LA-MARRVEL: A Knowledge-Grounded and Language-Aware LLM Reranker for AI-MARRVEL in Rare Disease Diagnosis

TL;DR

LA-MARRVEL tackles the bottleneck in rare-disease diagnosis by pairing a knowledge-grounded, language-aware LLM reranker with a high-recall first-stage pipeline. It uses expert-built phenotype and disease context, multiple LLM runs, and Tideman's ranked-pairs voting to produce a stable, explainable gene ranking that outperforms Exomiser, LIRICAL, and naive LLMs across three independent patient cohorts. The approach yields high top-k precision while preserving near-ceiling recall and provides an explainer trace detailing phenotype fit, inheritance context, and ACMG-like evidence for each gene. The work demonstrates practical benefits for clinical review, highlighting the importance of phenotype-grounded prompts, ensemble aggregation, and integration with existing diagnostic stacks. Prospective evaluations and richer prompting sources are identified as future directions to further embed LA-MARRVEL in real-world clinical workflows.

Abstract

Diagnosing rare diseases requires linking gene findings with often unstructured reference text. Current pipelines collect many candidate genes, but clinicians still spend a lot of time filtering false positives and combining evidence from papers and databases. A key challenge is language: phenotype descriptions and inheritance patterns are written in prose, not fully captured by tables. Large language models (LLMs) can read such text, but clinical use needs grounding in citable knowledge and stable, repeatable behavior. We explore a knowledge-grounded and language-aware reranking layer on top of a high-recall first-stage pipeline. The goal is to improve precision and explainability, not to replace standard bioinformatics steps. We use expert-built context and a consensus method to reduce LLM variability, producing shorter, better-justified gene lists for expert review. LA-MARRVEL achieves the highest accuracy, outperforming other methods -- including traditional bioinformatics diagnostic tools (AI-MARRVEL, Exomiser, LIRICAL) and naive large language models (e.g., Anthropic Claude) -- with an average Recall@5 of 94.10%, a +3.65 percentage-point improvement over AI-MARRVEL. The LLM-generated reasoning provides clear prose on phenotype matching and inheritance patterns, making clinical review faster and easier. LA-MARRVEL has three parts: expert-engineered context that enriches phenotype and disease information; a ranked voting algorithm that combines multiple LLM runs to choose a consensus ranked gene list; and the AI-MARRVEL pipeline that provides first-stage ranks and gene annotations, already known as a state-of-the-art method in Rare Disease Diagnosis on BG, DDD, and UDN cohorts. The online AI-MARRVEL includes LA-MARRVEL as an LLM feature at https://ai.marrvel.org . We evaluate LA-MARRVEL on three datasets from independent cohorts of real-world diagnosed patients.

Figures (8)

• Figure 1: Schematic illustration of LA-MARRVEL AI-MARRVEL first generates high-recall, variant-bearing candidate genes with annotations. LA-MARRVEL then composes knowledge-grounded prompts using HPO terms, disease and gene summaries, and variant-level evidence, queries an LLM multiple times, and aggregates the resulting partial rankings using Tideman's ranked-pairs voting. The final output is a reranked gene list with an explainable trace that integrates phenotype match, inheritance, and ACMG-style variant assessment.
• Figure 2: Comparative Performance Analysis of LA-MARRVEL. (A) Barplot of Recall@K (K = 1, 5, 10, 100) comparing two prompt-only LLM settings (Claude and Claude-Thinking, both without AI-MARRVEL context) against classical phenotype-driven tools (LIRICAL, Exomiser) and LA-MARRVEL. Prompt-only LLMs recover only 12--15% of causal genes at Recall@1 and remain below 55% even at Recall@10, whereas classical tools reach 50--75% and LA-MARRVEL attains 78% at Recall@1 and 95---98% by Recall@10, demonstrating that an LLM alone is insufficient and that coupling it to a high-recall first-stage ranker is essential for clinically useful performance. (B) Recall@K (K=1-10) is shown for LA-MARRVEL, AI-MARRVEL, Exomiser, and LIRICAL across the BG, DDD, and UDN cohorts. LA-MARRVEL consistently achieves the highest recall at clinically salient low K (Top-1/Top-3) while maintaining near-ceiling recall by Top-10, demonstrating improved prioritization of causal genes over both the first-stage ranker and established phenotype-driven tools.
• Figure 3: Ratio of Improved and Harmed Case by Original Rank For each cohort (BG, DDD, UDN), bars show the fraction of cases in which LA-MARRVEL improved, harmed, or left unchanged the rank of the causal gene, stratified by its initial AI-MARRVEL position. LA-MARRVEL most often rescues cases where the causal gene was originally ranked lower $(\geq 3)$, while a modest fraction of harms is concentrated among cases where the baseline already ranked the causal gene at the very top.
• Figure 4: Performance Gain of Context Engineering Recall@K loss is shown when removing key phenotype and disease information from the reranker input. Removing all HPO information yields the largest loss in Recall@K, followed by removing disease phenotypes, disease names, and HPO text descriptions (while retaining their identifiers). These results highlight the critical importance of structured HPO phenotypes and disease context for optimal performance in LA-MARRVEL.
• Figure 5: Performance across different Top-$G$ candidate sets Barplots of Recall@K for LA-MARRVEL using varying numbers of first-stage candidate genes ($G = 10, 20, 30, 50, 100$). Performance monotonically improves with larger candidate sets, reflecting the value of high-recall upstream ranking. Gains are most pronounced at low $K$, where small $G$ restricts recovery of causal genes.