The Provenance-2 core estimates anatomical site across 25 sites, reports uncertainty, and abstains when the evidence is insufficient. Separate research-preview signals can add context, but they do not have equivalent validation and do not replace orthogonal assays.
Site of origin is the flagship task with published cohort-level evidence. Confidence, abstention, and attribution support that task. Subtype, immune, and target signals remain exploratory and carry their own limits.
Estimates one of 25 anatomical sites (macro-F1 0.940 on the 2,700-sample internal held-out test, in-distribution). Under platform shift, the served decision policy abstains on roughly three in five profiles rather than force a single-site call.
Provides a within-site transcriptomic subtype signal only where a research head exists, including breast, colorectal, brain, and stomach. Evidence varies by site; it is not an assay result.
Links over-expressed genes to ChEMBL target and compound context. Over-expression is not dependency or predicted efficacy; these are research leads, never treatment recommendations. ChEMBL-derived context remains free research data, not a paid capability.
Summarizes literature-signature expression patterns from immune-cold to inflamed. It is a descriptive transcriptomic estimate, not IHC, flow cytometry, or a validated clinical biomarker.
The site-of-origin call carries calibrated confidence and a conformal candidate set in-distribution. Low-confidence or unfamiliar inputs can be flagged, widened to a set, or abstained on.
Gene attributions show which inputs most influenced a site call. They support review, but they are not causal drivers, biomarkers, or evidence about any single gene.
Not a score in isolation. A ranked, calibrated read you can interrogate, with the alternatives it considered and the confidence behind the call.
Illustrative example, not a live model output. On the held-out test set the model reads the right site at macro-F1 0.940 (in-distribution), and it abstains rather than force a call when no site is clearly ahead.
Schematic reliability diagram. Temperature scaling cut the model's calibration error roughly 8x (ECE 0.091 to 0.011), so a stated 80% means about 80% in practice, in-distribution.
Start from a single tumor expression profile, gene-level counts or TPM.
The pipeline maps to a common gene space, normalizes, and screens quality before any prediction.
The flagship core returns a site estimate, alternatives, uncertainty, and attributions. Any exploratory research layer is labeled separately with its own caveat.
The guided demo is illustrative, not a live model run. The evidence ledger publishes cohort-level validation and states why a reproducible accession-linked sample bundle is not yet public.