Skip to main content

Conference Program

Subpage Hero

Loading

LLM-Aided Script Development for APOE Gene Haplotype Analysis in Large-Scale Genomics  

Laboratory Genetics and Genomics
  • Primary Categories:
    • Laboratory Genetics
  • Secondary Categories:
    • Laboratory Genetics
Introduction:
The rapid development of Large Language Models (LLMs) like ChatGPT, Cursor AI, and Llama 3 offers promising opportunities to automate coding in bioinformatics, potentially accelerating research while reducing development time. This project explores the feasibility of using LLMs to generate a Python script for APOE gene haplotype calling, a common and key genetic factor genetic research. Our primary objective is to evaluate and gain practical experience with these AI tools, focusing on the quality, usability, and applicability of AI-generated code within a large-scale genomic data framework. This approach emphasizes a structured review of the outputs from three distinct LLMs, without aiming to formally benchmark their performance—a potential direction for future research. 

 

Methods:
This study utilizes real-world, pre-processed NGS datasets, including aligned reads and variant calls from the 1000 Genomes and a subset of the NIH All of Us cohort. These datasets ensure compatibility with downstream haplotype analysis workflows. LLMs, including ChatGPT, Cursor AI, and Llama 3, are tasked with generating Python scripts for APOE haplotype calling, focusing on zygosity determination and haplotype prediction. 

The scripts are integrated into a modular Nextflow pipeline to facilitate reproducibility and scalability. Key Python packages, such as pysam for BAM file manipulation, pandas for data handling, and argparse for command-line integration, are specified in prompts to LLM models to streamline data processing. 

A team of three senior bioinformaticians reviews the scripts, evaluating their structure, efficiency, and adherence to bioinformatics standards. After refinement, the optimized scripts are tested on cloud platforms, including Google Cloud and NIH’s All of Us infrastructure, to process over 100,000 genomes. The results are validated against reported haplotype calling methods, ensuring accuracy and robustness. 

 

Results:
Preliminary findings indicate that LLMs can generate functional bioinformatics scripts with minimal input, significantly expediting the initial development phase. However, expert refinement is necessary to optimize performance for domain-specific tasks. The integrated pipeline processes large genomic datasets efficiently, with haplotype predictions showing high concordance with established methods. Additionally, computational cost analyses demonstrate the feasibility of deploying these workflows in cloud environments for large-scale genomic studies. 

 

Conclusion:
This study demonstrates the potential of LLMs to accelerate bioinformatics script development, providing a strong foundation for complex genomic analyses. While LLM-generated scripts require expert adjustments for precision and scalability, they offer significant time savings in the development phase. Future research could focus on benchmarking LLMs across diverse bioinformatics tasks, improving their adaptability to heterogeneous data, and refining their integration into automated genomic pipelines

 

Agenda

Sponsors