Auto-VCF Interpretation Agent

Multi-Agent AI System for Genomic Variant Analysis

Retaj Alshaiabn-Etab Alotaibi-Sara Alsalmi-Jana Alghoraibi

Genomics · AI Agents · Scientific Research

The Challenge: Manual Variant Interpretation




Complex Data

VCF files contain dense variant information, including genes, genomic positions, and annotations, making the data difficult to process manually.

Time-Consuming Research

Finding relevant scientific and clinical evidence across databases such as PubMed and ClinVar is a manual, time consuming process.

Summarization Burden

Reviewing and synthesizing findings from multiple sources requires significant effort and increases the risk of inconsistencies or omissions.

Proposed Solution

The Auto-VCF Interpretation Agent is a multi-agent AI system that automates genomic variant interpretation. A VCF file enters a specialized pipeline where each agent handles a distinct stage.

Automate parsing and validation

Retrieve scientific evidence automatically

Generate structured, reviewed reports

System Architecture


Analysis Agent

The Analysis Agent is the entry point of the pipeline. It receives the uploaded VCF file and prepares structured data for downstream agents.

✅ VCF Validation

Checks file structure and format compliance

🧬 Gene Extraction

Identifies relevant genes from variant data

🔬 Variant & Annotation Extraction

Parses variant positions and associated annotations

Research Agent

The Research Agent takes genes and variants identified by the Analysis Agent and searches for scientific evidence across authoritative databases.

PubMed / ClinVar

Primary source for peer-reviewed genomic literature and variant evidence

Extensible Architecture

Designed to integrate Ensembl, and dbSNP in future iterations.

Evidence-Only Scope

Retrieves scientific context does not independently validate the VCF file

Writer & Critic Agents

✍️ Writer Agent

Combines analysis results and research findings into a clear, structured genomic report.

  • Summarizes scientific evidence
  • Produces a readable, organized output
  • Integrates variant data with literature findings

🔍 Critic Agent

Reviews the generated report before delivery, ensuring quality and accuracy.

  • Checks factual consistency and unsupported claims
  • Validates source relevance and evidence support
  • Returns PASS or NEEDS_REVISION with feedback

Demo

The system is accessible through a clean Gradio web interface, guiding users through each stage of the analysis pipeline.


Technology Stack

Python

Core programming language for all agent logic and data processing

CrewAI

Multi-agent orchestration framework coordinating specialized AI agents

Gradio

Web-based UI for VCF upload and report display

PubMed / ClinVar

Primary scientific evidence source for genomic literature retrieval

VCF Format

Standard Variant Call Format for genomic variant data input

Limitations & Future Improvements

⚠️ Current Limitations

Database coverage currently limited to PubMed / ClinVar

Designed for research and educational purposes, not clinical diagnosis.

LLM-generated results require human review.

🔭 Future Improvements

Ensembl & dbSNP

Expanded genomic reference data

Human-in-the-Loop Review


Extend Datatypes

Support for additional genomic data formats

Backend Integration

Add a dedicated backend for data management and processing.

Made with