Biomarker Discovery

Detect biological signals hidden inside complex data.

Disease biology produces subtle signals spread across genomics, proteomics, metabolomics, clinical data, and records. AI connects patterns that are not visible one dataset at a time.

Work

What we look for

Patient stratification

Find the subgroups a trial or a clinic is actually treating.

Disease signatures

Multi-omic signatures that hold up outside the cohort they were found in.

Treatment response prediction

Separate responders from non-responders before exposure.

Diagnostic and prognostic markers

Signals with a decision attached, not only a correlation.

Multi-omics integration

Joint models across modalities with different noise and missingness.

Clinical phenotype discovery

Phenotypes learned from records rather than assumed in advance.

Find the signal inside biological complexity.

Discipline

The hard part is not finding a signal

It is showing that the signal survives a cohort it has never seen.

Hold out by scaffold and by site

Random splits flatter every model. We split the way deployment will.

Pre-register the gate

The threshold is written before the result, and it is not widened to fit the outcome.

Report the smallest detectable effect

A non-significant result on thin data is not evidence of no effect, and we do not present it as one.

Why AI

Why a model sees what a clinic does not

The signal is spread across features no person can hold at once

One prognostic signature emerged from 6,642 unfiltered morphological measurements of tumour slides; organ-specific ageing clocks emerged from roughly 5,000 plasma proteins. In both cases the information was in the combination, not in any single standout marker.

It is not limited to the features medicine has already named

Clinical vocabulary encodes what the unaided eye could resolve and name. A model optimises against an outcome and is free to use whatever in the raw pixels or peptides correlates with it.

It reads together what the clinic reads in separate rooms

A retina goes to ophthalmology and a haemoglobin to haematology. Those boundaries describe how care is organised, not where the information lives.

It turns imaging patients already receive into screening

If a routine photograph or a ten-second trace carries a signal, the marginal cost of the test is zero. That is a different economics from ordering a new assay.

Evidence

This has already happened.

Findings that were in the data all along, and that no one had asked the data for.

AUC 0.93weak heart pump from an ECG

A century-old test turned out to carry a signal nobody read

A ten-second ECG was never thought to reveal how strongly the heart pumps. A network trained on 44,959 patients found that it does, and in a randomised trial across 45 sites and 22,641 patients the tool increased new diagnoses of low ejection fraction. Patients flagged despite a normal echo had a fourfold higher risk of developing dysfunction later.

Attia et al., Nature Medicine 2019; Yao et al., Nature Medicine 2021 ↗

AUC 0.97sex from a retinal photo

A retina holds cardiovascular risk factors, and more

Trained on 284,335 patients, a model read age, blood pressure and smoking status off ordinary retinal photographs. It also identified sex almost perfectly. An independent replication states plainly that clinicians remain unaware of the retinal features that differ between sexes.

Poplin et al., Nature Biomedical Engineering, 2018 ↗

6,642features measured blind

The prognosis was in the tissue around the tumour

An automated system measured thousands of features of breast cancer slides without being told which mattered. The features that best predicted survival were not in the cancer cells at all, but in surrounding stroma that pathologists do not grade. The score held on two independent cohorts.

Beck et al., Science Translational Medicine, 2011 ↗

0.63 g/dLhaemoglobin error

An eye photograph used as a blood test

Anaemia is diagnosed by drawing blood. Deep learning read haemoglobin concentration directly from retinal photographs in 11,388 patients, turning imaging that diabetic patients already receive into an opportunistic screen.

Mitani et al., Nature Biomedical Engineering, 2020 ↗

11organs aged separately

Your organs age at different speeds, and plasma says so

From about 5,000 plasma proteins in a single blood draw, machine learning estimated a biological age for each of eleven organs across 5,676 adults. Roughly one person in five has one organ ageing far faster than the rest, and an accelerated heart carries a markedly higher risk of heart failure.

Oh et al., Nature, 2023 ↗

The honest other half: the same freedom produces artefacts, and metrics alone cannot tell them apart. Models predict a patient's self-reported race from chest X-rays at 0.91 to 0.99 AUC through image corruption and filtering, which no clinician can do and nobody can explain. Networks identified the source hospital of a radiograph with 99.95 percent accuracy, and one pathology model fell from 0.84 to 0.69 when moved to a cohort from another country. A signal counts only once it survives an independent cohort from a different site, which of the cases above only the ECG work has fully cleared.

Bring the cohort.
We bring the search.

Start a Discovery Program → Read our research