LA-MARRVEL: A Knowledge-Grounded, Language-Aware LLM Framework for Clinically Robust Rare Disease Gene Prioritization

Imagine you are a detective trying to solve a very difficult mystery: Why is this patient sick?

In the world of rare diseases, the "culprit" is usually a tiny, broken piece of DNA (a gene variant) hidden among millions of other pieces. The problem is that modern DNA tests are like a giant magnifying glass that finds too many suspects. It might return a list of 1,000 possible genes, but only one is actually the criminal.

For years, doctors have had to manually sift through this massive list, reading thousands of medical papers and cross-referencing patient symptoms. It's slow, exhausting, and sometimes they miss the real culprit because the evidence is buried in complex text.

Enter LA-MARRVEL, a new AI tool designed to be the detective's super-smart assistant. Here is how it works, broken down into simple concepts:

1. The Two-Step Detective Process

Think of finding the right gene like finding a needle in a haystack.

Step 1: The Net (The First Stage): First, you throw a wide net to catch everything that might be a needle. The paper uses an existing tool called AI-MARRVEL for this. It's great at casting a wide net and catching 95% of the possible suspects, but the list is still too long to read carefully.
Step 2: The Interrogation (The LA-MARRVEL Stage): This is where the new system shines. Instead of just looking at a list of names, LA-MARRVEL acts like a brilliant detective who sits down with the patient's file. It reads the patient's specific symptoms (like "short stature" or "seizures") and compares them directly to the "mugshots" of the diseases associated with each gene.

2. The "Recipe" for Success: Context is King

The paper discovered that simply asking an AI, "Which gene is wrong?" doesn't work well. The AI gets confused or guesses based on famous genes it knows from the internet.

LA-MARRVEL changes the game by giving the AI a structured, detailed recipe (called a "prompt").

The Old Way: Telling the AI, "The patient has 'Stormorken Syndrome'." (This is like giving a detective a nickname; it doesn't tell them what the criminal actually looks like).
The LA-MARRVEL Way: Telling the AI, "The patient has short stature, muscle weakness, and bleeding from the nose." (This is like showing the detective a photo of the criminal's specific features).

By feeding the AI the actual symptoms rather than just disease names, it can make a much smarter connection between the patient and the gene.

3. The "Council of Judges" (Voting System)

Large Language Models (the brains behind LA-MARRVEL) can sometimes be a bit fickle. If you ask them the same question twice, they might give slightly different answers.

To fix this, LA-MARRVEL doesn't just ask the AI once. It asks the AI 10 times to rank the suspects. Then, it uses a special voting system (called Tideman's method) to combine those 10 opinions into one final, agreed-upon list.

Analogy: Imagine you have 10 expert judges. If one judge is having a bad day and ranks the wrong person #1, the other 9 judges will likely agree on the right person. The voting system ignores the outlier and picks the consensus winner. This makes the result much more stable and reliable.

4. The "Explainable" Report

In a hospital, you can't just say, "The computer says this gene is the answer." The doctor needs to know why.

LA-MARRVEL doesn't just give a rank; it writes a detective's report for every top gene. It says things like:

"We ranked Gene X as #1 because the patient has seizures and muscle weakness, which perfectly matches the known symptoms of Gene X. Also, the patient's DNA shows a broken copy of this gene from both parents, which fits the pattern of the disease."

It even explains why it demoted other genes:

"We moved Gene Y down to #20 because, although the DNA looks broken, the patient's symptoms don't match what Gene Y usually causes."

The Results: Why Does This Matter?

The researchers tested this on real patient data from three different hospitals.

Without the new system: The AI alone was only right about the #1 suspect 12–15% of the time.
With the old tools: They were right about 50–75% of the time.
With LA-MARRVEL: They got the #1 suspect right 78% of the time, and within the top 10 suspects, they were right 90–95% of the time.

The Bottom Line

LA-MARRVEL isn't trying to replace the doctor or the existing DNA tools. Instead, it acts as a smart filter that sits on top of the existing system. It takes the long, messy list of suspects, uses the patient's specific story to rank them, and gives the doctor a short, verified list with a clear explanation of why.

It turns a 7-year diagnostic odyssey into a much faster, more accurate process, ensuring that the right gene is found before the patient's condition worsens. It's not about the AI replacing the human; it's about the AI giving the human a superpower to see the truth hidden in the noise.

Here is a detailed technical summary of the paper "LA-MARRVEL: A Knowledge-Grounded, Language-Aware LLM Framework for Clinically Robust Rare Disease Gene Prioritization."

1. Problem Statement

Rare disease diagnosis is a time-intensive process where clinicians must match patient phenotypes (symptoms) to causal genes from vast, heterogeneous evidence sources. While modern sequencing identifies thousands of candidate genes, current computational tools often return long lists with high recall but low precision at the top ranks.

Limitations of Existing Tools: Traditional bioinformatics tools (e.g., Exomiser, LIRICAL) rely on structured features and statistical models, often struggling to synthesize unstructured clinical narratives or complex inheritance patterns effectively.
Limitations of Standalone LLMs: While Large Language Models (LLMs) excel at processing medical text, they suffer from bias toward well-studied genes, instability (outputs vary based on input ordering), and hallucinations. When used as standalone diagnostic tools, they fail to achieve the high recall necessary for clinical safety, recovering only ~12–15% of causal genes at the top rank.
The Gap: There is a need for a system that leverages the reasoning capabilities of LLMs to synthesize heterogeneous data (phenotypes, inheritance, variant evidence) while maintaining the robustness, recall, and interpretability required for clinical decision support.

2. Methodology: LA-MARRVEL Framework

LA-MARRVEL is a two-stage, knowledge-grounded reranking framework designed to sit on top of existing high-recall pipelines (specifically AI-MARRVEL). It does not replace initial retrieval but refines the candidate list.

A. Two-Stage Architecture

First Stage (High-Recall Retrieval): Uses AI-MARRVEL to generate an initial list of candidate genes (typically the top 100, denoted as $G=100$ ). This ensures the causal gene is retained in the pool with high probability.
Second Stage (LLM Reranking): LA-MARRVEL takes the top $G$ candidates and uses an LLM (Claude-4-Sonnet) to re-rank them based on a synthesized, knowledge-grounded prompt.

B. Knowledge-Grounded Prompt Construction

The core innovation lies in how the prompt is constructed. Instead of relying on disease labels alone, the system generates a structured context containing:

Patient Phenotypes: Detailed Human Phenotype Ontology (HPO) terms and descriptions.
Disease Annotations: Disease-specific HPO profiles and inheritance modes.
Variant Evidence: ACMG/AMP criteria data (e.g., ClinVar status, CADD scores, allele frequencies, zygosity).
Reasoning: The prompt explicitly asks the LLM to evaluate genes by comparing patient phenotypes against disease profiles and assessing variant pathogenicity.

C. Deterministic Aggregation (Ranked Voting)

To address LLM instability and stochasticity:

The system runs the LLM 10 times (10 independent inference runs).
It extracts the ordered gene lists from each run.
It aggregates these lists using Tideman's Ranked Pairs method (a Condorcet-compliant voting algorithm). This method locks pairwise victories in order of margin to produce a single, consensus ranking that is robust to individual run variations.

D. Explainability

For every gene, the system generates an auditable reasoning trace. This includes:

ACMG evidence codes (e.g., PM2, PP3).
Zygosity checks.
Phenotype matching scores.
Plain-language rationales explaining why a gene was promoted or demoted (e.g., "Phenotype mismatch" or "Inheritance model inconsistency").

3. Key Contributions

Architectural Innovation: Demonstrates that LLMs are most effective as second-stage rerankers rather than standalone diagnostic tools. This hybrid approach preserves the high recall of traditional tools while significantly boosting top-rank precision.
Contextual Grounding: Proves that structured, phenotype-rich prompts are superior to disease labels alone. Removing patient phenotype information caused the largest performance drop, highlighting the necessity of explicit context.
Stability via Aggregation: Introduces a deterministic voting mechanism (Tideman's method) that stabilizes LLM outputs, ensuring consistent rankings across repeated runs—a critical requirement for clinical deployment.
Transparency: Provides "explainable traces" that mimic manual curation steps, allowing clinicians to verify the logic behind gene prioritization.

4. Results

The framework was evaluated on three independent real-world cohorts: Baylor Genetics (BG), Deciphering Developmental Disorders (DDD), and the Undiagnosed Diseases Network (UDN).

Performance Gains:
- Recall@1: LA-MARRVEL achieved ~78%, compared to ~12–15% for prompt-only LLMs, ~50% for Exomiser, and ~31% for LIRICAL.
- Top-10 Recall: LA-MARRVEL reached ~90–95% across all cohorts, outperforming Exomiser (~~75–85%) and LIRICAL (~~70–80%).
- Improvement Magnitude: LA-MARRVEL improved top-rank accuracy by 10–30 percentage points over established tools and 45–50 points over LIRICAL.
Case Studies:
- Successfully corrected major ranking errors, such as promoting VARS2 from rank 26 to rank 1 and CHD8 from rank 13 to rank 1, with detailed reasoning explaining the shift based on phenotype concordance and inheritance models.
Ablation Studies:
- Removing patient phenotype data caused the most significant performance loss.
- Disease names alone were insufficient; explicit phenotype descriptors were required for effective reasoning.
Candidate Set Size: Performance saturated around G=100. Increasing the pool beyond this yielded diminishing returns while increasing noise.
Net Effect: While some cases were "harmed" (causal gene rank dropped), the number of "improved" cases significantly outweighed harmed cases across all cohorts.

5. Significance and Future Directions

Clinical Impact: LA-MARRVEL reduces the downstream review burden for clinicians by surfacing the correct causal gene at the very top of the list, where clinical attention is focused. It transforms a "black box" LLM into a transparent, auditable decision-support tool.
Workflow Integration: It is designed to integrate seamlessly into existing diagnostic pipelines without altering the initial high-recall retrieval step.
Limitations: The system is retrospective; performance may vary with incomplete phenotypic data. It also incurs higher computational costs due to multiple inference runs.
Future Work: The authors suggest adaptive ensembling (adjusting the number of runs based on case uncertainty), incorporating more structured evidence (e.g., family history), and conducting prospective clinical trials to measure impact on diagnosis time and accuracy.

In summary, LA-MARRVEL represents a significant step toward safe, reliable, and explainable AI in rare disease diagnostics, proving that combining high-recall bioinformatics with knowledge-grounded LLM reasoning can overcome the limitations of both approaches used in isolation.