An interpretable and adaptive autoencoder for efficient tissue deconvolution

Deconvolution models are a powerful tool for extracting cell-type-specific information from bulk gene expression profiles. Current methods leverage advanced machine learning models and high-resolution sequencing, like single-cell RNA-sequencing, showing promising results across diverse tissues and conditions. However, they still present important limitations: First, many depend on selecting a robust reference, which can strongly affect the deconvolution. Second, pseudobulk data used for training and real bulk RNA-seq samples often exhibit strong distribution shifts, which are currently unaccounted for. Finally, most deconvolution approaches behave as black boxes, which can compromise the reliability of the results. Here, we present Sweetwater, an adaptive and interpretable autoencoder that efficiently deconvolves bulk samples leveraging multiple classes of reference data. Moreover, we propose an improved way of generating training data from a mixture of FACS-sorted FASTQ files, reducing platform-specific biases and outperforming current single-cell-based references. Furthermore, we introduce a gold standard dataset to facilitate fair and accurate evaluation of deconvolution approaches. Finally, we demonstrate that Sweetwater adapts effectively to deconvolved samples during training, uncovering biologically meaningful patterns and enhancing result's reliability. Sweetwater is available at https://doi.org/10.6084/m9.figshare.29609180, and we anticipate it will expedite the accurate examination of high-throughput clinical data across diverse applications.