Abstract B037: Revolutionizing stem cell classification and genotypic profiling with large language models: scGPT-driven advances in single-cell genomics

Abstract The rapid advancement of deep learning technologies has revolutionized various scientific fields, and bioinformatics is no exception. Generative pretrained transformers (GPTs), known for their breakthrough success in natural language processing (NLP), are now being explored in the context of biological data analysis. This study introduces scGPT, a specialized adaptation of the GPT architecture tailored for single-cell genomics. By utilizing scGPT, we aim to address some of the most pressing challenges in single-cell data analysis, including high-resolution cell-type clustering, genotypic profiling, and efficient dose-effect correction. Single-cell RNA sequencing (scRNA-seq) technologies have opened new avenues for understanding cellular heterogeneity and molecular mechanisms at unprecedented resolution. However, this progress is hindered by issues such as batch effects, rarity, and the need for robust annotation methods that can be generalized across different datasets. Traditional methods like scVI and Harmony, often used alongside tools like Scrublet for deduplication, excel at specific tasks but lack the scalability and adaptability needed to handle new datasets without additional training. scGPT overcomes these limitations with its zero-shot learning capability, allowing seamless adaptation to unseen datasets while preserving biologically relevant variation. In this study, we used scGPT to perform dose effect correction, cell type clustering, and genotypic profiling across multiple publicly available scRNA-seq datasets. Comparative analyses were performed using state-of-the-art tools such as scVI, Harmony, and Scrublet. These data sets cover a variety of experimental conditions and biological systems and provide a robust test environment for evaluating scGPT performance. All analyses were performed on an Nvidia A100 GPU with CUDA to ensure computational efficiency, enabling high throughput and high-resolution processing. Our findings show that scGPT excels in batch integration, achieving superior cell alignment across batches while preserving critical biological signals. Unlike conventional methods, scGPT does not require retraining or fine-tuning when applied to new datasets, demonstrating its scalability and adaptability. In addition, the ability of scGPT to integrate different datasets and identify rare cell populations was validated through t-SNE and UMAP visualizations, illustrating its ability to distinguish between cell populations at high granularity clearly. Rare cell types often missed by traditional clustering methods were identified with remarkable accuracy, demonstrating the potential of scGPT in revealing subtle biological variations. Citation Format: Pei-Hsuan Chen, Yen-Jhen Chen, Pei-Lung Chen. Revolutionizing stem cell classification and genotypic profiling with large language models: scGPT-driven advances in single-cell genomics [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Functional and Genomic Precision Medicine in Cancer: Different Perspectives, Common Goals; 2025 Mar 11-13; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(5 Suppl):Abstract nr B037.