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 provide deep biological insights but pose significant computational challenges.
• Existing subsampling techniques can improve efficiency but may compromise performance in downstream machine learning and deep learning (ML/DL) analyses.

Purpose of the Study:

• Introduce scValue, a novel cell-ranking approach for efficient and effective subsampling of large scRNA-seq data.
• To enhance ML/DL workflows by preserving critical biological signals and improving cell-type representation in subsamples.

Main Methods:

• Developed scValue, a method that ranks cells by 'data value' using random forest out-of-bag estimates.
• Prioritized high-value cells and oversampled cell types with greater data value variability.
• Benchmarked scValue on cell-type annotation, label transfer learning, cross-study label harmonization, and bulk RNA-seq deconvolution tasks.

Main Results:

• scValue consistently outperformed existing subsampling methods in automatic cell-type annotation tasks, achieving performance close to full-data analysis.
• Demonstrated superior preservation of T-cell annotations and accurate reproduction of T-cell subtype relationships in case studies.
• Evaluated on 16 public datasets, scValue showed fast execution, balanced cell-type representation, and distributional properties similar to uniform sampling.

Conclusions:

• scValue offers a robust and scalable solution for subsampling large scRNA-seq datasets in ML/DL applications.
• The method effectively preserves biological signals and enhances the performance of downstream analyses.
• scValue is available as an open-source Python package.