[Paper Review] COEM: Cross-Modal Embedding for MetaCell Identification

Mao, Haiyi, Jia, Minxue+3 more|arXiv (Cornell University)|Jul 15, 2022

Single-cell and spatial transcriptomics4 citations

TL;DR

COEM introduces a cross-modal variational auto-encoder to jointly embed scRNA-seq and scATAC-seq data into a shared low-dimensional space, enabling more accurate and well-separated metacell identification than single-modality methods. It reduces spurious negative peak-to-gene correlations by balancing chromatin accessibility and gene expression dynamics, improving regulatory inference and achieving faster clustering than SEACells.

ABSTRACT

Metacells are disjoint and homogeneous groups of single-cell profiles, representing discrete and highly granular cell states. Existing metacell algorithms tend to use only one modality to infer metacells, even though single-cell multi-omics datasets profile multiple molecular modalities within the same cell. Here, we present extbf{C}ross-M extbf{O}dal extbf{E}mbedding for extbf{M}etaCell Identification (COEM), which utilizes an embedded space leveraging the information of both scATAC-seq and scRNA-seq to perform aggregation, balancing the trade-off between fine resolution and sufficient sequencing coverage. COEM outperforms the state-of-the-art method SEACells by efficiently identifying accurate and well-separated metacells across datasets with continuous and discrete cell types. Furthermore, COEM significantly improves peak-to-gene association analyses, and facilitates complex gene regulatory inference tasks.

Motivation & Objective

• To address the limitation of existing metacell methods that rely on only one omic modality, which can introduce bias due to temporal lags between chromatin opening and transcription.
• To improve metacell identification accuracy by jointly modeling scRNA-seq and scATAC-seq data in a shared latent space.
• To reduce false negative peak-to-gene associations caused by the time lag between chromatin accessibility and gene expression.
• To enhance downstream gene regulatory network inference by leveraging more biologically consistent metacell representations.

Proposed method

• COEM employs a modified multi-view variational auto-encoder (VAE) to learn a shared low-dimensional latent representation from joint scRNA-seq and scATAC-seq data.
• The model assumes conditional independence of gene expression and chromatin accessibility given the latent representation and cell type, enabling joint generative modeling.
• A KNN graph is constructed from the shared latent space to capture local cell neighborhood relationships.
• Spectral clustering is applied to the KNN graph to identify high-granularity metacells with strong separation and compactness.
• Metacell features are aggregated from original single-cell profiles based on cluster assignments for downstream analysis.
• The method outperforms SEACells in metacell purity, separation, and computational efficiency, particularly in datasets with continuous and discrete cell types.

Experimental results

Research questions

• RQ1 Can joint modeling of scRNA-seq and scATAC-seq data improve metacell identification accuracy compared to single-modality approaches?
• RQ2 Does integrating multi-omic data reduce spurious negative peak-to-gene correlations caused by temporal lags in regulatory dynamics?
• RQ3 Can COEM identify more biologically coherent and well-separated metacells across both discrete cell types and continuous developmental trajectories?
• RQ4 How does COEM’s computational efficiency compare to state-of-the-art methods like SEACells in large-scale single-cell datasets?

Key findings

• COEM achieved a metacell purity of 0.91 on the sci-CAR dataset, significantly higher than SEACells-ATAC (0.67) and SEACells-RNA (0.74), indicating improved cell-type resolution.
• In the CD34+ bone marrow dataset, COEM reduced negative peak-to-gene associations to 1.3%, compared to 25.1% in SEACells-ATAC, aligning better with known biological ratios.
• For the GATA2 gene, COEM achieved a peak-to-gene correlation of 0.73, outperforming SEACells-ATAC (0.62) and vastly exceeding the single-cell level correlation of 0.1.
• Spectral clustering in COEM reduced running time by an order of magnitude compared to archetypal analysis on the same latent space across four datasets.
• COEM identified metacells with superior separation and compactness across multiple datasets, including those with continuous developmental trajectories.
• The method enables more accurate gene regulatory network inference, as demonstrated by stronger and more biologically plausible peak-to-gene correlations.