uniPort

uniPort integrates heterogeneous single-cell multi-omics and spatially resolved transcriptomic datasets into a shared latent space to enable joint embedding, cross-dataset gene imputation, and label transfer.

Key Features:

• Coupled Variational Autoencoder (coupled-VAE): Employs a coupled variational autoencoder to jointly embed diverse single-cell datasets (e.g., single-cell transcriptomics and chromatin accessibility) into a shared latent space that captures common and dataset-specific signals.
• Minibatch Unbalanced Optimal Transport (Minibatch-UOT): Incorporates minibatch unbalanced optimal transport to align heterogeneous dataset distributions and perform scalable alignment computations across large datasets.
• Gene Imputation: Constructs a reference atlas within the shared latent space to impute genes across different datasets.
• Flexible Label Transfer Framework: Uses an optimal transport plan for label transfer and deconvolution of heterogeneous spatial transcriptomic data rather than relying solely on latent-space nearest-neighbor mappings.

Scientific Applications:

• Integration of Multi-Omics Data: Combines single-cell transcriptomics with other modalities such as chromatin accessibility for joint analysis of cellular regulatory states.
• Spatial Transcriptomic Analysis: Deconvolutes spatially resolved transcriptomic data to map cell-type composition and gene expression patterns within tissue contexts.
• Reference Atlas Construction: Builds reference atlases in latent space to enable cross-study comparison and support gene imputation across datasets.

Methodology:

Uses a coupled variational autoencoder for joint embedding, applies Minibatch Unbalanced Optimal Transport to align distributions and compute optimal transport plans for label transfer, and constructs a reference atlas in the shared latent space for gene imputation.