Evidence-first genomics
Every variant, judged by its evidence.
Omevra reads the evidence behind a genetic variant and shows its working: what supports the call, what is missing, and how far to trust it.
Live application
Run the Variant Evidence Agent — right here
The public demo runs on this page — nothing to install, no account.
Public demo — please do not upload patient data. Use example or de-identified variants only.
Every variant, judged by its evidence.
Curated example. On the full test set, 37.3% of variants are summarised (10,357 of 27,748); the rest are declined for insufficient evidence.
Research Use Only.
After a quiet period the app can take up to a minute to wake. It loads on its own — no need to refresh.
How it works
Data in. Pipelines run. Evidence out.
One standard across every product: measure first, publish the result with its limits, and never let a confident-sounding answer outrun its evidence.
Real data, real scale — thousands of variants, dozens of sequencing runs, full task suites. No toy demonstrations.
Governed, reproducible runs with a frozen test set, so a result can be re-checked instead of merely believed.
Every claim cites its source, or the system declines. Calibration is reported alongside accuracy, because both matter.
The rule behind everything here is cite-or-decline. A variant report is only as good as the evidence it can point to. When the evidence is thin, declining is the correct answer — and it is the answer the Variant Evidence Agent gives 17,391 times out of 27,748.
Calibration is reported alongside discrimination: the final model reaches AUROC 0.970 with a calibration error of ECE 0.025, measured on the 10,357 variants it scored out of the 27,748 in the frozen temporal test set — a test set the model never trained on.
Validation
Tested on the future, not the training set
Most variant tools are tested on data they were tuned on. Omevra’s headline numbers come from a temporal split: the model was fitted on ClinVar as frozen in January 2024, and tested on variants whose classifications were made or changed after that freeze — a test set the model never trained on. A temporal split cannot leak the way a random split does.
AUROC 0.970, measured on the 10,357 variants the model scored out of the 27,748-variant temporal test set — with a calibration error of ECE 0.025 on the same scored variants: when the model says 0.80, it is right about 80% of the time.
Coverage is 37.3%. The model scores the missense variants that carry predictor scores and declines the rest — most of them non-missense variants with no predictor score at all — rather than inventing a probability.
Every claim in an evidence report cites the record retrieved for it. Where the evidence is too thin to support a claim, the system declines — and the decline is shown as a decline, never dressed up as a result. Cite-or-decline.
See full results
| Model | n (test) | AUROC | ECE | Brier |
|---|---|---|---|---|
| AlphaMissense alone (raw) | 10,373 | 0.921 | 0.040 | 0.113 |
| Ensemble (AlphaMissense + gene prior) + isotonic | 10,373 | 0.958 | 0.017 | 0.078 |
| From-scratch MLP + isotonic | 10,373 | 0.947 | 0.013 | 0.090 |
| M2+R (AlphaMissense + gene prior + REVEL) + isotonic | 10,357 | 0.968 | 0.023 | 0.066 |
| Random forest + REVEL + isotonic | 10,357 | 0.970 | 0.024 | 0.063 |
| GBM+R (gradient boosting + REVEL) + isotonic — the final model | 10,357 | 0.970 | 0.025 | 0.062 |
Measured on the same temporal test set; every row is labelled with its own n.
Use cases
Built for the VUS problem
Three ways research teams use calibrated, cited variant evidence — and the limits of each, stated plainly.
An evidence pack for each candidate variant: a calibrated probability where the evidence supports one, a citation behind every claim, and an explicit decline where it does not — so a researcher starts from the evidence, not from a blank search page.
Score a batch list in one run: the borderline band (0.40–0.60) surfaces for human review first, and declines stay visible in the dashboard and the exports — the whole cohort triaged, including the honest “not enough evidence” answers.
Gene-level evidence summaries from the variant record, and reproducible pipeline runs through the Atlas — the same code, run again, returning byte-identical results that a reviewer can verify.
Pricing
Plans that scale with your variants
Start free. Paid plans open with the first pilot cohort, and pricing is confirmed with you before anything is ever charged.
FREE
$0
Trying it out — 100 variants per month · 1 seat · Evidence packs (PDF, CSV, XLSX).
See the benefits
- Variants per month — 100
- Seats — 1
- Evidence packs (PDF, CSV, XLSX) — Yes
- Saved projects and history — 7 days
- Shared team workspace — No
- Re-check alerts on new ClinVar evidence — No
- API access — No
- Pipeline runs (FASTQ to VCF) — No
- Support — Docs
- Single sign-on (SSO) — No
RESEARCHER
$79 / month
$63 / month billed annually
Individual researchers — 5,000 variants per month · 1 seat · Unlimited saved projects and history.
See the benefits
- Variants per month — 5,000
- Seats — 1
- Evidence packs (PDF, CSV, XLSX) — Yes
- Saved projects and history — Unlimited
- Shared team workspace — No
- Re-check alerts on new ClinVar evidence — Monthly
- API access — No
- Pipeline runs (FASTQ to VCF) — Credits
- Support — Email, 2 business days
- Single sign-on (SSO) — No
LAB
$499 / month
$399 / month billed annually
Research labs and core facilities — 50,000 variants per month · 5 seats · Shared team workspace · API access.
See the benefits
- Variants per month — 50,000
- Seats — 5 (add more at $49 / seat)
- Evidence packs (PDF, CSV, XLSX) — Yes
- Saved projects and history — Unlimited
- Shared team workspace — Yes
- Re-check alerts on new ClinVar evidence — Weekly
- API access — Yes
- Pipeline runs (FASTQ to VCF) — Credits
- Support — Email, 1 business day
- Single sign-on (SSO) — No
ENTERPRISE
Custom
Biotech, pharma, CROs — Unlimited variants · Unlimited seats · API access with higher limits · Single sign-on (SSO) · Dedicated contact and SLA.
See the benefits
- Variants per month — Unlimited
- Seats — Unlimited
- Evidence packs (PDF, CSV, XLSX) — Yes
- Saved projects and history — Unlimited
- Shared team workspace — Yes
- Re-check alerts on new ClinVar evidence — Custom
- API access — Yes, with higher limits
- Pipeline runs (FASTQ to VCF) — Included allowance
- Support — Dedicated contact and SLA
- Single sign-on (SSO) — Yes
Early pilot partners: a free Lab plan for 90 days in exchange for feedback.
The builder
Built by Umesh Linga
Get started
Start free. Scale when the work does.
The Variant Evidence Agent runs in public. Open it, run the example variants, and read the evidence pack it returns — every claim cited, every gap shown. Public demo — please do not upload patient data. Use example or de-identified variants only.
Book a demo — a 30-minute call, in three parts:
- A live run on de-identified example variants — your own de-identified list, or the public demo sets.
- The validation method — the temporal split, the calibration, and the coverage limits, with the numbers behind them.
- Pilot terms — how early access works for the first cohort of labs.
Your message goes only to Umesh. Never sold, never added to a mailing list. Privacy