Ascleon

Inspiration

Rare disease patients face an exhausting diagnostic odyssey, waiting an average of 5-7 years for answers. Existing genomic analysis tools like Exomiser excel at prioritizing gene variants but often miss crucial clinical context. I built Ascleon to bridge this gap between computational genomics and real-world clinical presentation, helping clinicians reach accurate diagnoses faster. And helping patients in understanding their problems independent of which language they speak and where they come from and which doctors they are able to see

What it does

Ascleon intelligently evaluates and reranks results from state-of-the-art genomic tools by incorporating four critical dimensions of patient data:

• Excluded phenotypes (symptoms explicitly ruled out in a patient)
• Age of onset (matching disease timing with patient presentation)
• Phenotype frequency (weighing how common symptoms are in each disease)
• Diagnostic recommendations (suggesting targeted tests from literature)

It also includes a genetic analysis of sequence data provided or recommends further tests to refine a ranked list of diseases given a phenotypic and genetic set of patient data.

How we built it

I created a multi-agent architecture using pydantic-ai where specialized agents work in parallel to analyze different aspects of patient data:

• HPOA agent analyzes phenotype frequency across diseases
• HPO agent processes phenotype data and relationships
• OMIM agent retrieves onset information
• Literature agent suggests diagnostic tests
• Visual Literature Agent that analysis graphs, and figures from medical papers

Challenges we ran into

I faced significant engineering challenges integrating multiple external data sources and APIs while maintaining real-time performance. The pydantic-ai implementation required careful refactoring to follow the appropriate agent-tool pattern. Managing the balance between detailed analysis and timely results proved particularly challenging with larger datasets, required me to implement fallback mechanisms for API timeouts and rate limits.

Accomplishments that we're proud of

I successfully developed a system that demonstrates measurable improvements in disease candidate rankings compared to baseline genomic tools. My implementation efficiently handles real-world phenopacket data and integrates multiple knowledge sources seamlessly. The architecture built is modular and extensible, allowing for easy integration of additional data sources and analysis methods as they become available

What we learned

Building Ascleon deepened the understanding of how AI can augment clinical decision-making through multi-faceted analysis. We learned that effective clinical tools must balance technical sophistication with usability and integrate diverse data sources while maintaining privacy and security. The project also highlighted the importance of handling uncertainty and providing explanations alongside AI-generated recommendations.

What's next for Ascleon

I planned to expand Ascleon's capabilities by integrating additional knowledge sources and improving performance with larger datasets. Our roadmap includes:

• Expanding literature analysis to incorporate full-text publications
• Developing a cloud-based deployment for broader clinical accessibility
• Creating an API for integration with electronic health record systems
• Implementing additional analysis modules for structural variants and copy number variations
• Conducting formal clinical validation studies to measure diagnostic improvement rates
• Include a visual model to check for variants and genes in Ensembl and analyse specific regions further to minimise mistakes and strengthen its outputs.

Built With

• pydanticai
• python