New BLASE tool maps bulk RNA-seq samples onto single-cell trajectories

Researchers have built a tool that answers a question bulk RNA sequencing has never been able to settle on its own: where, along a developmental journey, do the cells inside a bulk sample actually sit? The method, called BLASE for bulk linkage analysis for single cell experiments, takes a single-cell RNA-seq dataset with a trajectory, slices that trajectory into pseudotime bins, and then uses Spearman correlation to find the bin whose gene expression profile most closely resembles a bulk RNA-seq sample. Bootstrapping supplies confidence intervals around each correlation, and a mapping is called strong only when the lower bound of the best bin’s correlation exceeds the upper bound of the next best bin. The result is published in BMC Methods as an open-access methodology paper.

The motivation comes from a long-standing gap between the two dominant transcriptomic technologies. Single-cell RNA-seq can reveal how individual cells change, for example during cell type differentiation, but it is expensive and technically demanding. Bulk RNA-seq is cheaper and easier, yet it averages across a population and cannot elucidate cell type specific trajectories. Deconvolution methods such as CIBERSORTx, DWLS and MuSiC can estimate the proportions of annotated cell types inside a bulk sample by building signature matrices from a single-cell reference, but the authors argue these tools are poorly suited to trajectories. Differences along a trajectory are gradual rather than discrete, so two points close in pseudotime can look nearly identical, and in the authors’ testing, cell type deconvolution techniques gave unrealistic results, for example estimating that pseudobulks from multiple windows over pseudotime were mostly comprised of cells from the first pseudotime bin.

Only one prior option existed for deconvolving bulk RNA-seq over a trajectory: MeDuSA, a mixed model-based method that estimates the distribution of cells over pseudotime. In the authors’ testing, however, MeDuSA could not always resolve the correct distribution of simulated samples based on a Caenorhabditis elegans single-cell dataset from its own paper, or of synchronised Plasmodium falciparum bulk RNA-seq samples, instead overestimating the period of pseudotime through which a bulk was distributed. BLASE takes a different approach. It discretises the pseudotime of a single-cell trajectory, pseudobulks the cells in each bin, and scores the similarity between each bin and a bulk sample using Spearman’s Rho of their normalised counts. The bin with the highest correlation is considered the match, and bootstrapping over genes determines whether that call is strong. The tool also offers heuristics for choosing hyperparameters, including a convexity measure of how distinct each bin is from its neighbours, and several gene selection strategies, such as gene peakedness selection, which favours genes whose expression peaks sharply at particular points in pseudotime.

The validation began with simulated data carrying a known ground truth. Using Dyngen, the team generated 2000 cells with 400 target genes, calculated pseudotime with Slingshot, grouped cells into five bins by pseudotime range, and created two replicate pseudobulks per bin. BLASE aligned every pseudobulk sample to its correct pseudotime bin with high correlation and strong calls. The competing tools were asked to name the cell category they predicted to have the highest proportion. CIBERSORTx was unable to make predictions for this dataset despite troubleshooting and was excluded. DWLS mapped a mean of 91.3 percent of cells to the correct bins and MuSiC 84.2 percent, while MeDuSA gave the highest proportion to the correct bin but assigned a mean of only 42.3 percent of cells, with a much wider distribution than expected.

Real data told a similar story. In a myeloid differentiation lineage from a published haematopoiesis single-cell dataset, split into five equal pseudotime bins and pseudobulked, BLASE correctly and strongly mapped each pseudobulked bin to the correct reference bin. DWLS predicted that every pseudobulked bin should map to the first bin, achieving a mean of only 20.9 percent correct mapping; CIBERSORTx managed 46.3 percent; MuSiC 75.0 percent; and MeDuSA 43.3 percent. The authors note that the other tools were designed for discrete cell types rather than the continuum of pseudotime, and that only BLASE could strongly map a bulk sample to one time point. Across the simulated and experimental benchmarks, the paper reports that BLASE correctly identified the correct pseudotime bin for 10 pseudobulked pseudotime bins in simulated data, while other tools ranged from around 40 to 90 percent accuracy, and in experimental data BLASE correctly mapped all pseudobulked pseudotime bins compared with a range of around 20 to 75 percent for the alternatives.

The most striking demonstration came from malaria. The intraerythrocytic stage of the Plasmodium falciparum lifecycle repeats every 48 hours, progressing through ring, trophozoite and schizont stages, and both microarray and bulk RNA-seq time series of this cycle exist. The team mapped a microarray time course of synchronised parasites, sampled hourly, onto a single-cell reference of 6000 P. falciparum cells from the Malaria Cell Atlas. BLASE correctly identified the predominant cell type of the synchronised cells in 48 out of 48 samples, 30 of them strong calls. All strong calls made by BLASE were correct, and all non-strong predictions were correct as well. A gene list built from TradeSeq gave the highest number of strong mappings at 30, using every gene gave 27, and the two gene peakedness methods gave 15 and 21. Among the competing tools, DWLS equalled BLASE by also correctly mapping all 48 samples, CIBERSORTx missed one, while MuSiC and MeDuSA managed only 34 and 33 correct mappings respectively.

The same machinery works in reverse and in space. On 10x Visium spatial transcriptomics data of psoriatic skin, BLASE resolved the spatio-temporal process of keratinocyte differentiation, predicting developed granular keratinocytes in the thicker epidermal layer of lesions and less mature cells just below the outermost layer. The authors are candid about the limits: BLASE struggled to attribute spot mappings with high confidence, likely because each Visium spot contains few transcripts and a mixture of cell types, and on a second healthy skin dataset the roughly 50 micrometre spot size against an epidermis around 100 micrometres thick gave only two spots of resolution, yielding no convincing results. They also flag spot swapping, where RNA does not stay localised to its original spot, as a possible explanation for unlikely mappings deep in the dermis.

The third use case may prove the most consequential for working biologists. In many knock-out cell lines, development is altered, so a bulk RNA-seq comparison between mutant and wild type can be dominated by cell cycle differences rather than the true effect of the mutation. The team revisited a published study of a P. falciparum mutant, PB31, carrying a Piggybac insertion in the LRR5 gene, which renders it sensitive to heat shock. Mapping bulk samples onto the Malaria Cell Atlas reference revealed developmental shifts: the NF54 wild type under normal conditions mapped to bin 3 while heat shock mapped to bin 4, and the PB31 mutant under normal conditions mapped to bins 1 and 8, a difference as large as 18 hours given the 48-hour lifecycle. Under heat shock, PB31 sat at a similar point to NF54 under the same condition. Because the lifecycle is cyclic, heat shock itself can synchronise the parasites, and the mapping scores suggested the heat shock samples were indeed more synchronised.

These discrepancies matter for differential expression analysis. Comparing PB31 under normal and heat shock conditions returned 345 differentially expressed genes, while a single-cell differential analysis between the relevant pseudotime bins returned 282 genes attributed to cell cycle differences. The overlap between the two lists was 142 genes, which the authors attribute to developmental rate differences, leaving 203 genes in the corrected bulk list that should represent the true signal better. Gene ontology enrichment on the corrected list revealed new terms, including lipid metabolic process and other metabolic processes, that had been obscured before correction, because a smaller gene list faces less stringent multiple testing correction. The authors propose BLASE as a more general solution than existing malaria-specific correction methods, which rely on discrete references and the cyclic regulation of blood stage gene expression that most other cells and pathogens do not share.

BLASE is released as an open-source R package under the GPL3 license and is available through Bioconductor, with source code on GitHub and reproducibility data on Zenodo. It is fast: in a haematopoiesis example with 5573 cells, 19178 genes and 5 samples, BLASE finished within a few seconds while DWLS took over 25 minutes, and its computation scales with bins times samples times bootstrap iterations, with parallel execution available for large jobs such as mapping more than 1000 Visium spots. The authors position the tool around three uses: annotating single-cell data from existing bulk RNA-seq, identifying the progress of bulk data through a process captured in single-cell data, and correcting developmental differences in differential expression analysis. They caution that BLASE, like every deconvolution method evaluated, depends on the reference dataset being correct and appropriate, and that pseudotime is not necessarily linearly related to real time, a caveat that applies to trajectory inference methods generally. With atlas-scale reference datasets proliferating, the approach is likely to become more robust, letting researchers extract trajectory-level insight from bulk experiments without paying for new single-cell ones.

Subject of Research: A computational method that maps bulk RNA-seq samples onto single-cell pseudotime trajectories using correlation and bootstrapping

Article Title: BLASE: bulk linkage analysis for single cell experiments – teasing out the secrets of bulk transcriptomics with trajectory analysis

Article References: McCluskey, A., Kettlewell, T., Smith, A. M., Kundu, R., Gunn, D. A., & Otto, T. D. (2026). BLASE: bulk linkage analysis for single cell experiments – teasing out the secrets of bulk transcriptomics with trajectory analysis. BMC Methods, 3(1), Article 29. https://doi.org/10.1186/s44330-026-00082-7

Image Credits: AI Generated

DOI: 10.1186/s44330-026-00082-7

Keywords: BLASE, bulk RNA-seq, single-cell RNA-seq, pseudotime, trajectory inference, deconvolution, Spearman correlation, bootstrapping, Plasmodium falciparum, keratinocyte differentiation, spatial transcriptomics, differential expression