Adaptive resampling for improved machine learning in imbalanced single-cell datasets

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Machine learning models for single-cell transcriptomics offer biological insights.
• Current tools struggle with underrepresented or out-of-distribution cellular data.
• Effective representation learning is crucial for analyzing complex single-cell data.

Purpose of the Study:

• To introduce a generalizable Adaptive Resampling (AR) approach for single-cell transcriptomics.
• To enhance single-cell representation learning by addressing limitations in current models.
• To improve the performance of machine learning models on diverse single-cell data.

Main Methods:

• Developed an online, adaptive resampling strategy based on learned latent data structure.
• Integrated Adaptive Resampling (AR) concurrently with model training.
• Evaluated AR on gene expression reconstruction, cell type classification, and perturbation response prediction tasks.

Main Results:

• Adaptive Resampling (AR) significantly improved downstream performance across various tasks and datasets.
• The AR training approach enhanced the quality of learned cellular embeddings.
• Demonstrated superior performance compared to standard training methods in single-cell transcriptomic analysis.

Conclusions:

• Adaptive Resampling (AR) is a valuable technique for improving machine learning models in single-cell transcriptomics.
• AR effectively addresses challenges with underrepresented and out-of-distribution cellular data.
• The proposed method enhances representation learning and predictive accuracy for biological insights.