Ensemble learning models that predict surface protein abundance from single-cell multimodal omics data.

Area of Science:

• Single-cell biology
• Computational biology
• Proteomics and transcriptomics

Background:

• Single-cell protein abundance is crucial for understanding cell states but difficult to measure directly.
• Single-cell RNA sequencing (scRNA-seq) is a proxy but can be inaccurate due to various regulatory factors.
• Emerging multimodal omics technologies (e.g., CITE-seq) enable simultaneous RNA and protein profiling.

Purpose of the Study:

• To compare machine learning models for predicting single-cell protein abundance from multimodal omics data.
• To evaluate the performance of tree-based ensemble methods against neural networks.
• To enhance the accuracy and interpretability of protein abundance prediction models.

Main Methods:

• Utilized single-cell multimodal omics data (RNA and protein abundances).
• Compared deep neural network models with tree-based ensemble methods, including Random Forest.
• Employed Random Forest's feature importance scores for biological interpretation.

Main Results:

• Ensemble learning methods generally outperformed neural network models.
• Random Forest demonstrated the best overall prediction performance.
• Feature importance analysis provided insights into biological mechanisms driving predictions.

Conclusions:

• Ensemble learning, especially Random Forest, is highly effective for accurate protein abundance prediction from multimodal single-cell data.
• This study validates ensemble learning for large-scale multi-omics data mining and knowledge discovery.
• The findings pave the way for more reliable characterization of cell states using integrated omics approaches.