Building longitudinal models of disease.
AI learned language, then vision, then the rules of proteins. The next frontier is disease over time, not as static snapshots, but as living systems that evolve. Cancer is the proving ground, and drug resistance is where those trajectories are most learnable.
Three things just became true at once.
-
Expanded
Targeted therapies
Standard of care has shifted from chemo to a large and growing arsenal of targeted drugs, each with its own resistance pathways, and nearly all eventually stop working.
-
Validated
Serial ctDNA
TRACERx tracked ctDNA a median of about 151 days ahead of clinically detected relapse in resected NSCLC (Abbosh et al., Nature 2023), and PADA-1 showed that acting on a rising ESR1 ctDNA signal before radiographic progression roughly doubled progression-free survival — 11.9 months against 5.7 (Bidard et al., Lancet Oncol 2022).
-
Ready
Next-line drugs
Therapies built to overcome resistance already exist. The missing piece is finding the right patient before the window closes.
The technical risk is bounded. The data risk is the whole risk.
Span is a method waiting for the data it was built to read. That sentence is the company, and it is the one worth stress-testing.
- 1The test that would falsify it. Stated in the methods paper, and sharp enough to fail: on real serial ctDNA, at a matched false-alarm rate, change-point detection on the sequence of detection calls should catch more impending progressions, and catch them earlier, than thresholding the reported variant fraction. If it does not, the thesis is wrong and the benchmark will say so.
- 2Where it should not help, said first. The same paper reports the regimes where the advantage collapses: when resistance emerges fast, and when draws are too sparse to observe the flicker below the limit of detection. The synthetic evaluation is reported by regime for exactly that reason.
- 3What it is worth if it lands. Conditional, and worth stating once: PADA-1 randomised patients on a rising ESR1 ctDNA signal ahead of radiographic progression and roughly doubled progression-free survival, 11.9 months against 5.7. Lead time on that decision is the thing being bought. Span has shown no such result and is not claiming one.
- 4What has to be collected. No public serial-ctDNA breast-cancer cohort exists — the methods paper opens by saying so, and OncoTraj v1 is the argument for why one has to be built. The method is specified, the splits are frozen and the evaluation is written down. What it needs next is a serial arm to read, which is the conversation we are here to have with trial sponsors and cancer centres.
All four points are drawn from the CP-BLG methods paper and OncoTraj v1. Both are linked in full on the research page
Span does not run an assay. That decides most of the business.
Span produces no sample and no sequencing read. The input is a series of detection calls that somebody else’s assay has already produced, on somebody else’s draw schedule, ordered by somebody else’s clinician.
That rules out the models Span is most often mistaken for. Span is not a laboratory-developed test, not a CLIA laboratory and not a competing panel, and is not working toward becoming one.
It leaves three counterparties, and only three — everyone who already holds a serial ctDNA series:
- 1Trial sponsors. Serial draws at six-to-eight-week intervals are already routine inside trials. The read is on an arm they have collected and can score against outcomes they already hold.
- 2Assay makers and MRD programmes. They hold the series and today report each point in it separately. Span is a reader of the sequence, which is the positioning argument.
- 3Cancer centres running serial monitoring inside their own protocols, where the series exists but nothing reads it as a series.
In all three cases the first engagement is the same, and it is retrospective: one existing serial arm, the change-point test scored against a pre-specified naive rule. That experiment, specified in full
What is not decided, and is not going to be asserted here: whether the eventual form is a licence, a co-development or a service; what it costs; and the regulatory route. Span’s software is research use only, is not a medical device and has not been submitted to any regulator. Those are decisions that follow the first result on real serial data, and that result does not exist yet.
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.
Disease moves as a trajectory. So should the medicine.
We're building a longitudinal understanding of how each patient's disease evolves, learned one trajectory at a time. Our ambition is the world's most accurate models of disease over time for precision oncology, so every patient receives the right therapy at the right time.
Those two sentences sit at a different altitude from the artefact, and the gap is worth closing rather than leaving for a reader to notice. The detector itself has no trainable parameters, so it does not get sharper by being run — the advantage is structural, and it is deliberately fixed because there is no serial data to fit it to. What use would accumulate is the calibration around it: decision thresholds fitted per growth regime and per assay, and priors over how trajectories move. None of that exists yet. Span is pre-clinical, and this is the work ahead, not a position already held.