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

Expansive repositories of scRNA-seq data are now available. These are often analysed assuming that mRNA abundance reflects expression of the cognate protein. However, post-transcriptional/translational regulation and the sparsity of measurements in single-cell data make mRNA an inadequate proxy for protein. Methods to quantify surface proteins alongside scRNA-seq exist but are less widely adopted. Machine learning approaches for protein imputation from scRNA-seq data have been published, which learn transcriptome-wide patterns that predict protein expression where data for both is available. These models can then be applied to infer surface protein expression on scRNA-seq only data sets, increasing their utility.

We test 9 machine learning methods for predicting single-cell protein expression, comparing the accuracy between methods and compared to using cognate mRNAs alone. Overall, machine learning -based protein predictions across methods outperform direct inference from mRNAs, including cases where proteins absent by mRNA are successfully predicted by the wider transcriptome. When comparing models trained on restricted cell types and across different datasets/tissues, we find that the overlap in cell type composition of training and test data is an important determinant of prediction accuracy. We also compare computational resource requirements to guide method selection.

These results reiterate that single-cell mRNA abundance is not a reliable proxy of cognate protein expression and that whole-transcriptome based imputations can improve upon them given appropriately trained models. However, limitations to the generalisability of these methods persist, notably a requirement for highly similar training data, which may limit the current scope of applications.