UNMIRI opens genomics schema and warns of AI variant-matching errors

UNMIRI has released an open-source FHIR Genomics R4 schema and a preprint showing that cosine-similarity search can collapse clinically distinct cancer variants into the same result. The company says a typed knowledge graph avoids that failure mode and returned zero wrong-variant retrievals on the same benchmark.

Why it matters: - Clinical AI tools that confuse one cancer variant with another can surface the wrong evidence, which can affect treatment selection and downstream decision-making. - UNMIRI’s release gives developers an open schema and a retrieval approach designed to keep variant identity exact before any ranking or search step.

What happened: - UNMIRI LLC published an open-source FHIR Genomics R4 schema for cross-vendor next-generation sequencing data. - The company also published a preprint titled “Cosine Similarity Conflates Clinically Distinct Cancer Variants: A Case for Typed-Graph Retrieval in Precision Oncology Decision Support” on bioRxiv. - The preprint is available at the preprint DOI. - The schema is available on GitHub and archived on Zenodo under DOI 10.5281/zenodo.20042352.

The details: - The study tested a retrieval method common in retrieval-augmented AI: cosine similarity of embeddings. - On a curated benchmark of clinically distinct variant pairs, every pair was treated as nearly identical at a 0.95 similarity threshold under two biomedical encoders. - The paper says that meant genuinely different variants were judged the same by the math. - UNMIRI’s platform resolves variant identity through a typed knowledge graph and matches standardized identifiers exactly before retrieval or ranking. - On the same benchmark, that approach reduced wrong-variant retrieval to zero. - The open schema maps reports from labs including Foundation Medicine, Tempus, Caris, Guardant, and Natera into one structured FHIR R4 Genomics representation. - The schema includes HGVS variant conventions, biomarker fields, and identifier-based citation patterns. - UNMIRI says the schema is a working implementation of the public FHIR Genomics specification that other teams can adopt, fork, or contribute to. - The company’s broader platform includes APIs for cross-vendor NGS interpretation, genomics-aware clinical decision support, variant-grounded trial matching, and prior-authorization decisions. - UNMIRI is pre-revenue and is onboarding design partners through synthetic-data sandboxes.

Between the lines: - The release argues that variant identity is not a similarity problem and should be handled as a deterministic matching problem. - That distinction matters because two variants in the same gene can require different drugs, so a “close enough” match can become a clinical error. - The preprint is also a critique of a broader pattern in medical AI, where embedding-based retrieval can work well for text but fail on precise biomedical entities.

What’s next: - UNMIRI is using the schema and retrieval layer to support its precision oncology infrastructure for healthcare software companies, biotech medical affairs teams, and clinicians. - The company said it will continue onboarding design partners while its platform remains in synthetic-data testing. - More information is available at UNMIRI’s website.