We built GenomixIQ the other way round. Compute what can be computed. Use a language model only for what is left. Flag everything for a person. This post explains the split and where it still falls short.
The problem in one paragraph
The ACMG/AMP guideline from 2015 (Richards et al., Genetics in Medicine) sorts variants into five tiers: pathogenic, likely pathogenic, uncertain significance, likely benign and benign. It does this by combining evidence criteria drawn from population, computational, functional and segregation data. A lot of that work is lookup and arithmetic. Some of it needs judgment. Those two things should not be handled by the same component.
What we compute by rule
Six criteria: BA1, BS1, PM2, PVS1, PP3 and BP4. The inputs are Ensembl VEP annotation on MANE transcripts, gnomAD v4.1 filtering allele frequencies, Ensembl transcript structures and ClinGen reference snapshots. Where a ClinGen expert panel has published its own rules, we use the strengths that panel allows.
Because this step is deterministic, it can be tested. We compared it with codes that ClinGen expert panels applied to 13,447 classifications across 175 genes.
Agreement on cases both sides decided:
BA1 94.5%
PP3 92.5%
BP4 92.3%
BS1 90.0%
PM2 88.7%
PVS1 86.4%
his is agreement, not accuracy. A case only counts when both sides made a call, and it only covers genes with an expert panel. Remaining disagreements include gene specific PVS1 decision trees, panels that set their own SpliceAI or REVEL cutoffs, and curations made against older gnomAD versions. Full method: genomixiq.com/validation
What the language model does
It assesses the evidence the rules cannot reach, such as functional, segregation and phenotype evidence, and proposes a classification next to the rule output. We have not validated that part against expert panels. We treat it as a proposal.
Why not trust the model with citations
A 2025 bioRxiv benchmark of generative AI tools for variant literature retrieval found that hallucination robustness dropped for most tools when a variant had little published literature. The authors link less literature to a higher risk of unsupported claims. That is exactly the situation where a lab most needs help, so we do not let a model's citation stand on its own.
Scoring every result
Each result carries a G-ARVIS score on six dimensions: Groundedness, Self-confidence, Reliability, Variance, Inference cost and Safety. These are internal quality signals, not accuracy measurements.
The Safety rule is simple. A Pathogenic or Likely Pathogenic call with no citations scores zero and gets flagged for human review.
That check only confirms a citation is present. It does not confirm the citation supports the call. A geneticist still has to read it.
The scoring idea is open source. The engine is on PyPI as argus-llm under Apache 2.0. GenomixIQ applies its own genomics specific scoring on top.
What the output looks like
Field names only, not a real result:
{
"classification": "",
"acmg_criteria": ["", ...],
"pathogenicity_score": ,
"confidence": ,
"requires_human_review": ,
"citations": [...],
"garvis_score": {...}
}
Known limits
Classification takes tens of seconds per variant in our test setup. The first call to the free demo server can take up to a minute. Jobs live in server memory and can be lost on restart. The free trial is capped and for public or de-identified data only. Nothing here is for clinical use.
Try it
Free API trial, public or de-identified variants only: genomixiq.com. API docs: api.genomixiq.com/api/docs
If you work on variant curation and you find a case where the rules disagree with your panel, I want to hear about it. anil@ambharii.com
Sources
Richards S, et al. Standards and guidelines for the interpretation of sequence variants. Genet Med. 2015;17(5):405-424. https://pubmed.ncbi.nlm.nih.gov/25741868/
Benchmarking generative AI tools for literature retrieval and summarization in genomic variant interpretation. bioRxiv 2025. https://www.biorxiv.org/content/10.1101/2025.09.29.679212.full.pdf
About the author: Anil Prasad is co-founder and CTO of GenomixIQ and founder of Ambharii Labs. He builds production AI systems, and for the last stretch that has meant variant interpretation: rule computed evidence, model proposals, and a human in the loop. The code is early, the limits are on the website, and bug reports on edge cases are welcome. Reach him at anil@ambharii.com or anilsprasad.com.
Top comments (0)