CSI-GEP: A GPU-based unsupervised machine learning approach for recovering gene expression programs in atlas-scale...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Exploratory analysis of single-cell RNA sequencing (scRNA-seq) often uses hard clustering on 2D projections (e.g., UMAP), which can distort data and involve arbitrary parameters.
• Gene Expression Programs (GEPs) offer a more robust way to model scRNA-seq data, but existing methods lack scalability, consistency, and clear parameter selection guidelines.

Purpose of the Study:

• To develop a scalable, consistent, and GPU-based unsupervised learning approach for inferring gene expression programs (GEPs) from scRNA-seq data.
• To improve the accuracy and efficiency of GEP identification compared to existing methods, particularly for large-scale datasets.

Main Methods:

• Developed Consensus and Scalable Inference of Gene Expression Programs (CSI-GEP), a GPU-accelerated unsupervised learning algorithm.
• Evaluated CSI-GEP's performance on simulated and real-world atlas-scale scRNA-seq datasets, comparing it against state-of-the-art methods, including GPT-based neural networks.

Main Results:

• CSI-GEP successfully recovered ground truth GEPs in both simulated and real scRNA-seq datasets.
• The method demonstrated superior performance over cutting-edge techniques, including GPT-based neural networks.
• Applied to a 2.2 million-cell mouse brain atlas, CSI-GEP identified endothelial cell subtypes missed by other approaches.
• Analysis of an integrated human tumor and cell line atlas revealed cancer-specific mesenchymal-like GEPs.

Conclusions:

• CSI-GEP provides a powerful and scalable solution for unsupervised GEP discovery in large scRNA-seq datasets.
• The method enhances the ability to resolve cellular heterogeneity and discover novel biological insights, as demonstrated in brain and cancer atlases.
• CSI-GEP represents a significant advancement in the analysis of single-cell genomics data.