Technical Summary: Genosolver

Problem Statement

Despite advancements in molecular medicine and sequencing technologies (WGS/WES), a significant diagnostic bottleneck persists in rare disease care: the interpretation of vast genomic data remains challenging, leaving many patients without a diagnosis. Current computational approaches rely heavily on standardized clinical features, primarily the Human Phenotype Ontology (HPO). While HPO provides necessary standardization, it suffers from inherent limitations in capturing complex, unstructured clinical information. Specifically, HPO often fails to encode:

• Temporal and longitudinal symptom changes.
• Detailed family history and inheritance patterns.
• Negated findings (e.g., "absent EEG findings").
• Specific attributes of symptoms (e.g., "therapy-resistant myoclonic-astatic epilepsy").
• Therapeutic agents and specific diagnostic results not yet mapped to HPO terms.

Existing machine learning implementations often fail to bridge the entire diagnostic pipeline, typically limiting themselves to converting features into HPO terms (losing natural language nuance) or superficially integrating genetic data. There is a lack of an open-source pipeline that unifies phenotype extraction from unstructured narratives, rare disease diagnosis, and genetic variant prioritization into a single workflow capable of handling both standardized and non-standardized data.

Methodology

The authors present Genosolver, an integrated, multi-stage workflow utilizing Large Language Models (LLMs) and Large Reasoning Models (LRMs) to address these gaps. The system is designed to operate in data-secure local environments and consists of three primary modules:

1. Phenotype Extractor:

   • Processes unstructured electronic health records (EHRs), including clinical notes and genetic results.
   • Utilizes local Small Language Models (SLMs) and translation models (for German-to-English normalization) to extract phenotypic features, demographic data, and family history.
   • Benchmarks extraction by testing the model's ability to derive HPO terms from raw text, while also preserving the original unstructured narrative.

2. Disease Designator:

   • Acts as a coarse phenotype filter to reduce computational load.
   • Utilizes a vector database (FAISS) containing curated genotype-phenotype relations (derived from HPO-Jax, GeneReviews, and UniProt).
   • Employs BERT-based sentence transformers and Maximal Marginal Relevance (MMR) to retrieve relevant gene-phenotype associations based on the extracted features.

3. Variant Prioritizer:

   • Matches the filtered list of potential diagnoses with the patient's filtered genetic variants (SNVs, indels, and potentially CNVs/repeats).
   • Uses a Retrieval-Augmented Generation (RAG) approach, where the LLM/LRM accesses a curated literature knowledge base (GeneReviews, UniProt) to evaluate whether a specific variant explains the patient's phenotype.
   • Generates a reasoning summary explaining the ranking process, incorporating context such as age, sex, and inheritance patterns.

Experimental Setup:

• Models Evaluated: The study tested four local, open-source models: Llama-3.3-70B-Instruct, MedGo (Qwen-based), DeepSeek distilled Llama 70B, and GPT OSS 120B.
• Baselines: Performance was compared against Exomiser, a renowned variant prioritization tool, using various configurations (standard settings, ACMG classification filtering, and gene-variant score filtering).
• Datasets:
  • Retrospective Validation: 233 previously genetically solved cases (2023–2025) across four disease categories (neurodevelopmental, organ-specific, neuromuscular, hematological).
  • Reanalysis: 1,875 previously unsolved cases from the same center.
  • Proof of Concept: Small cohorts for repeat expansions (n=5) and Copy Number Variants (CNVs) (n=19).

Key Results

1. Variant Prioritization Performance

In the evaluation of 233 solved cases, the Genosolver workflow utilizing the GPT OSS 120B model and unstructured clinical notes (Experiment 1) demonstrated superior performance compared to Exomiser:

• Top 1 Hit Rate: Genosolver achieved 72% (Exomiser: 63%).
• Top 10 Hit Rate: Genosolver achieved 94% (Exomiser: 85%).
• This represents a 9% improvement over the baseline Exomiser tool.
• Other models (Llama, MedGo, DeepSeek) generally underperformed compared to Exomiser when used without the specific fusion strategies or unstructured data inputs.

2. Value of Unstructured Data

The study identified 522 phenotype terms extracted by Genosolver that could not be mapped to existing HPO terms. These included missed features, specific attributes, temporal components, and negative findings.

• The workflow utilizing unstructured narratives outperformed workflows using only expert-assigned HPO terms or machine-generated HPO terms.
• The model successfully utilized contextual information (family history, age, sex) in its reasoning. In 48 cases, family history was used to infer inheritance patterns (47/48 correct); in 43 cases, age was correctly applied; and in 6 cases, sex was correctly used to justify a diagnosis.

3. Large-Scale Reanalysis

A semi-automated reanalysis of 1,875 unsolved cases yielded an additional diagnostic rate of 1.7% (31 new diagnoses).

• Reasons for new diagnoses:
  • 21 cases: Missed due to previous rigid prioritization/filtering strategies.
  • 6 cases: Changed ClinVar annotations (VUS to pathogenic).
  • 2 cases: New gene-phenotype associations discovered.
• The Genosolver workflow (using Llama for efficiency) successfully identified these cases where standard pipelines had previously filtered them out.

4. Expansion to Other Variant Types

As a proof of concept, the workflow was adapted for:

• Repeat Expansions: Achieved a hit rate of 0.8 (k=1,3,5,10) on 5 cases.
• CNVs (Single/Multi-exon): Achieved hit rates of 0.8 (k=1) and 0.9 (k=3,5,10) on 10 cases.
• Multi-gene CNVs: Successfully retained 7 out of 9 relevant multi-gene CNVs by evaluating the collective relevance of genes within the CNV.

Significance and Claims

The paper claims that Genosolver offers a scalable, data-secure workflow that significantly increases molecular diagnoses in rare diseases by integrating unstandardized clinical data with local LLMs and reasoning models.

• Holistic Integration: Unlike previous tools, Genosolver unifies phenotype extraction, diagnosis, and variant prioritization in a single pipeline, capable of processing both HPO terms and raw clinical narratives.
• Beyond HPO Limitations: The study provides evidence that unstructured phenotypic features carry additional clinical value (e.g., temporal context, negations) that current HPO standards cannot capture, leading to higher diagnostic accuracy.
• Privacy and Adaptability: By utilizing local, open-source models and a RAG approach rather than fine-tuning, the system maintains data privacy (crucial for healthcare) and remains adaptable to rapidly evolving medical knowledge without the need for constant model retraining.
• Reanalysis Utility: The workflow demonstrates that LLM-based phenotypic prioritization can effectively uncover missed diagnoses in large cohorts of previously unsolved cases, particularly those missed due to rigid filtering strategies in traditional pipelines.

The authors conclude that this approach represents a viable path forward for expanding rare disease diagnostics to include complex variant types (CNVs, repeats) and leveraging the full depth of electronic health records.