AIVA: An Agentic Platform for Phenotype-Aware Variant Analysis, Interpretation and Clinical Decision Support in Rare Disease

Background Rare disease diagnosis remains slow, with patients facing a diagnostic odyssey averaging 5 to 7 years. Genomic sequencing has shifted the bottleneck to interpretation: the challenge is identifying the pathogenic variant(s) among ranked candidates. Existing classifiers and phenotype-driven prioritization tools filter and rank candidates, but the review still occurs outside the platform, where analysts manually curate the literature and databases to weigh the evidence for pathogenicity. In addition, current automated classifiers apply generic ACMG/AMP rules rather than gene-specific expert-panel specifications, and most confine users to their own fixed annotations. We present AIVA (AI-powered Variant Analysis), an agentic platform that addresses this interpretive gap through a conversational interface, performing phenotype-aware prioritization and ACMG/AMP classification, grounding every assessment in tool-retrieved evidence with attached citations. Results Benchmarking variant classification on 8,387 ClinGen expert-panel-curated variants across 108 genes and 35 Variant Curation Expert Panels, AIVA reached a macro F1 of 80.5%, ahead of the rule-based classifiers BIAS-2015 (75.3%) and InterVar (60.6%). In an assembly stress test, AIVA detected all 644 corrupted genome-build records and self-corrected them, resulting in 68% of these mislabeled variants being correctly classified. Benchmarking prioritization across 1,396 simulated rare disease cases, AIVA ranked the causal gene first in 66.7% of cases, versus 59.5% for LIRICAL and 28.4% for Exomiser, by leveraging its real-time literature review and agentic capabilities. Conclusions To our knowledge, AIVA is among the first agentic platforms to perform both variant prioritization and ACMG/AMP classification, outperforming established tools on both tasks. By retrieving functional data, VCEP specifications, and phenotype evidence at analysis time, it applies the full set of ACMG/AMP criteria, including the literature-dependent ones that rule-based classifiers cannot evaluate. AIVA can identify and self-correct errors in the input like assembly mismatch, keeping results consistent. In one conversational, literature-grounded workflow, AIVA closes the gap between automated classifiers and expert tertiary review. AIVA is available at https://chat.aivaportal.com.