scValue: value-based subsampling of large-scale single-cell transcriptomic data for machine and deep learning tasks

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Large single-cell RNA sequencing (scRNA-seq) datasets offer deep biological insights but pose significant computational challenges.
• Existing subsampling techniques may compromise performance in downstream machine learning and deep learning (ML/DL) analyses.

Purpose of the Study:

• To introduce scValue, a novel cell-ranking approach for efficient and effective subsampling of large scRNA-seq datasets.
• To demonstrate scValue's superiority over existing methods in preserving biological signals for ML/DL tasks.

Main Methods:

• scValue ranks cells based on 'data value' derived from random forest out-of-bag estimates.
• Prioritizes high-value cells and oversamples cell types with greater data value variability.
• Benchmarked against existing subsampling methods on cell-type annotation, label transfer, label harmonization, and deconvolution tasks.

Main Results:

• scValue consistently outperformed existing subsampling methods in automatic cell-type annotation tasks.
• Achieved performance comparable to full-data analysis across various ML/DL models and datasets.
• Effectively preserved T-cell annotations, reproduced T-cell subtype relationships, and constructed reliable single-cell references for deconvolution.

Conclusions:

• scValue offers a robust and scalable solution for subsampling large scRNA-seq data in ML/DL workflows.
• Demonstrates fast execution, balanced cell-type representation, and distributional properties similar to uniform sampling.
• Available as an open-source Python package.