We select candidate blood biomarkers from published research, measure them as a panel, and use specialist models to interpret the pattern. Each disease programme has its own evaluation.
The work starts with published research and biological expertise. Which signals are associated with a disease? In which people? Under which conditions?
We use those questions to investigate candidate biomarkers and the role each could play in a panel.
Some biomarkers move together. Others capture a different part of the biology. We look for complementary information and study what each measurement adds.
The goal is a useful combination: signals that help find the disease, distinguish it from other conditions, and work together in the model.
For the multi-cancer panel, machine learning and scientific review narrowed about 2.5 million PubMed articles published since 2018 to 729 studies for detailed review and 18 candidate markers.
A tumour leaves three kinds of trace in blood. We weighed all three before choosing proteins. Early in the disease, tumour cells and tumour DNA are scarce in a blood draw. Proteins are read as levels across the whole sample rather than counted as rare particles, and clinical laboratories already measure them every day. Whether a protein panel finds early-stage disease is what our studies are designed to test.
Whole cancer cells shed from a tumour into the bloodstream.
1 in 1 billionblood cells is a tumour cell, by one estimate, even in advanced cancer.Source 1
Measured with Cell capture and imaging
Ten thousand dots and not one tumour cell. At one in a billion, you would need about 100,000 fields like this to expect one.
Published study data, not VarSee results.
Fragments of tumour DNA among the cell-free DNA every blood sample carries.
1 in 10,000DNA fragments, or fewer, came from the tumour in half of the stage I lung cancers analysed.Source 3
Measured with Deep sequencing or digital PCR
One tumour fragment among ten thousand. In half of the stage I cancers analysed, there was less than this.
Published study data, not VarSee results.
Molecules whose levels in blood change as a cancer develops and the body responds.
36 proteinswere linked to lung cancer in blood drawn up to three years before diagnosis, across six cohorts.Source 5
Measured with Automated immunoassays
Not a rare particle to find, but a change in level across the whole sample, read across a panel.
Published study data, not VarSee results.
No single protein is specific to one cancer, and several also rise with inflammation or other illness. That is why we measure a panel and read it with specialist models, in context.
How specialist models read a panelPublished research on each kind of signal, from academic studies. These figures are not results of VarSee and do not describe any company’s test.
A single measurement can make two patterns look alike. Add another dimension and the relationship becomes easier to see.
In research, we test that contribution against the individual markers in the same population.
Of the people with the condition, how many does the test find? Everyone past the threshold is flagged; squares short of it are missed.
Of the people without the condition, how many does the test correctly clear? Circles past the threshold become false alarms.
A second measurement can pull the groups apart, so fewer are missed and fewer are flagged by mistake. Whether a real panel does this is what validation has to show.
Our mixture-of-experts approach gives specialist models different parts of the problem. A gating layer brings their outputs together.
