A machine learning framework for scRNA-seq UMI threshold optimization and accurate classification of cell types.

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cancer cell and tumor microenvironment diversity.
• Technical artifacts and stringent filtering (e.g., unique molecular identifier [UMI] cutoffs) can obscure biologically relevant cell populations, including rare or low-expression cells.

Purpose of the Study:

• To develop a machine learning framework for optimizing UMI thresholds in scRNA-seq data.
• To improve the accurate classification and recovery of diverse cell types, including rare populations, in cancer studies.

Main Methods:

• A novel machine learning framework was developed to train cell lineage and subtype classifiers using validated marker genes.
• The framework systematically assesses optimal UMI thresholds to balance cell recovery and classification accuracy.
• The method assigns accurate labels to cells recovered at lower read depths by applying the optimal UMI threshold.

Main Results:

• The framework was applied to breast cancer scRNA-seq datasets, demonstrating its effectiveness.
• In a breast cancer dataset, the minimum UMI threshold was lowered from 1500 to 450, increasing cell recovery by 49%.
• Classification accuracy exceeded 0.9, even with the reduced UMI threshold.

Conclusions:

• The developed framework offers a robust approach for determining optimal UMI thresholds in scRNA-seq data.
• This method enhances the recovery and accurate classification of cells, enabling more comprehensive downstream analyses in cancer research.
• It provides a roadmap for future scRNA-seq studies to maximize data utility and biological insights.