binding data. Peptides sit in an awkward middle, less catalogued, and the record of which ones stick to which proteins, and how hard, is thin.

So the first job is not prediction. It is measurement. You cannot teach a model to guess well on a subject nobody has written down.

The loop they're building

The design is a loop. One lab produces the target proteins using cell-free synthetic biology, a way of making proteins without growing whole cells. Brown's lab then uses mass spectrometry to test large libraries of peptides against those proteins and measure whether they bind, and how strongly. That experimental data trains the model to predict binding for proteins it has not been tested against. Those predictions get checked in the lab, and the new results feed back in to sharpen the next round.

Brown's framing of what the AI is for is the part worth keeping. "AI won't replace experiments," he said, "but it will help us choose better ones." That is a smaller claim than the usual one, and a more defensible one. The model is not the oracle. It is the thing that narrows a vast space of possible experiments down to the few worth running next.

How to hold it

This is early, and the announcement is careful to say so. There is no drug here, no clinical result, no validated therapeutic peptide. The database will open to the public only once it is established, and no timeline is given. Every prediction the model makes is a hypothesis until an experiment agrees with it. The benefits on offer, faster discovery and shorter paths to a candidate, are stated goals, not demonstrated ones.

What makes it worth a post anyway is the shape of it. Most of the AI-and-biology headlines promise the finished molecule. This one promises the infrastructure underneath, the open data and the honest loop, and treats the shortage of data as the thing to fix rather than the thing to paper over. Open resources have a way of outlasting the projects that make them, because other labs can build on a public database long after the original grant runs out.

None of that guarantees a useful peptide at the end. It just means the work is starting where the work actually is. The most valuable output of research like this, for a long while, will not be a drug. It will be a better list of which experiments to run, which is slower to celebrate and more honest than the promise it replaces.

---

This post is educational and general in nature. It is not medical advice. For guidance about your own health, talk to a qualified clinician.