A non-detect is not a zero. Span reads the whole series of blood draws for tumour DNA (ctDNA), not the single value, and counts the draws that came back empty as evidence about what is there.
Twelve simulated patients, each plotted against its own limit of detection · synthetic, method demo only
One simulated patient
Every dot above the line is a draw that found tumour DNA. Everything the line cuts off below it is a draw that found none, and in the clinic, that is the end of the story.
Weeks 20 to 42
Not detected. Not detected. Not detected. In the chart, in the trial database, in the patient’s own copy of the result, each of these is stored as a blank. Five blanks in a row look like five months of good news.
Underneath the line
Nothing found does not mean nothing was there. It means the tumour DNA was below what the test can see, on that day, which is a measurement: a number and a date. Fourteen of them, in order, pin down a curve that the handful of positive results never could on their own.
The model calls the turn
It flags the change here: eighteen weeks before the tumour DNA is even strong enough for the test to report it, and off a run of results that were mostly blanks. Nothing was visible yet. The shape of the series was.
Change point fitted at week 56.1 by a censored-Poisson Bayesian latent-growth detector; the latent fraction clears the assay floor at week 74.5. Simulated patient; the method is not simulated.
Same blood, different reader
A scan found the progression at week 95. The standard way of reading these same tubes, wait for one value to cross the line, does not raise a flag until week 122, nearly six months after the scan: off tubes that had already been drawn, with no new sample and no new assay.
One simulated patient, the 90th-percentile case: not the result. The cohort result is directly below.
Synthetic patient data: method demo only
| Week | Result |
|---|
Trained and scored across all three sources in our benchmark, a random forest predicts time to resistance at a C‑index of 0.656, well clear of the 0.500 floor. (The C‑index is how often a model puts two patients’ times to resistance in the right order: 0.500 is a coin toss, 1.000 is perfect.) Re-fit and scored inside a single registry, where knowing which dataset a patient came from can no longer stand in for biology, the same model falls to 0.432, with an interval that contains chance.
The binding constraint is the data modality (single-timepoint snapshots) and not the model. That is the argument for reading a series, and it is the one result on this page that was measured rather than illustrated.
OncoTraj v1, Task B (time to resistance). Bootstrap 95% intervals. All three tasks, with the majority baseline beside each
The story above followed the 90th-percentile case. Across the whole simulated cohort, Span flags 25–27% of progressions at least 12 weeks before imaging, against 11–22% for a single-draw rule. The gap comes from the single-draw rule, not from Span: Span holds near a quarter in every regime, while the single-draw rule falls from 22% to 11% the longer the subclone stays below the limit.
The figure opens on that cohort: one mark per simulated patient who progressed, each rule’s catches filled first. Switch to Patient to scrub through single cases; the first is the median, where Span is late. Everything in it is synthetic.
A series of ctDNA draws goes in and a patient trajectory comes out. Two of these are outputs the method has been evaluated on. The third is listed with what backs it, because it is the one still to be demonstrated.
Roadmap, not built: EHR and imaging ingestion, and ranked next-line options with trial matching.
The next section is the experiment that would change the first of these.
Span is a method waiting for the data it was built to read, and the experiment that would settle it is small, retrospective, and already specified. It needs no new samples, no protocol change, and no patient to be treated on a Span alarm.
Span sits on top of the tests, not against them. Commercial liquid-biopsy panels report a molecular result from a single draw, and they do that well; Span reads the series those draws already produce.
Preprint. The censored-Poisson latent-growth change-point detector, arXiv:2606.11876.
Preprint. 813 patients, frozen leakage-audited splits, arXiv:2606.11144.
Four snapshot models across three tasks. None identifies the resistance mechanism.
For anyone new to the idea: why resistance is usually found late, what an empty blood test still carries, and what changes if care can see it coming.
Every year, cancer takes nearly ten million lives. Most of us know one of them. And for many, the drugs worked, until one day they didn’t.
Resistance. The cancer quietly evolves around the treatment. It is not bad luck. It is evolution, and on targeted therapy it comes for almost every patient.
Here is the cruel part. We usually find out from a scan, months after the change began. Medicine is reading yesterday’s news.
Care is decided from snapshots. One scan, one test, one frozen moment. But a disease is not a photograph. It is a movie.
So start from first principles. Tumours shed fragments of their DNA into the blood. Every blood draw is one frame of that movie.
And here is the part everyone throws away. When a test comes back empty, that is not a zero. It is a whisper. Each faint miss, each faint hit, is a clue about how fast something is growing below the line we can measure.
Span listens to every whisper across the whole series, and marks the day resistance begins, while it is still invisible on a scan.
We tested the snapshot approach. Remove the shortcuts, and it predicts resistance no better than a coin flip. The snapshot itself is the limit.
Why does this matter? Because in cancer, time is the one thing you cannot buy back. Months of warning mean months of options.
Imagine cancer care that steers instead of reacts. Every patient corrected before they drift, until a deadly disease becomes one people live with.
The data to prove it already exists. We just have to read it. Predict the trajectory, not the snapshot. This is Span AI.
Oncology decides care from snapshots of a moving disease. If we can learn the trajectory itself, we can act before the window closes, for every patient, not just the ones scanned at the right moment.
Aarchi Singh Thakur, Co-founder & CEO
If you hold a serial-ctDNA arm, the experiment above is the whole ask. If you are evaluating Span as an investment, ask for the deck: it comes back in the thread, and we will answer the questions this page leaves open. We reply to every message, and nothing here is gated: the methods paper is a direct download on the research page.
