RareCollab -- An Agentic System Diagnosing Mendelian Disorders with Integrated Phenotypic and Molecular Evidence

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Millions of children suffer from rare Mendelian disorders, often facing prolonged diagnostic odysseys due to limitations in current genomic sequencing interpretation.
• Existing computational tools struggle to integrate diverse data types like genomics, transcriptomics, and phenotypes for comprehensive rare disease diagnosis.

Purpose of the Study:

• To develop and evaluate RareCollab, an agentic diagnostic framework designed to bridge the gap in rare disease diagnosis by integrating multi-modal data.
• To improve the accuracy and efficiency of identifying causative variants in rare genetic disorders.

Main Methods:

• Developed RareCollab, an AI framework combining a quantitative Diagnostic Engine with Large Language Model (LLM)-based specialist modules.
• Integrated genomic data, transcriptomic sequencing (RNA-seq) data, phenotype information, variant databases, and scientific literature for analysis.
• Validated RareCollab on a benchmark dataset of Undiagnosed Diseases Network (UDN) patients with paired genomic and transcriptomic data.

Main Results:

• RareCollab achieved 77% top-5 diagnostic accuracy in a benchmark of UDN patients.
• The framework demonstrated a ~20% improvement in top-1 to top-5 diagnostic accuracy compared to existing variant-prioritization methods.
• RareCollab provides high-resolution, interpretable assessments by operationalizing multi-modal evidence.

Conclusions:

• RareCollab represents a significant advancement in rare disease diagnostics by effectively integrating multi-modal data using modular AI.
• The framework offers a scalable and accurate solution to reduce the diagnostic odyssey for children with rare genetic disorders.
• This approach highlights the potential of AI in revolutionizing the interpretation of complex genomic and transcriptomic information for clinical applications.