PanelAppRex aggregates disease gene panels and facilitates sophisticated search

PanelAppRex is an openly available, harmonized resource and interactive search tool that aggregates over 58,000 curated gene-disease panel associations to facilitate sophisticated, natural language-based queries and seamless integration into bioinformatic pipelines for genomic diagnostics.

The Gist

Imagine you are a detective trying to solve a complex medical mystery. You have a patient with a strange set of symptoms, and you suspect a genetic culprit is behind it. To find the answer, you need to consult a massive library of "suspect lists" (called gene panels). These lists tell doctors which genes are known to cause specific diseases.

The problem? Right now, this library is a mess. The lists are scattered across different buildings (websites), written in different languages (formats), and some pages are torn or missing. If you want to search for a specific clue, you have to run from building to building, manually copying notes, and hoping you didn't miss anything. It's slow, exhausting, and prone to human error.

Enter PanelAppRex.

Think of PanelAppRex as a super-powered, magical librarian who has spent years gathering every single suspect list from every building, translating them all into the same language, and organizing them into one giant, perfectly indexed master database.

Here is how it works, broken down into simple concepts:

1. The "One-Stop Shop" Library

Instead of visiting Genomics England, Australia, and other sources separately, PanelAppRex brings over 58,000 gene-disease connections into one place.

• The Analogy: Imagine if you could ask a single librarian, "Show me all the books about heart problems in children," and they instantly pulled every relevant book from every library in the world and stacked them on your desk. That's what this tool does for genetic data.

2. Speaking Human, Not Robot

Usually, searching these databases requires you to know the exact technical code for a gene (like a secret password). PanelAppRex lets you speak in plain English.

• The Analogy: You don't need to know the ISBN number of a book to find it in a library; you can just say, "I'm looking for a story about a boy who can fly." Similarly, a doctor can type, "A child with a weak immune system and a skin rash," and PanelAppRex understands the context, even if the doctor doesn't know the exact gene names. It connects the dots between symptoms and genes automatically.

3. The "Detective's Test" (Benchmarking)

The authors tested this new tool to see if it actually works. They took 15 real-life medical cases where the answer was already known (like a solved mystery) and asked PanelAppRex to find the culprit gene using only the symptoms described in the case.

• The Result: It was a home run.
  • If you looked at all the lists the tool suggested, it found the correct gene 100% of the time.
  • Even if you just picked the "best guess" list that a human would choose, it was still 100% accurate.
  • On average, the correct gene showed up in 85 out of 100 lists the tool suggested.
• The Takeaway: It's like a detective who, when given a vague description of a crime, never fails to bring the right suspect to the interrogation room.

4. The "AI Assistant" (The Future)

The researchers also added a special feature using Artificial Intelligence (AI).

• The Analogy: Imagine the librarian doesn't just hand you the books; they also read them for you and write a one-page summary of the most important facts.
• PanelAppRex can now generate these summaries for doctors. It takes thousands of pages of complex scientific research and condenses them into a few bullet points explaining how a group of genes causes a disease. This is designed to help doctors and AI systems work together faster in the future.

Why Does This Matter?

For a long time, finding the right genetic test has been like trying to find a needle in a haystack while wearing blindfolds. PanelAppRex takes off the blindfolds and gives you a metal detector.

• For Doctors: It saves hours of work and reduces the chance of missing a diagnosis.
• For Researchers: It provides a clean, organized dataset they can plug into their computer programs to discover new patterns.
• For Patients: It means faster, more accurate diagnoses, which is the first step toward getting the right treatment.

In short, PanelAppRex is the universal translator and organizer for the chaotic world of genetic disease data, making it easier for humans and computers to work together to solve medical mysteries.

Technical Summary

1. Problem Statement

Disease-gene panels are critical resources for clinical diagnosis and genomic research, providing curated lists of genes associated with specific phenotypes. However, existing resources (such as Genomics England's PanelApp and PanelApp Australia) suffer from significant fragmentation:

• Data Silos: Panel data is distributed across multiple platforms and formats.
• Inconsistency: Annotations vary in structure and completeness, making programmatic aggregation difficult without data loss.
• Accessibility: Manual selection and cross-referencing of panels are labor-intensive, hindering integration into automated bioinformatic pipelines and AI-driven workflows.
• Lack of Standardization: There is no unified, machine-readable dataset that harmonizes gene identifiers, inheritance modes, and disease annotations across sources.

2. Methodology

The authors developed PanelAppRex, a harmonized dataset and interactive search tool designed to aggregate and standardize gene-disease panel data.

• Data Aggregation & Sources:

  • Core Data: Aggregated 58,592 entries representing gene-disease panel associations from sources including Genomics England (PanelApp), PanelApp Australia, and others.
  • Reference Data: Integrated external datasets for validation and enrichment, including gnomAD v4 (for allele frequencies), ClinVar (for variant classifications), UniProt (for protein mechanisms), and Ensembl (for identifier validation).
  • Identifier Standardization: Mapped all entries to standard identifiers (HGNC symbols, Ensembl IDs, OMIM IDs). Missing identifiers were programmatically recovered using the biomaRt R package.

• Implementation:

  • Language/Framework: Implemented in R.
  • Data Formats: Distributed as analysis-ready datasets in TSV (Tab-Separated Values) and RDS (R Data Serialization) formats, alongside a pre-compiled web interface.
  • Search Engine: Features a natural language-style search interface (HTML/JavaScript) that supports progressive filtering. It allows queries based on gene names, phenotypes, disease groups, Mode of Inheritance (MOI), and genomic locations. The system handles partial matching and aliases (e.g., mapping "Factor XII" to the gene F12).

• AI Extension (RAG):

  • To demonstrate AI applicability, the authors created an optional "Info" layer using Retrieval-Augmented Generation (RAG).
  • They retrieved mechanistic text from UniProtKB for each gene, merged it with panel metadata, and used an LLM (gpt-4.1-mini) to generate concise, clinician-facing summaries of shared mechanisms and phenotypes for each panel.

• Benchmarking Strategy:

  • Dataset: 15 published genetic case reports (5 immunology, 5 neurology, 5 cross-disciplinary) selected from JACI, Neurology, and PubMed.
  • Protocol: Queries were constructed using only pre-genetic clinical narratives (phenotypes) without knowledge of the causal gene.
  • Metrics: Evaluated retrieval accuracy by checking if the true causal gene was present in:
    1. The union of all returned panels.
    2. The user-selected "best-fit" panel(s).

3. Key Contributions

• Unified Dataset: Created the first large-scale, harmonized dataset of over 58,000 gene-panel associations with standardized identifiers and structured annotations (MOI, disease terms, literature support).
• Dual-Interface Tool: Provided both a user-friendly web interface for clinicians/researchers and a machine-readable dataset for bioinformatic pipelines.
• AI-Ready Architecture: Demonstrated the integration of structured panel data with LLMs to generate context-aware summaries, paving the way for RAG-based diagnostic support systems.
• Open Availability: The dataset, source code, and a demonstration page are openly available via Zenodo and GitHub, with a commitment to maintenance for at least two years.

4. Results

• Data Completeness: After automated recovery, core identifiers (HGNC/Ensembl) were complete for all valid gene entries. The dataset covers 451 distinct panels, with gene counts ranging from ~167 to ~1,675 genes per panel.
• Benchmarking Performance:
  • Causal Gene Recovery: In 100% of the 15 benchmarked case studies, the true causal gene was recovered when considering the union of all returned panels.
  • Precision: The causal gene was present in 85.6% of all individual panels returned by the search.
  • User Selection: In manual interface testing, the causal gene was present in the user-selected "best-fit" panel(s) in 100% of cases.
  • Alias Handling: The system successfully retrieved correct panels even when queries used aliases (e.g., "Factor XII") or phenotypic descriptions without explicit gene names.
• AI Summary Layer: The RAG workflow successfully compressed a 6.6-million-word gene knowledge base into ~135,000 words of panel-level summaries (approx. 50x compression), generating structured outputs (analysis, overview, bullet points) for each panel.

5. Significance and Future Impact

• Clinical & Research Utility: PanelAppRex significantly reduces the time and effort required to select appropriate gene panels for genomic analysis, supporting both manual diagnostic workflows and automated variant interpretation pipelines.
• Statistical Modeling: The structured nature of the data (specifically MOI and gene-disease links) enables future probabilistic modeling to estimate prior probabilities for variant classification, addressing a gap in current clinical genetics regarding "true negatives" and unobserved variants.
• AI Integration: By providing a structured, machine-readable foundation, PanelAppRex serves as a critical substrate for next-generation AI tools in genomics, including Retrieval-Augmented Generation (RAG) for diagnostic decision support.
• Limitations & Future Work: The authors acknowledge that not all genes are linked to panels (prioritizing high-confidence data) and that some panels may be over-represented due to historical biases. Future work aims to expand query types, integrate more variant annotations, and conduct broader real-world evaluations.

Conclusion: PanelAppRex bridges the gap between fragmented, expert-curated panel data and the needs of modern computational genomics, offering a robust, open-source solution for both human and machine-driven disease gene discovery.