A deep generative framework with embedded vector arithmetic and classifier for sample generation, label transfer, and...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Rapid accumulation of multiple-source single-cell datasets necessitates advanced computational methods for integration and decomposition.
• Existing methods often lack flexibility in handling diverse analytical tasks and auxiliary information.

Purpose of the Study:

• To introduce inClust, a flexible deep generative framework for integrated clustering and data decomposition of single-cell data.
• To demonstrate the versatility of inClust in supervised, semi-supervised, and unsupervised learning scenarios.

Main Methods:

• inClust utilizes a deep generative framework incorporating auxiliary information embedding and latent space arithmetic.
• The framework's modular design allows for independent implementation of functional parts while ensuring runtime interrelation.
• The method supports supervised, semi-supervised, and unsupervised analytical modes.

Main Results:

• inClust demonstrated superior performance compared to existing data integration methods on benchmark datasets.
• The framework successfully performed conditional out-of-distribution generation (supervised), label transfer (semi-supervised), and spatial domain identification (unsupervised).
• inClust accurately expressed covariate effects, distinguished cell types, and segmented spatial domains.

Conclusions:

• inClust is an excellent general framework for harmonizing and decomposing multi-source single-cell data.
• The method offers a powerful and flexible approach for various computational biology tasks.
• inClust advances the analysis of complex single-cell datasets by enabling multi-task learning and data decomposition.