Leveraging Gpt4 for Enhanced Cell-Type Deconvolution in Immunological Research

---

Presentation Time: 11:30 AM - 12:45 PM
Poster Board Number: B941
Abstract ID: 6013

Presenting Author:
Reuben Sarwal , Bioinformatics Programmer at UCSF

Abstract:

This study seeks to validate the efficacy of GPT-4, an advanced machine learning language model, in the realm of computational biology, specifically for the task of cell-type deconvolution from whole blood RNA-seq data. The goal is to establish a novel, user-friendly approach that simplifies the complex process of identifying cell types within mixed cell populations, which is crucial for advancing immunological studies. For this purpose, we uploaded a full set of raw annotated TPM count data (n = 178,801 transcripts) and CIBERESORTx’s cell-type reference matrix. GPT-4 was then prompted to estimate cell-type proportions, and proportions were estimated using Non-Negative Least Squares (NNLS) and validated against standard deconvolution methods: CIBERSORTx (Stanford), Granulator (Vincent Kuettel et al.,) and xCell (UCSF). GPT-4 demonstrated a 70% agreement with established deconvolution methods, most notably CIBERSORTx, successfully identifying monocytes, macrophages, mast cells, and eosinophils. The comparison with conventional tools showed that GPT-4 can serve as a rapid and efficient alternative for cell-type deconvolution, potentially accelerating the pace of immunological research and discovery. It offers a fast and user-friendly method for dissecting complex transcriptomic data, empowering immunologists to conduct in-depth analyses with reduced computational overhead and time investment, paving the way for broader adoption of artificial intelligence in biomedical research.