Machine learning for cell type classification from single nucleus RNA sequencing data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell/nucleus RNA sequencing (sc/snRNA-seq) generates high-resolution transcriptomic data for cell phenotyping.
• Classifying cells into specific types using sc/snRNA-seq data is challenging due to transcriptional similarities between related cell types and data sparsity.
• Matching cell types across different sc/snRNA-seq experiments remains a significant hurdle.

Purpose of the Study:

• To evaluate supervised machine learning methods for robust cell type classification using snRNA-seq data.
• To identify optimal machine learning models and feature selection strategies for granular cell type identification.
• To address challenges in distinguishing closely related cell types and improving cross-experiment cell type matching.

Main Methods:

• Applied supervised machine learning algorithms including logistic regression, support vector machines, random forests, neural networks, and LightGBM.
• Utilized snRNA-seq datasets from human brain middle temporal gyrus (MTG) and human kidney.
• Evaluated classification performance using an F-beta score, emphasizing precision to mitigate gene expression dropout effects.
• Investigated the impact of hyperparameter optimization and feature selection on model performance.

Main Results:

• Multinomial logistic regression emerged as the best-performing model for granular cell type classification across both tested datasets.
• An effective feature selection step was identified as the most critical factor for optimizing machine learning pipeline performance.
• The chosen F-beta score weighting successfully accounted for technical artifacts like gene expression dropout.

Conclusions:

• Supervised machine learning, particularly multinomial logistic regression coupled with robust feature selection, significantly enhances cell type classification accuracy from snRNA-seq data.
• Feature selection is paramount for maximizing the performance of machine learning models in sc/snRNA-seq analysis.
• These findings provide a data-driven approach to overcome limitations in current cell phenotyping techniques using sc/snRNA-seq.