Not a pitch-deck number — a tested result
Before a single commercial sample was run, the analysis engine was built on public patient data and then tested on studies it had never seen. Here are the actual numbers, including the ones that went down.
The score below (statisticians call it AUC) runs from 0.50 to 1.00. 0.50 means the model is guessing — no better than a coin flip. 1.00 would be perfect. The number that matters is how a model does on new studies it has never seen, because that is what a real customer's data will be. A model always looks better on the data it was built on.
These two groups of genes move in opposite directions between patients and controls: neutrophil/inflammation markers up, lymphocyte markers down. The score is simply one group minus the other. There is no hidden model, so it is fully interpretable.
The main tests were planned before the results were seen, and every result is reported, including the ones that didn't work. The same standard applies to every dataset the platform generates. View the code and the live app on GitHub →
What we got wrong
The first number was too good. On the study where the genes were found, the signature scored 0.76–0.77. Tested on seven new studies, it scored 0.61. The signal is real (it beat chance in every study), but it is smaller than the first result suggested. This is the most common trap in biomarker research, and we chose to measure it rather than hide it.
Most of the signal is the white-blood-cell mix. People with depression in these studies tend to have slightly more neutrophils and fewer lymphocytes. A standard blood count captures this as well as our genes do. Nine more complex models we tried, from small gene panels to machine learning on all genes, did not beat it on new studies.
Not all data is good data. In one large public study, patients and controls fell into separate processing batches, so any model could "detect depression" just by spotting the batch. Another used a different blood fraction (PBMCs), in which the signal flips direction. We now run a quality check on every dataset before it is used.
What this tells us: the blood signal for depression is real but small, and it is mostly about which immune cells are present. To go further, you need either a sharper question (for example, which patients have an inflammatory form of depression) or a more direct tissue: the patient-derived neurons the platform is built to profile.
How this compares to published work
Published DNA-based risk scores for major depression (iPSYCH cohort) reach AUC 0.57–0.63. Our blood signature reaches 0.61 on new studies, in the same range. The two measure different things: a DNA score estimates lifelong risk, while a blood signature reflects the body's current state.
In other words, we did not beat the state of the art, and we don't claim to. What this proof of concept shows is that the pipeline finds real, interpretable biology and tells the truth about how far it generalises. That is exactly the discipline the laboratory engine needs.
This is a proof of concept on public blood data from adults with depression. It is not a clinical test, it is not regulator-approved, and it has not yet been run on patient-derived neurons, which is what the laboratory engine is being built to do. On new studies the signature performs at AUC 0.61, about the same as a standard blood count. What it demonstrates is that the analytical pipeline works end to end, finds real and interpretable biology, and is tested honestly on data it has never seen.
What we're testing next
We are now asking a sharper question: not "does this person have depression?" but "is this an inflammatory form of depression?" Roughly 1 in 4 to 5 patients show low-grade inflammation, and they may respond differently to treatment. Across all 8 studies we found the same inflammation pattern, and about 1 in 5 patients carries it. Whether it predicts treatment response is still open: our first test, in two public treatment trials, was not conclusive, and the next step is to check it against measured CRP.
Why this approach is grounded, not speculative
None of the underlying claims are novel on their own. Each one is where the peer-reviewed literature already points. Click through the evidence behind each layer.
Carrying a pathogenic variant is not the same as developing disease. Modifier genes, epigenetics and environment all shape whether a mutation produces illness — which is why penetrance in neurodevelopmental conditions is so often incomplete. Most strikingly, a 2026 re-analysis found that many published penetrance figures were substantially inflated, with some variants dropping from an assumed 10–40% to near zero. That is the gap NeuroPenetrance exists to measure directly.
Growing a patient's own cells into neurons is now an established way to study brain conditions in human biology rather than in animal proxies. The field's own recent framing is that combining these models with multi-omics measurement and computational analysis is precisely what moves the work toward patient-specific diagnostics — which is the thesis this company is built on.
Genome-wide studies have linked thousands of variants to psychiatric conditions. Each one carries a tiny effect alone, but combined into a polygenic risk score they produce a reproducible signal already used for research-grade risk stratification. Recent appraisals describe the remaining barrier to clinical use as validation and communication rather than discovery — and that is the layer functional cellular data is meant to strengthen.
No clinical guideline currently includes a molecular test for depression — diagnosis still rests entirely on symptom interviews. Protein and gene-expression profiling is the leading candidate to change that: a decade of work has produced reproducible candidate signatures, and researchers increasingly describe large-scale validation, not discovery, as the remaining obstacle.