Integrating single-cell RNA-seq datasets with substantial batch effects.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) data integration is crucial for biological insights.
• Conditional variational autoencoders (cVAEs) are popular for scRNA-seq integration.
• Existing methods struggle with harmonizing diverse datasets (e.g., cross-species, cross-protocol).

Purpose of the Study:

• To develop and compare novel regularization strategies for cVAE-based scRNA-seq data integration.
• To address limitations of current methods in handling substantial technical and biological variation.
• To propose an optimal cVAE strategy for complex biological systems.

Main Methods:

• Implementation and assessment of alternative regularization techniques for cVAEs.
• Comparison of VampPrior and Gaussian priors for data integration.
• Evaluation of cycle-consistency loss against adversarial learning (GLUE model).
• Assessment of Kullback-Leibler (KL) divergence regularization strength tuning.

Main Results:

• VampPrior significantly improves biological variation preservation and batch correction compared to Gaussian prior.
• Cycle-consistency loss outperforms adversarial learning in preserving biological information.
• KL regularization strength tuning alone is not recommended due to indiscriminate removal of biological and batch information.
• A novel model combining VampPrior and cycle-consistency loss demonstrates superior performance.

Conclusions:

• A new cVAE-based integration strategy combining VampPrior and cycle-consistency loss is proposed.
• This optimal strategy enhances downstream interpretation of cell states and biological conditions in complex scRNA-seq datasets.
• The proposed model, sysVI, is available in the scvi-tools package for broader accessibility.
• The regularization techniques offer potential improvements for other cVAE-based models.