An interpretable and adaptive autoencoder for efficient tissue deconvolution

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Deconvolution models extract cell-type information from bulk gene expression.
• Current methods face challenges with reference selection, data distribution shifts, and lack of interpretability.

Purpose of the Study:

• To present Sweetwater, an adaptive and interpretable autoencoder for efficient bulk sample deconvolution.
• To improve training data generation and establish a gold standard dataset for deconvolution evaluation.

Main Methods:

• Developed an adaptive and interpretable autoencoder (Sweetwater).
• Proposed an improved method for generating training data from FACS-sorted FASTQ files.
• Introduced a gold standard dataset for evaluating deconvolution approaches.

Main Results:

• Sweetwater leverages multiple reference data classes for efficient deconvolution.
• The new training data generation method reduces platform biases and outperforms single-cell references.
• Sweetwater adapts during training, uncovering biologically meaningful patterns and improving reliability.

Conclusions:

• Sweetwater offers an interpretable and adaptive solution for deconvolution.
• The study provides a new benchmark dataset for evaluating deconvolution methods.
• Sweetwater is expected to accelerate the analysis of high-throughput clinical data.