USADAE: a deep learning approach to disentangle hidden covariates in RNA-seq data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• RNA-sequencing (RNA-seq) data analysis is challenged by hidden technical variations (confounders) that obscure true biological signals.
• Existing methods like SVA, PEER, and RUVSeq often fail due to their reliance on linear assumptions, inadequate for complex nonlinear confounding patterns.
• Current deep learning methods primarily address known batch effects, not the disentanglement of unknown confounders.

Purpose of the Study:

• To develop a novel computational framework for robustly distinguishing biological signals from unmeasured confounders in RNA-seq data.
• To create a method capable of handling complex nonlinear interactions between biological and technical variables.
• To enhance the accuracy of downstream analyses, including differential gene expression and expression quantitative trait loci (eQTL) studies.

Main Methods:

• Developed the UnSupervised Adversarial Deconfounding AutoEncoder (USADAE), an autoencoder framework utilizing adversarial learning.
• Implemented adversarial disentanglement to encode RNA-seq data into separate biological and confounder latent spaces.
• Validated the method through comprehensive simulations and real-world data applications in cancer genomics and eQTL studies.

Main Results:

• USADAE significantly outperformed existing methods in simulations for covariate extraction while preserving biological signals.
• Demonstrated the model's robustness and effectiveness across diverse real-world RNA-seq datasets.
• Successfully enabled accurate downstream differential expression and eQTL analyses after confounder correction.

Conclusions:

• USADAE offers a powerful new approach for addressing nonlinear confounding in RNA-seq data analysis.
• The method effectively disentangles hidden technical variability, leading to more reliable biological insights.
• USADAE shows significant promise for improving the analysis of various genomic datasets, including cancer and eQTL studies.