bulk2scDiff: A Pseudobulk-Conditioned Diffusion Model for Bulk-to-Single-Cell RNASeq Generation

In the realm of genomic research, bulk RNA sequencing is the go-to method for analyzing large clinical cohorts, despite its tendency to blend transcriptional signals across cell populations and overlook cellular heterogeneity. Inferring this hidden diversity from pre-existing bulk transcriptomic data could enhance studies on large cohorts, but it is an underdetermined issue, as a single bulk profile could correspond to numerous underlying cellular populations.

Traditional computational deconvolution methods tackle this challenge primarily by estimating cell-type proportions or cell-type-averaged expression profiles, rather than resolving expression at the cellular level.

Researchers have now introduced bulk2scDiff, a conceptual diffusion framework that transforms bulk-to-single-cell inference into conditional generation of single-cell expression profiles from pseudobulk transcriptomic inputs. This approach was assessed on two breast cancer and acute myeloid leukemia single-cell RNA sequencing datasets, where pseudobulk profiles were derived from the single-cell data and served as conditioning inputs. The ground truth for each evaluation was provided by matching single-cell populations.

The results showed that bulk2scDiff accurately reconstructed populations from training samples and generated biologically relevant single-cell populations for unseen samples, demonstrating a strong ability to generalize to recurrent immune features. An additional pseudobulk-swap control further validated sample-specific conditioning, with each sample's pseudobulk yielding the closest match to its observed population in nearly every instance.

This groundbreaking work confirms the feasibility of conditional diffusion for generating single-cell populations from pseudobulk transcriptomic profiles, paving the way for future evaluations using clinical bulk RNA sequencing data.