Developer

This pipeline is developed by BMAP team.

Workflow

Data Agent

1.1 Variant Filtering Module: Annotates genetic variants in VCF files using ANNOVAR and snpEff, and applies multi-level filtering based on HGMD, ClinVar, MutationTaster2021, SIFT, and PolyPhen-2. Genes unrelated to the phenotype are further filtered using LLM.
1.2 Variant Prioritization Module: Ranks candidate mutations using a negative binomial test-based prioritization algorithm.
1.3 Natural Language Explanation Module: Utilizes LLM to generate explanations of the ranking results, including pathogenic mechanisms, gene functions, and phenotype associations, facilitating downstream debate and clinical interpretation.

Knowledge Agent

2.1 Performs symptom-based reasoning on the top 20 candidate pathogenic genes from the result of Data Agent using LLM, integrating gene function, phenotype relevance, and gene-gene interactions, and outputs ranking results in JSON format.

Debate Agent

3.1 Employs a debate framework grounded in Bayesian reasoning, where the Knowledge and Data Agents debate over a deduplicated and randomized list of candidate pathogenic genes.
3.2 Coordinates iterative rounds of reasoning between agents based on patient-specific information and supporting evidence from the Data Agent. Within a maximum of N rounds (configurable parameters), it guides convergence of opinions and synthesizes a final interpretable ranking report incorporating both annotations and knowledge-based insights.

workflow diagram

Background

The patient presented with weakness in both lower limbs, frequent urination, pectus carinatum, rib valgus and lumbar kyphosis as indicated by magnetic resonance spectroscopy at the L1–L2 segment. The patient sought a genetic diagnosis and a precise treatment strategy.

Research Plan

Whole-genome sequencing was performed to identify the pathogenic genes and mutations responsible for the patient’s clinical manifestations. The goal was to accurately diagnose the genetic condition and enable targeted therapeutic interventions.

Results

The debate analysis suggested that the pathogenic mechanism in this case may involve a synergistic effect of multiple functionally relevant gene mutations. The GALNS mutation was most closely associated with the skeletal abnormalities and was identified as the primary contributor. The PKD1 mutation may be related to the patient’s urinary symptoms, while the HYDIN mutation could indirectly contribute by affecting ciliary function. Although AXIN1 lacked direct pathogenic evidence, its role in the Wnt signaling pathway suggests potential involvement in skeletal development. Mutations in other genes lacked clear phenotypic associations and pathogenic annotations and were thus ranked lower in pathogenic potential. This indicates that GALNS is the most probable pathogenic gene.

v1.0: variant annotation, candidate pathogenic genes, explanation

Test Data

The dataset RD147 includes whole-genome sequencing (WGS) data in VCF format and clinical phenotypic abnormalities. The WGS data integrates the clinical confirmed pathogenic variant (GALNS variant p.Ter523Gluext*?/c.1567T>G) identified in the patient with variants from healthy control individual.

Reproducibility

The pipeline utilizes code CRS1110254509596356608 and Docker images (SRS1106541051004596224, SRS1106607526105067520, SRS1106549768446029824, SRS1106547058258751488),both of which are stored in the BMAP repositories.

Pipeline

Genetic variants are extracted following the GATK Best Practices guidelines for short variant discovery, which can be found here. The variants are then filtered and ranked by the Data Agent, Knowledge Agent, and Debate Agent.

Usage

1. Select the application 'Genetic disease diagnosis (Debate)'

2. Click on the 'Analysis' button

3. Click on the 'WGS and Phenotype'

4. Fill in the request parameters

Note: the two file names should include the subfolder name and the extension name (eg. data/R147.vcf, data/R147_hpo.txt).

5. Upload the compressed data in gz or rar format.

6. View the results