Credit: Pixabay

Technological advances within the past decade have allowed researchers to generate multiomic profiles within single-cells: gene expression and chromatin accessibility (SNARE-seq, sci-CAR), gene expression and protein epitopes (CITE-seq), or chromatin accessibility and protein epitopes (Pi-ATAC, ASAP-seq). While these methods enable multimodal analyses of gene regulation with single-cell resolution, the protocols are both technically challenging and expensive, potentially restricting their feasibility among researchers.

As an alternative approach to these multiomic methods, Wu et al.1 at Stanford University recently developed BABEL, a deep learning tool that can effectively translate one single-cell modality (chromatin accessibility) into a paired dataset (gene expression) for downstream analysis. The authors first trained BABEL on three cell types that had been separately profiled using single-cell (sc)ATAC-seq (chromatin accessibility) or scRNA-seq. They reported that BABEL was capable of using scATAC-seq data to robustly predict gene expression, maintaining a strong correlation with complementary scRNA-seq datasets and outperforming tools like ArchR or MAESTRO that impute gene activity from chromatin accessibility. Furthermore, BABEL-derived gene expression data preserved similar clustering and cell type-specific expression patterns observed in complementary scRNA-seq datasets. The authors also observed that BABEL could infer gene expression from scATAC-seq data in multiple human samples that were not used when training the algorithm, such as lymphoblastoma cell lines or clinical basal cell carcinoma isolates. Interestingly, BABEL also maintained a high level of accuracy when analyzing cerebral cortex or skin samples taken from adult mice, suggesting that it may be applicable for single-cell analyses across multiple species. While the authors primarily focused on translating human scATAC-seq results into analogous gene expression data, they also demonstrated that BABEL could work in reverse, inferring chromatin accessibility from scRNA-seq data. Similarly, BABEL could translate scATAC-seq input into protein epitope expression, highlighting its flexible framework.

Altogether, BABEL offers a robust and versatile approach to translating single-cell datasets between modalities. Given the practical and financial barriers to performing multiple single-cell experiments, BABEL represents a promising tool to make multimodal analyses more accessible to researchers and thereby maximize the potential of their data.