Automating Candidate Gene Prioritization with Large Language Models: Development and Benchmarking of an API-Driven Workflow Leveraging GPT-4

In this exploratory study, we developed an automated workflow that leverages Large Language Models, specifically GPT-4, to prioritize candidate genes for targeted assay development. The workflow automates interaction with OpenAI models and enables prompt creation, submission. It features customizable prompts designed to evaluate candidate genes based on criteria such as association with biological processes, biomarker potential, and therapeutic implications, which can be tailored for specific diseases or processes. Benchmarking experiments comparing the performance of the Application Programming Interface (API)-based automated prompting approach with manual prompting demonstrated high consistency and reproducibility in gene prioritization results. The automated method exhibited scalability by successfully prioritizing genes relevant to sepsis from the BloodGen3 repertoire, comprising 11,465 genes, distributed among 382 modules. The workflow efficiently identified sepsis-associated genes across the repertoire, revealing distinct gene clusters and providing insights into their distribution within module aggregates and individual modules. This proof-of-concept study demonstrates how LLMs can enhance gene prioritization, streamlining the identification process for targeted assays across various biological contexts. However, it also reveals the need for further validation and highlights the exploratory nature of this work due to scoring inconsistencies and the necessity for manual fact-checking. Despite these challenges, the automated workflow holds promise for accelerating targeted assay development for disease management and paves the way for future research.