What does "a non-detect is not a zero" actually mean?

A negative liquid biopsy means the variant was below the assay's limit of detection, not that it is absent. Each draw samples a finite number of DNA molecules, so at low variant fractions detection behaves like a weighted coin: detect, non-detect, detect. Discarding the non-detects throws away most of the information in that period.

How far ahead of imaging can serial ctDNA show resistance?

The most-cited figure is TRACERx: a median lead of about 151 days, roughly five months, for ctDNA over clinically detected relapse in resected NSCLC, using personalised multi-mutation tracking rather than an off-the-shelf panel (Abbosh et al., Nature 616:553-562, 2023, doi:10.1038/s41586-023-05776-4). That is post-surgical surveillance, not on-treatment resistance in metastatic disease, so treat it as evidence that the window exists rather than as the window Span would operate in. Span's own methods paper demonstrates about six months of lead on a synthetic case. Neither figure is a Span clinical result.

A censored-Poisson Bayesian latent-growth change-point detector. It models the assay's detection process directly rather than fitting a curve to reported values, runs a sequential generalised-likelihood-ratio test for an upward shift in detection rate, aggregates evidence across competing resistance pathways, and fires at a calibrated threshold.

Is Span clinically validated?

No. Span is pre-clinical. There is no cleared or CE-marked assay, no prospective trial, and no patient has been treated on the basis of a Span alarm. Published results are a methods demonstration on synthetic and public longitudinal data, plus an open benchmark.

An open benchmark for longitudinal resistance prediction in EGFR-mutant NSCLC on first-line osimertinib. It harmonises 813 patients from two real-world registries and a published trial supplement into one schema with frozen, leakage-audited splits and reproducible baselines. Its headline result is negative: re-scored inside a single source, the timing task collapses to chance — random forest falls to a C-index of 0.432 [0.360, 0.514] against a 0.500 floor. The mechanism task has not been re-scored within-source.

The science

From scattered blood draws to a trajectory.

Serial ctDNA carries a signal that any single test throws away. Span reads the pattern of detections over time, and is designed to flag resistance while the variant is still below the limit of detection.

The core idea

A non-detect is not a zero.

When a liquid biopsy comes back negative, it doesn't mean the resistant clone is gone. It means the clone is below the assay's limit of detection (LoD). At low variant allele frequency, detection flickers, detect, non-detect, detect, because each draw samples a finite number of molecules (Poisson sampling).

A single value is blind during this sub-LoD dwell. So is its slope. But the rising rate of detections is not, and that is what Span models.

Fig. 1Schematic. Each draw is a coin weighted by how much tumour DNA is present, so single results flicker. The rate of detections turns upward while every individual value is still a non-detect — that turn is what Span models.

The method

The Span detector: a censored-Poisson Bayesian latent-growth (CP-BLG) change-point test.

It models the biology of the assay directly, rather than fitting a curve to noisy numbers.

Fig. 2The detector, end to end. Each stage is described in full below.

Model the detection process

Each draw's detect / non-detect is a Bernoulli outcome under Poisson sampling with LoD left-censoring, so non-detects inform the model instead of being discarded as zeros.

Test for a change-point

A sequential generalised-likelihood-ratio (GLR) test watches for an upward shift in each variant's detection rate, the earliest statistical sign a resistant clone is expanding.

Aggregate across mechanisms

Evidence is combined across competing resistance pathways (ESR1, PIK3CA, RB1, HER2), so the alarm reflects the whole resistance landscape, not one marker.

Fire at a calibrated threshold

The alarm is tuned to a matched false-alarm rate, so lead time is reported honestly against a controlled rate of false positives.

No trainable parameters. The advantage is structural, not learned.

Span is a transparent decision rule, there is nothing to overfit, and every alarm traces back to the detection pattern that produced it. That interpretability is a feature, not a limitation: it's what makes the method defensible to clinicians and regulators.

Common questions

What does "a non-detect is not a zero" actually mean?: A negative liquid biopsy means the variant was below the assay's limit of detection, not that it is absent. Each draw samples a finite number of DNA molecules, so at low variant fractions detection behaves like a weighted coin: detect, non-detect, detect. Discarding the non-detects throws away most of the information in that period.
How far ahead of imaging can serial ctDNA show resistance?: The most-cited figure is TRACERx: a median lead of about 151 days — roughly five months — for ctDNA over clinically detected relapse in resected NSCLC, using personalised multi-mutation tracking rather than an off-the-shelf panel (Abbosh et al., Nature 616:553–562, 2023). That is post-surgical surveillance, not on-treatment resistance in metastatic disease, so treat it as evidence that the window exists rather than as the window Span would operate in. Span's own methods paper demonstrates about six months of lead on a synthetic case. Neither figure is a Span clinical result.
What is CP-BLG?: A censored-Poisson Bayesian latent-growth change-point detector. It models the assay's detection process directly rather than fitting a curve to reported values, runs a sequential generalised-likelihood-ratio test for an upward shift in detection rate, aggregates evidence across competing resistance pathways, and fires at a calibrated threshold.
Is Span clinically validated?: No. Span is pre-clinical. There is no cleared or CE-marked assay, no prospective trial, and no patient has been treated on the basis of a Span alarm. Published results are a methods demonstration on synthetic and public longitudinal data, plus an open benchmark.
What is OncoTraj?: An open benchmark for longitudinal resistance prediction in EGFR-mutant NSCLC on first-line osimertinib. It harmonises 813 patients from two real-world registries and a published trial supplement into one schema with frozen, leakage-audited splits and reproducible baselines. Its headline result is negative: re-scored inside a single source, the timing task collapses to chance — random forest falls to a C-index of 0.432 [0.360, 0.514] against a 0.500 floor. The mechanism task has not been re-scored within-source.

What it produces

A decision, ahead of time.

Span turns a series of ctDNA draws into a patient trajectory, and reads three things off it. Two carry an evaluation. The third carries only a synthetic one, and is listed third and labelled for that reason; everything past all three is roadmap, and is marked as such below.

1Estimated time-to-resistance. How long the current line is likely to keep working.
2A calibrated threshold. The alarm fires at a level set for a chosen false-alarm rate, so the call is auditable.
3A ranked mechanism hypothesis — synthetic evidence only. The aggregation across competing resistance pathways is part of the method and has been demonstrated on synthetic cohorts. It has not been shown on real serial data, and OncoTraj's mechanism task — snapshot tissue NGS — has no within-source re-scoring at all. A serial-ctDNA cohort with sequenced progression biopsies is what would settle it.

Roadmap, not built: EHR and imaging ingestion, and ranked next-line options with trial matching. Span is pre-clinical and none of this is a cleared device.

See the evidence Read the paper (PDF)