Variant classification is the bottleneck of clinical genomics. Every whole-genome sequencing run produces 4-5 million variants, of which 50,000-100,000 are coding or regulatory. Classifying each variant according to the ACMG/AMP guidelines requires integrating evidence from population databases, functional studies, computational predictions, and clinical literature, a process that takes an experienced molecular pathologist 15-30 minutes per variant.
Today we release NutraGenome's automated ACMG classifier, a deep learning system trained on 2.3 million expert-classified variants from ClinVar, LOVD, and proprietary clinical laboratory datasets. The model achieves 96.2% concordance with expert pathologist classifications on a held-out test set of 45,000 variants, matching inter-pathologist agreement rates reported in the literature.
The model architecture is a multi-modal transformer that integrates: variant-level features (amino acid change, conservation scores, protein domain annotations), population frequency data from gnomAD v4 and internal databases, functional evidence from high-throughput assays (MAVE data, splicing assays), literature features extracted from PubMed abstracts using a biomedical language model, and structural context from NutraFold-predicted protein structures.
In clinical validation at two partner institutions, the classifier reduced manual review time by 87%. Pathologists reported that the model's evidence summaries and confidence scores were helpful even for the 3.8% of variants where they disagreed with the automated classification, as the structured evidence display highlighted relevant information they might otherwise have missed.
Importantly, the classifier produces interpretable outputs. For each variant, it generates a structured evidence report listing the ACMG criteria that support its classification (e.g., PS1, PM2, PP3), the specific data sources used, and a confidence score. This transparency is essential for clinical adoption and regulatory compliance.
The classifier is integrated into the NutraGenome pipeline and runs automatically on all variant calls. Results are available in the platform's variant browser alongside manual review tools for cases requiring expert adjudication. We believe this combination of AI automation and expert oversight represents the optimal approach for clinical genomics in 2026 and beyond.