Overview
Clinical genomics teams already have tools to identify genetic variants and predict whether they may cause disease. The bottleneck comes next: reviewing published evidence for each variant, applying established guidelines, and documenting the findings so another expert can verify and approve the conclusion. Today that means:
- Three-to-five hours of literature work per variant, mostly in full text papers and supplementary spreadsheets that generic search indexes poorly.
- A growing pile of variants of uncertain significance (VUS), because free annotation engines skip the criteria that decide a VUS.
- Two curators reaching two conclusions, with no shared record a third can use to resolve it.
- An SOP pinned to a ClinGen or ACGS specification that has since been revised, with no view of what changed.
- Internal case archives and cohort data that cannot leave your environment, so no SaaS tool sees them.
- No traceable path from a signed classification back to the sentence that justified it.
Loka builds this pipeline, deploys it in your AWS account, and hands it over. We start with ALMA, so you begin with a working framework.
PHASE 1: SPECIFICATION AND EVALUATION DESIGN
We pin your rulebook: the base framework, your national or panel specification (e.g. ACMG/AMP 2015, ACGS, CanVIG-UK, ClinGen VCEP), your combining rules, and your gene panel. These become versioned data files. We select an evaluation set from the ClinGen Evidence Repository for your genes and design the data boundary in your account.
PHASE 2: BUILD, INTEGRATION, AND REVIEWER WORKFLOW
Two lanes, split by whether a criterion is computable or needs reading prose.
- Computed lane: normalization to MANE Select, gnomAD filtering allelic frequency at 95% confidence, Ensembl VEP consequence, and ClinVar same-residue evidence. No language model states a number.
- Literature lane: proband counts, co-segregation, de novo status, functional assay validity, and phenotype specificity, read from full text and supplements. Cohorts reported in more than one paper are surfaced before any count is asserted, so no family is counted twice.
- Evidence: that cannot be weighed under your specification returns as indeterminate with its source, so the reviewer sees the gap.
- Deterministic scoring: so a reviewer can re-derive every conclusion.
- Citation checks: every cited identifier must resolve, and every quote must appear in the retrieved text. Claims that fail are dropped, not regenerated.
- Reviewer workflow: accept, re-weight, reject, or add evidence per code. The classification recalculates live.
- Agent orchestration on Amazon Bedrock: with per-run cost, timings, and a full step log.
- Optional: one internal evidence source, with explicit handling of evidence independence.
PHASE 3: EVALUATION, DEPLOYMENT, AND HANDOVER
We score against expert-panel curations on classification agreement, per-code precision, recall, and evidence attribution against the curator’s cited publications. Over-calling pathogenic is worse than over-calling uncertain, so we report error asymmetry separately. The phase closes with deployment, documentation, runbooks, and knowledge transfer.
WHAT YOU OWN AT THE END
- The pipeline in your AWS account.
- Your rule set as versioned data, plus the diff to the current published release.
- A per-variant evidence record including code, strength, verdict, sources with locators, reasoning, and the reviewer’s decision.
- An evaluation report scored per code and per cited source.
- A reviewer interface with agent steps, tool calls, per-paper views and CSV, Markdown, and JSON export.
- Documentation, runbooks, and knowledge transfer.
SCOPE AND LIMITS
ALMA assembles evidence and drafts a record for a qualified reviewer to sign. It does not classify variants, call variants, or predict variant effect. It is not a certified in-vitro diagnostic. Baseline scope is germline coding SNVs and small indels in dominant conditions. Recessive phase, copy number, and somatic tiering are scope extensions.
Highlights
- An evidence record for your reviewer to sign. Every criterion arrives with its strength, sources, exact locators, and reasoning, versioned against the specification that produced it.
- Your specification loads as versioned data. An ACGS, CanVIG-UK, or ClinGen VCEP specification is a configuration file. We run your pinned SOP and show you the diff to the current release.
- Your boundary, your pipeline, proven per evidence code. Deployed in your AWS account on Amazon Bedrock, so your data stays inside and the pipeline is yours with no license. Built on a Loka Labs framework scored against ClinGen expert-panel curations criterion by criterion.
Details
Pricing
Custom pricing options
How can we make this page better?
Legal
Content disclaimer
Support
Vendor support
- Email: hcls@loka.com
- Website: <www.loka.com >
- Response time: 24 business hours for standard inquiries