OmevraOmevra · Capacity
Capacity
What each product accepts as input, the verified limits, and what it does not take yet — how much Omevra takes at a time.
Variant Evidence Agent
Variant lists — a gene and a variant in HGVS text, pasted in directly or uploaded as a file (.txt / .csv / .tsv, up to 2 MB).
500 variants in one quick batch; 25,000 in one queued job. A demonstrated full run processed all 27,748 variants of the test universe at about 2,779 variants per second.
The scored universe holds 27,748 variants across 3,084 genes (ClinVar, January 2024 freeze). Coverage: the model scores 10,357 of those 27,748 variants (37.3%) — the missense variants with predictor scores — and those receive an evidence summary. The other 17,391 are declined, most of them non-missense variants with no predictor score at all. A decline is always shown as a decline, never dressed up as a result. Extending coverage — splice and non-coding variants — is next on the roadmap.
Outputs: a visual dashboard (charts and an evidence heat map), a per-variant evidence pack with a citation on every claim, and exports in PDF, CSV and XLSX.
Raw sequencing files (FASTQ / BAM) — those belong to the NGS Pipeline Atlas. Variants outside the scored universe get a live evidence check or a decline — never an invented score.
NGS Pipeline Atlas
Raw sequencing runs (FASTQ). The Atlas runs where the data already lives — the lab’s own cluster or cloud storage — not as a browser upload.
33 real E. coli K-12 sequencing runs processed end to end: 27 passed quality control, and 6 were stopped at the ≥10× coverage gate instead of producing unreliable results. A second full run came out byte-identical. The Atlas is built on Nextflow, the field-standard pipeline engine.
A run at the 10 TB scale has not been demonstrated. The same pipeline and the same quality gates are what a run of that size would use, and accuracy does not change with volume because every sample runs the same code — but 10 TB is not yet demonstrated, and this page says so plainly.
Hybrid Perturb-Bench
Nothing to upload — it is an explorer over the published Norman 2019 Perturb-seq dataset (111,445 cells).
Your own data. The explorer covers that one fixed, published dataset.
Multi-Agent Evidence
Variant evidence questions in plain words.
Measured head to head against a single agent: both systems scored 1.00 on citation accuracy — the ceiling of that benchmark, so the test cannot separate them — and the multi-agent system cost more. A harder benchmark is needed.
Web Agent
A public web address plus a task in plain words.
At most 2 public runs at a time. In the final evaluation, 36 of 36 tasks completed with guardrails on (95% CI 90.3–100%); the adversarial suite scored 9 of 9 (95% CI 66.4–100%) against 0 of 9 (95% CI 0–33.6%) with guardrails off.
Platform notes
The Atlas turns raw reads into variant lists, and the Variant Evidence Agent takes it from there.
Evaluation universe v1, built on the ClinVar monthly release, January 2024 freeze. The universe is refreshed as a new versioned release, at most quarterly.