Robust probabilistic modeling for single-cell multimodal mosaic integration and imputation via scVAEIT.

Area of Science:

• Computational Biology
• Single-cell Multi-omics Analysis
• Bioinformatics

Background:

• Single-cell technologies allow joint profiling of multiple omics, revealing cellular regulatory complexity.
• Integrating datasets with shared and exclusive features (mosaic integration) presents challenges in imputation and latent space alignment.
• Existing matrix factorization methods struggle with nonlinear embeddings and accurate imputation of missing molecular data.

Purpose of the Study:

• To develop a robust method for mosaic integration and imputation of multimodal single-cell datasets.
• To address limitations of existing methods in handling nonlinear latent spaces and missing data.
• To create a self-consistent single-cell atlas from diverse omics measurements.

Main Methods:

• Proposed scVAEIT, a probabilistic variational autoencoder model for multimodal dataset integration and imputation.
• Utilized a missing mask for learning conditional distributions of unobserved modalities and features.
• Applied regularization techniques to prevent overfitting during imputation.

Main Results:

• scVAEIT accurately imputes missing modalities and features across biologically diverse cells, including those unseen during training.
• The model effectively adjusts for batch effects while preserving biological variation, yielding improved latent representations.
• Demonstrated significant improvements in integration and imputation across varied cell types, technologies, and tissues.

Conclusions:

• scVAEIT offers a flexible and accurate end-to-end solution for mosaic integration of multimodal single-cell data.
• The imputation capability enhances the construction of comprehensive single-cell atlases.
• This approach advances the robust integration and analysis of complex single-cell multi-omics datasets.