Deep learning-based cell composition analysis from tissue expression profiles.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Cell deconvolution aims to determine the cellular makeup of tissues from bulk gene expression data.
• Existing methods often require extensive data preprocessing and feature selection, limiting their applicability.
• Single-cell RNA sequencing (scRNA-seq) provides high-resolution gene expression profiles crucial for improving deconvolution accuracy.

Purpose of the Study:

• To develop a robust and accurate deep learning model for cell deconvolution.
• To create a versatile tool that minimizes the need for complex data preprocessing.
• To enhance the understanding of tissue cellular composition across diverse biological contexts.

Main Methods:

• Scaden utilizes a deep neural network trained on scRNA-seq data to learn discriminative features.
• The model is engineered for inherent robustness against noise and batch effects common in gene expression data.
• A single trained Scaden network is applied to various data types, including bulk RNA-seq and microarray data.

Main Results:

• Scaden demonstrates superior precision and robustness compared to existing cell deconvolution algorithms.
• The model effectively deconvolves human and mouse tissue expression data from multiple sources.
• Scaden integrates information from diverse datasets, improving overall deconvolution performance.

Conclusions:

• Deep learning, exemplified by Scaden, offers a stable and flexible approach for cell deconvolution across various data types.
• Scaden's ease of use and broad applicability facilitate deeper molecular and cellular insights into development and disease.
• The Scaden software package and web application enable accessible analysis of public gene expression datasets.