uniPort: a unified computational framework for single-cell data integration with optimal transport

Abstract

Single-cell data integration can provide a comprehensive molecular view of cells. Here we introduce uniPort, a unified single-cell data integration framework which combines a coupled Variational Autoencoder (coupled-VAE) and Minibatch Unbalanced Optimal Transport (Minibatch-UOT). It leverages both highly variable common and dataset-specific genes for integration and is scalable to large-scale and partially overlapping datasets. uniPort jointly embeds heterogeneous single-cell multi-omics datasets into a shared latent space. It can further construct a reference atlas for online prediction across datasets. Meanwhile, uniPort provides a flexible label transfer framework to deconvolute spatial heterogeneous data using optimal transport space, instead of embedding latent space. We demonstrate the capability of uniPort by integrating a variety of datasets, including single-cell transcriptomics, chromatin accessibility and spatially resolved transcriptome data. uniPort software is available at https://github.com/caokai1073/uniPort.

MIDAS: a deep generative model for mosaic integration and knowledge transfer of single-cell multimodal data

“…MULTIMODAL DATA GLUE [21] and uniPort [22] were developed for the integration of unpaired trimodal data (i.e., datasets involving single specific modalities).…”

“…MOFA [15] and WNN [13] have been proposed for the integration of bimodal data with complete modalities, and totalVI [16], sciPENN [17], Cobolt [18], and MultiVI [19] are designed for bimodal integration with missing modalities. Fewer trimodal integration methods have been developed; MOFA+ [20] has been proposed for trimodal integration with complete modalities, and GLUE [21] and uniPort [22] were developed for the integration of unpaired trimodal data ( i.e. , datasets involving single specific modalities).…”

Linking cells across single-cell modalities by synergistic matching of neighborhood structure

“…The generalized unsupervised manifold alignment algorithm (GUMA) (Cui et al ., 2014), which uses a local geometry matching term, and MAGAN (Amodio and Krishnaswamy, 2018), which uses two generative adversarial networks (GANs), are examples of tools that find matchings of the cells between the two datasets. Recent methods that embed the two modalities into a common latent space and then attempt to align the embedded low-dimensional manifolds are SCOT (Demetci et al ., 2022), Pomona (Cao et al ., 2022a), and uniPort (Cao et al ., 2022b) all of which employ Gromov-Wasserstein optimal transport for alignment, and MMD-MA (Singh et al ., 2020), which aims to minimize the maximum mean discrepancy between the data sets in the latent space. Several methods rely on deep neural architectures to solve the manifold alignment task (Zuo and Chen, ???…”

“…MMD-MA (Singh et al, 2020) UnionCom (Cao et al, 2020) SCIM (Stark et al, 2020) scMVAE (Zuo and Chen, ????) SCOT (Demetci et al, 2022) Pomona (Cao et al, 2022a) uniPort (Cao et al, 2022b) scDART (Zhang et al, 2021) GLUE (Cao and Gao, 2021) Synmatch Table 1: Methods for unsupervised multi-model data integration. This multi-modal integration problem can be generally framed in two distinct ways: (1) finding a discrete mapping between cells in the two modalities or (2) embedding the disjoint measurements into a continuous shared latent space representing the intrinsic cellular structures across cellular modalities.…”