Machine learning predictions surpass individual mRNAs as a proxy of single-cell protein expression

Area of Science:

• Single-cell genomics
• Computational biology
• Proteomics

Background:

• Single-cell RNA sequencing (scRNA-seq) data is abundant but often assumes mRNA levels reflect protein expression.
• mRNA is an unreliable proxy for protein due to post-transcriptional/translational regulation and data sparsity.
• Existing methods for joint protein and scRNA-seq quantification are not widely adopted.

Purpose of the Study:

• To evaluate machine learning methods for predicting single-cell protein expression from scRNA-seq data.
• To compare the accuracy of machine learning imputation against direct mRNA inference.
• To assess factors influencing model generalizability and computational resource requirements.

Main Methods:

• Tested 9 machine learning algorithms for protein prediction from scRNA-seq.
• Compared prediction accuracy using machine learning versus cognate mRNA abundance.
• Evaluated model performance across different cell types, datasets, and tissues.

Main Results:

• Machine learning-based protein predictions significantly outperformed direct mRNA inference.
• Successful prediction of proteins with absent mRNA signals was achieved using whole-transcriptome data.
• Cell type overlap between training and test data was critical for prediction accuracy.

Conclusions:

• Single-cell mRNA abundance is an unreliable indicator of protein expression.
• Whole-transcriptome imputation models can enhance utility but require carefully selected training data.
• Generalizability remains a challenge due to the need for highly similar training datasets.