GeneResolver: A Traceable Hybrid Multi-Agent Pipeline for Etiology-Aware Gene Prioritization

Background: Rare-disease diagnosis requires the integration of clinical phenotypes, genomic variants, inheritance patterns, and heterogeneous biological evidence. Existing gene-prioritization tools commonly assume that a disorder is caused by a single gene and that a curated Human Phenotype Ontology profile is already available. We developed GeneResolver as a traceable, privacy-oriented hybrid multi-agent pipeline that determines the likely etiology before applying multi-evidence gene prioritization. Cases are classified as single-gene, chromosomal, multifactorial, or non-genetic so that presentations unlikely to have a single-gene cause can enter mechanism-specific evidence branches instead of gene ranking. Language models support contextual interpretation, whereas ontology resolution, evidence retrieval, variant annotation, and numerical ranking remain deterministic and traceable. Results: On 196 published case narratives, classification of raw narratives without few-shot examples achieved four-class accuracy of 77.0%, a macro-averaged F1 score of 0.767, and single-gene-versus-rest branch-separation accuracy of 86.2%. On 4,677 phenotype-only cases from the PhEval benchmark, GeneResolver achieved precision at rank one of 0.4368 and a mean reciprocal rank of 0.5154, compared with published Exomiser values of 0.391 and 0.458, respectively. On the LIRICAL spiked-exome benchmark, the percentage of cases ranked first increased from 45.80% with phenotype data alone to 54.57% when phenotype and variant call format data were combined; mean reciprocal rank increased from 0.5465 to 0.6421. Conclusions: GeneResolver provides a reference architecture for integrating etiological routing, clinical-text interpretation, phenotype grounding, optional genomic-variant analysis, and traceable gene prioritization. The results support further evaluation of etiology-aware hybrid systems as clinical decision-support tools. GeneResolver remains a proof-of-concept intended to assist, rather than replace, expert clinical interpretation.