Partner with us
The approach

Three engines, kept honest and kept separate.

We refuse to collapse cancer into a single score. Each engine answers a different question, is measured against the strongest available methods, and is tested on data it has never seen.

Engine A

Presentation

Which targets does a tumor actually display?

What we do
We model which fragments a specific tumor naturally shows on its surface — accounting for the fact that what a cell could display and what it does display are different things.
How we know we’re right
Better recovery of independently measured targets than strong baselines, on tumors the model never saw.
What we don’t claim
A list of what might bind is not a list of what is really there.
Engine B

Recognition

Which displayed targets can a T cell truly engage?

What we do
We predict whether an immune receptor can recognize a given target selectively — and, separately, whether that recognition would actually act on a real cancer cell rather than just bind in isolation.
How we know we’re right
On entirely withheld receptors and targets, recover more genuinely responsive combinations — and tell near-identical healthy and tumor targets apart.
What we don’t claim
Binding is not the same as function, and function is not the same as a safe, effective therapy.
Engine C

Design

When nature falls short, can we engineer recognition?

What we do
Where a patient has no natural response to a verified target, computational protein design offers an orthogonal route: building recognition molecules against that specific target.
How we know we’re right
Novel designs that bind the intended target and discriminate against a stringent panel of look-alikes — confirmed in a small, blinded experiment.
What we don’t claim
A molecule that binds in an assay is a tool, not a treatment.
Two directions at once

Forward and inverse.

Most pipelines run one way: from a tumor’s molecular data, predict a target. We also run the other way — starting from immune cells that already show signs of recognizing a tumor, and working back to what they see.

The two directions check each other. Where they disagree is often where the most valuable learning hides — and those disagreements become training signal.

Why measurement is the point

Prediction earns trust only against reality.

Models are cheap to make and easy to fool. The discipline that matters is comparing predictions to real, independently generated measurements — and counting the misses, not just the hits.

Before we spend on new experiments, we prove value on public data the model has never seen. Only a claim that survives that gets tested in the lab.

The discipline

Know. Believe. Prove.

Everything we publish is tagged by how sure we are. A hypothesis never wears the clothes of a result — and we try hardest to prove ourselves wrong.

See the research agenda →