🧬 biology

GeneResolver: A Traceable Hybrid Multi-Agent Pipeline for Etiology-Aware Gene Prioritization

GeneResolver is a traceable, privacy-oriented hybrid multi-agent pipeline that improves rare-disease diagnosis by first classifying case etiology to route evidence appropriately, thereby outperforming existing tools in both phenotype-only and combined phenotype-genomic benchmarks.

Original authors: Ivana Vichentijevikj, Kostadin Mishev, Monika Simjanoska Misheva

Published 2026-09-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ivana Vichentijevikj, Kostadin Mishev, Monika Simjanoska Misheva

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

For families navigating the labyrinth of a rare genetic disease, the path to a diagnosis is often a long and weary journey. Even with modern technology that can read a person's entire genetic code, finding the specific error responsible for an illness remains difficult. The challenge is not just in reading the code, but in understanding what it means. Doctors must sift through thousands of genetic variations, looking for the one that matches a patient's unique collection of symptoms. This process relies heavily on a precise description of those symptoms, often translated into a standardized list of medical terms. However, turning a doctor's messy, handwritten notes or a complex story of a patient's history into that clean, standardized list is a slow, manual task. Furthermore, many tools used to find the cause of a disease assume the problem comes from a single broken gene. This assumption can lead doctors down the wrong path if the illness is actually caused by a large chromosomal error, a mix of many small genetic factors, or something entirely outside the genes.

A new system called GeneResolver, developed by researchers in North Macedonia, attempts to solve these problems by acting as a smart, privacy-focused guide for doctors. Instead of jumping straight to ranking genes, this system first reads the patient's story to decide what kind of biological problem it is facing. It asks a simple but crucial question: is this likely caused by a single gene, or is it something else? If the answer is a single gene, the system proceeds to rank potential genetic culprits. If the answer is something else, it stops the gene search and instead offers evidence about the actual cause, such as a chromosomal issue or a non-genetic factor. The system is designed to work locally on a hospital's own computers, ensuring that sensitive patient stories and genetic data never leave the building, a critical feature for protecting privacy.

The researchers built this system as a team of specialized digital assistants working together. One assistant reads the raw clinical notes and carefully removes any personal details like names or addresses, replacing them with generic placeholders. Another assistant translates the remaining medical story into a structured list of symptoms using a standard medical vocabulary. A third assistant looks at genetic data if it is available, checking for errors in the genetic code. A central coordinator then takes all this information and makes a judgment call on the type of disease. If the system determines the case fits a single-gene pattern, it moves to a ranking phase where it weighs the symptoms against known genetic associations and the severity of any genetic errors found. If the system decides the case is not a single-gene disorder, it routes the information to a different branch that gathers evidence relevant to that specific type of problem, avoiding the mistake of forcing a single-gene explanation onto a complex case.

In testing this approach, the researchers found that the system could successfully distinguish between single-gene cases and other types of disorders with high accuracy. When asked to classify raw patient stories without any prior examples to learn from, the system correctly identified the type of disorder in 77 percent of cases. More importantly, it correctly separated single-gene cases from all other types 86 percent of the time. This ability to route cases correctly is vital because it prevents the system from wasting time or giving misleading results when a gene search is not the right tool. The system also performed well at the core task of finding the right gene when one was likely the cause. On a large set of nearly 4,700 cases where only symptom descriptions were available, the system placed the correct gene at the very top of its list more often than other widely used tools. When genetic data was added to the mix, the system's ability to find the correct gene improved even further, placing the right answer at the top in over half of the cases.

The study also highlighted the importance of keeping data private. The system uses a large language model to understand the context of medical stories, but this model runs entirely on local hardware within the hospital. This means the sensitive details of a patient's life and health are never sent to an external server. The researchers tested different methods for cleaning personal information from notes and found that a specific local tool was the most effective at removing identifying details while keeping the medical meaning intact. By combining this local privacy protection with a flexible workflow that adapts to the type of disease, GeneResolver offers a new way to approach rare diseases. It does not claim to replace the doctor; rather, it acts as a traceable, evidence-based partner that organizes complex information and suggests the most likely paths forward, leaving the final decision and interpretation to the human expert.

Technical Summary: GeneResolver

Problem Statement

Rare disease diagnosis faces a critical bottleneck: while genomic sequencing has accelerated variant discovery, the interpretation and prioritization of these variants remain difficult. Existing gene-prioritization tools often suffer from two primary limitations:

Rigid Etiological Assumptions: Most pipelines assume a disorder is caused by a single gene and require a pre-curated Human Phenotype Ontology (HPO) profile as input. They often fail to distinguish between monogenic, chromosomal, multifactorial, or non-genetic etiologies, leading to biologically inappropriate gene rankings for non-monogenic cases.
Privacy and Data Governance: Many Large Language Model (LLM) assisted workflows rely on external cloud services, posing risks for sensitive clinical narratives and genomic data. There is a lack of architectures that can perform contextual interpretation locally while maintaining strict data privacy.

Methodology

GeneResolver is a traceable, hybrid multi-agent pipeline designed to integrate etiological routing, clinical text interpretation, phenotype grounding, and optional genomic variant analysis. The system operates on a modular architecture where a rule-based orchestrator (LangGraph) coordinates specialist agents.

Core Architecture and Agents

The pipeline processes inputs (free-text clinical narratives, HPO terms, VCF files, or specific variant/gene identifiers) through five distinct modes. The workflow is divided into deterministic and LLM-assisted components:

Ingestion & Privacy: A local de-identifier masks Protected Health Information (PHI) using the OpenMed filter. Crucially, demographic extraction and inheritance pattern inference utilize a locally deployed MedGemma-27B-IT model to ensure unmasked family history is never transmitted externally.
Phenotyping: A hybrid approach extracts HPO terms. A deterministic dictionary-based recognizer (FastHPOCR) provides high-speed extraction, while a secondary LLM pass enriches the profile by proposing missed phenotypes, which are then grounded to existing HPO terms.
Etiological Routing (The Core Innovation): Before gene prioritization, an LLM-based Etiology Classifier analyzes the raw clinical narrative to assign one of four labels: single-gene, chromosomal, multifactorial, or non-genetic.
- Single-gene cases: Proceed to multi-evidence gene prioritization.
- Other etiologies: Are routed to mechanism-specific evidence branches that retrieve relevant literature and associations without generating a gene ranking list, avoiding biologically invalid outputs.
Variant Processing: For cases with VCF data, variants are annotated locally using Ensembl VEP. An inheritance agent infers Mendelian patterns from family history. A deterministic ACMG/AMP 2015 rule engine classifies pathogenicity, which feeds back into the gene score.
Gene Prioritization: For single-gene cases, a multi-evidence fusion engine ranks genes using a weighted score ( $S_{fusion}$ $S_{f u s i o n}$ ). The score combines:
- Phenotype evidence (HPO likelihood ratios).
- Variant pathogenicity (ACMG/AMP, impact, rarity).
- Gene-disease associations (Open Targets).
- Pathway evidence (Reactome).
- Loss-of-function constraints (gnomAD).
- Literature support (PubMed).
- GWAS evidence.
- Note: Weights are dynamic; in phenotype-only mode, the variant channel receives zero weight.

Traceability and Determinism

The system maintains a strict separation of concerns: LLMs handle contextual interpretation (synthesis, etiology classification, reasoning), while ontology resolution, evidence retrieval, variant annotation, and numerical ranking remain deterministic. Every intermediate result is recorded and attributable to the specific agent that produced it, ensuring the pipeline is fully traceable.

Key Contributions

Etiology-Aware Routing: The system introduces a pre-prioritization step that classifies the disorder's etiology. This prevents the application of monogenic gene-ranking algorithms to chromosomal or multifactorial cases, a common failure mode in existing tools.
Privacy-Preserving Hybrid Design: GeneResolver demonstrates a feasible architecture where sensitive clinical text and genomic data are processed locally using a deployed LLM (MedGemma), while only non-identifying, limited query terms are sent to external databases for evidence retrieval.
Traceable Multi-Agent Pipeline: The design explicitly separates deterministic evidence retrieval and scoring from LLM-based reasoning, providing a reference architecture for explainable and auditable clinical AI.
Unified Input Handling: The system accepts a flexible range of inputs, from raw clinical narratives to structured VCFs, automatically selecting the appropriate processing mode.

Results

The system was evaluated on multiple benchmarks and component-specific datasets:

Etiology Classification: On a silver-standard dataset of 196 case narratives, the raw-narrative classifier (without few-shot examples) achieved a 77.0% four-class accuracy and a 0.767 macro-averaged F1 score. Crucially, it achieved 86.2% accuracy in the binary task of separating single-gene cases from all other etiologies (branch separation).
Phenotype-Only Gene Ranking: On the PhEval benchmark (4,677 cases), GeneResolver achieved a Precision@1 of 0.4368 and a Mean Reciprocal Rank (MRR) of 0.5154. This outperformed published values for Exomiser (Precision@1: 0.391; MRR: 0.458), PhenoGenius, Phen2Gene, and GADO.
Genome-Level (Phenotype + VCF): On the LIRICAL spiked-exome benchmark, combining phenotype and VCF data improved performance significantly. The percentage of cases with the correct gene ranked first increased from 45.80% (phenotype only) to 54.57%, and MRR increased from 0.5465 to 0.6421.
Component Performance: The selected de-identification model (OpenMed) achieved a recall of 0.772 for PHI removal. The FastHPOCR phenotype extractor achieved an F1 score of 0.729, outperforming ClinPhen (0.541).

Significance and Claims

The authors position GeneResolver as a proof-of-concept and a reference architecture rather than a clinically approved diagnostic tool. The paper claims:

Feasibility: It demonstrates that privacy-oriented, etiologically aware genomic decision support is feasible using local LLMs and deterministic pipelines.
Clinical Utility: The results support the further evaluation of such hybrid systems as clinical decision-support tools that can assist, but not replace, expert clinical interpretation.
Limitations: The authors explicitly state that the system has not undergone prospective clinical validation, usability testing, or regulatory assessment. The etiology classifier relies on a silver-standard corpus due to the lack of public benchmarks, and the de-identification models used are not yet ready for full clinical deployment.

In conclusion, GeneResolver offers a novel approach to rare disease diagnosis by addressing the "garbage in, garbage out" problem of applying monogenic algorithms to non-monogenic cases and by solving the privacy constraints of using LLMs in clinical genomics.