Empowering single-cell genomics with scGPT: Precision clustering and genetic insights into acute lymphoblastic leukemia.

Abstract

e18514 Background: Acute lymphoblastic leukemia (ALL) is a hematological malignancy characterized by impaired differentiation and abnormal proliferation of lymphoblasts. Despite advances in the understanding of its molecular mechanisms, accurate characterization of cellular heterogeneity and genetic alterations remains a challenge. The aim of this study is to take advantage of the scGPT, generative pre-trained transformer model adapted to the genome of one cell, to improve the distinction of cellular clusters and profile gene expression in all. We have used data files including B-cell subtypes and T-cells of ALL. Comparative analyzes with conventional tools, such as scVI and Harmony, have shown an excellent ability to maintain biological variability while providing accurate cell type annotations. Methods: This study utilized single-cell RNA sequencing (scRNA-seq) to obtain transcriptomic profiles from individual cells. Biological samples were processed through optimized cell capture and library preparation protocols to reduce technical noise and batch effects. NVIDIA GPU powered the computational environment with CUDA acceleration, and key software tools included scVI, scrublet, harmony, and scGPT. Highly variable genes were identified using scVI, leveraging Variational Autoencoders (VAEs) to detect nonlinear expression variability. Dimensionality reduction was performed with scVI, producing embeddings corrected for batch effects. Clustering analysis using the Leiden algorithm identified cell subpopulations and visualization tools such as UMAP and t-SNE displayed relationships among cell groups. Results: Visualization techniques, including UMAP and T-SNE, have confirmed the excellent ability of scGPT to identify rare cell populations and subsets otherwise missed by traditional methods. For example, in B cell subtypes, scGPT successfully distinguished naive B cells, memory B cells, and plasmablasts with high accuracy. Similarly, T cell subtypes such as regulatory T cells and cytotoxic T cells have been clearly characterized. Integration of scGPT with zero-shot learning enabled accurate profiling of gene expression across cell clusters and identified key differentially expressed genes relevant to acute lymphoblastic leukemia (ALL) pathology. These were genes associated with cell proliferation, immune evasion and metabolic pathways. Conclusions: Advanced scGPT modeling has enabled the prediction of potential therapeutic targets and paved the way for tailored interventions. Our research applies scGPT as a key tool to advance single-cell transcriptomic studies in ALL, offering a better understanding of its cellular and molecular complexities. This approach holds promise for improving therapeutic strategies and personalized medicine in hematologic malignancies.