aiDIVA - Diagnostics of Rare Genetic Diseases Using Large Language Models

Genome sequencing (GS) enables the accurate identification of genetic variants in most genomic regions and is rapidly transforming routine diagnostics for rare diseases (RD). While streamlined data generation is scalable, efficient prioritization and correct clinical interpretation of detected alterations remain a challenge, often requiring manual classification by experts with years of training. Hence, there is a need for AI-driven clinical decision support systems that assist clinical experts in identifying causal variants or, in case of large-scale re-analysis of unsolved cases, fully automate the process. To this end, many tools have been developed to estimate the impact of variants on protein function. However, only a small number of tools combine genomic data, variant annotations, and phenotypic data to diagnose cases. Here we introduce aiDIVA, an ensemble-AI featuring a hierarchically organized set of statistical and machine learning models trained on genomic and phenotypic data to identify the causal variant(s) among tens of thousands of genetic variants of a patient. aiDIVA generates pathogenicity classifications for each variant using a random forest AI model and an evidence-based score for dominant and recessive diseases. It combines these predictions with additional clinical metadata to prioritize and rank the most likely causal variants. aiDIVA uses large language models (LLMs) to further improve and explain the results. Finally, the aiDIVA-meta model combines all scores to generate a ranked list of variants. In a benchmark analysis on more than 3,000 diagnostically solved RD patients, the causal variant was included in 97% of the cases among the top-3 candidate variants reported by aiDIVA-meta. Unlike comparative methods, aiDIVA provides interpretable explanations for the best candidates.