What does Provotics do?
Provotics runs Provenance-1, a model that reads a single tumor RNA-seq expression profile and estimates the tumor's likely site of origin across 25 anatomical sites. From the same transcriptome it also returns linked reads such as molecular subtype, over-expressed druggable targets as research hypotheses, immune context, and a calibrated confidence for each call, with the driver genes behind the prediction.
Is Provotics a medical device?
No. Provotics is a research and educational project. It is not a medical device and is not intended for clinical diagnosis or treatment decisions. Nothing it returns should be used to direct patient care.
How accurate is it?
On real held-out patients it reaches a macro-F1 of 0.908 and balanced accuracy of 0.911 across the 25 sites. On 381 independent tumors drawn from 7 separate studies (different patients, centres, and pipelines) it scores 96.1%. Just as important, it knows when it is unsure: at 90% conformal coverage it abstains on inputs it cannot place confidently rather than forcing a guess. These figures come from a research and educational project, not a medical device.
What input does it need?
One tumor RNA-seq expression profile, gene-level counts or TPM, mapped to the curated 3,882-gene panel the model reads. The pipeline harmonizes to a common gene space and screens quality before any prediction is made.
Does it need patient identifiers or PHI?
No. Provenance-1 works from expression values only. It does not require names, dates of birth, medical record numbers, or any protected health information to produce a report.
What does it mean when the model abstains?
Abstaining means the model has decided it cannot place a profile confidently, so it declines to commit to a site rather than returning a low-quality guess. We run conformal prediction at 90% coverage and out-of-distribution gating, so unfamiliar or low-quality inputs are flagged or held back. An abstention is a feature: it is the model telling you it does not know.
Can it diagnose or detect cancer?
No. Provenance-1 does not screen for, detect, or diagnose cancer, and it does not establish that a sample is a tumor. It estimates a likely site of origin for a profile that is already understood to be a tumor sample. "Site of origin" carries no diagnostic authority and is for research and educational use.
What can't it do yet?
It has a cross-platform generalization gap. On single tumors produced on sequencing platforms it has not seen, it abstains on roughly two-thirds of cases and is correct on about 98% of the ones it does call. We treat that as the honest behaviour: when the input is off-distribution, the model holds back rather than guessing. Closing that gap is active research.
How is my data handled?
We do not require or want PHI; expression values are enough. Data use is governed by our access agreement, and we keep handling minimal and purpose-limited. See the
Privacy page for the full detail.
Who can access it?
Provotics is in an invite-only research phase. Access is granted by application so we can confirm intended research and educational use and set up the access agreement.
Apply for access to be considered.
How much does it cost?
Access is apply-first, and pricing depends on who you are:
- Individual: $49/month.
- Students: free under a university plan.
- Research team: $6,000/year per seat.
- Pharma: from $150,000/year.
See the
Pricing page for what each tier includes.
Where can I read more?
The
Platform page covers what the model reads from one profile,
Research and
Safety cover how it is evaluated and where it abstains, and
DocsFAQ covers running your own profiles once you have access.