Reusability report: Exploring the transferability of self-supervised learning models from single-cell to spatial transcriptomics

Self-supervised learning (SSL) has emerged as a powerful approach for learning meaningful representations from large-scale unlabelled datasets in single-cell genomics. Richter et al. evaluated SSL pretext tasks on modelling single-cell RNA sequencing (scRNA-seq) data, demonstrating the effective use of SSL models. However, the transferability of these pretrained SSL models to the spatial transcriptomics domain remains unexplored. Here we assess the performance of three SSL models (random mask, gene programme mask and Barlow Twins) pretrained on scRNA-seq data with spatial transcriptomics datasets, focusing on cell-type prediction and spatial clustering. Our experiments demonstrate that the SSL model with random mask strategy exhibits the best overall performance among evaluated SSL models. Moreover, the models trained from scratch on spatial transcriptomics data outperform the fine-tuned SSL models on cell-type prediction, highlighting a domain gap between scRNA-seq and spatial transcriptomics data whose underlying causes remain an open question. Through expanded analyses of multiple imputation methods and data degradation scenarios, we demonstrate that gene imputation would degrade SSL model performance on cell-type prediction, an effect that is exacerbated by increasing data sparsity. Finally, integrating zero-shot random mask embeddings into chosen spatial clustering methods significantly enhanced their accuracy. Overall, our findings provide valuable insights into the limitations and potential of transferring SSL models to spatial transcriptomics and offer practical guidance for researchers leveraging pretrained models for spatial transcriptomics data analysis.