Understanding the genetic activity of different cell types is crucial for studying diseases and developing treatments. However, most RNA sequencing experiments measure gene activity from a mixture of many cells, making it difficult to know which signals come from which cell type. To solve this problem, scientists use computational methods called deconvolution models, which estimate the contribution of different cell types in a bulk RNA sequencing sample.

Researchers from the University of Navarra have developed a new tool called Sweetwater that makes this process more accurate and easier to interpret. Unlike older approaches that act like “black boxes,” Sweetwater explains how it reaches its results, giving researchers more confidence in the findings. The method adapts during training, learns from multiple reference datasets, and even reduces biases that often occur when combining data from different experimental platforms.

Model overview

Sweetwater requires an input reference dataset in order to build pseudobulk samples prior training. This reference can leverage scRNA-seq or snRNA-seq matrices (mode SW) by subsampling cells and aggregate them to create synthetic profiles, in a similar way to modern deconvolution approaches. On the other hand, Sweetwater also provides two novel modes (mode SWbulk) that allow to create pseudobulk samples from FACS-sorted FASTQ files by subsampling and concatenating reads from different cell types (see the “Materials and methods” section). Moreover, it allows to account for mRNA quantity expression biases if the mRNA quantities of FACS-sorted cell types is available (mode SWbulk+). Sweetwater’s training applies three different phases sequentially, where the first two account for building a common latent space between real and simulated samples, and the third phase performs the deconvolution task. Finally, Sweetwater can deconvolve both microarray and bulkRNA-seq samples.

The team also created a new gold-standard dataset for evaluating deconvolution methods fairly. Their results show that Sweetwater not only outperforms current models but also uncovers biologically meaningful patterns that could advance research across a wide range of clinical applications.

By improving how scientists analyze bulk RNA sequencing data, Sweetwater has the potential to accelerate discoveries in fields from cancer research to personalized medicine.

Availability – Sweetwater is available at https://github.com/ML4BM-Lab/Sweetwater.