Seurat WNN

WNN learns a unique weight for each modality in every cell, rather than applying a global weighting scheme. For a cell where RNA data is noisy but protein signal is strong, the protein modality receives higher weight in defining its neighbors. This is achieved by minimizing the within-cluster cross-modality prediction error using a weighted k-nearest neighbor graph.

The algorithm builds a shared nearest neighbor graph by first computing independent k-nearest neighbor graphs for each modality (e.g., RNA expression PCA space and protein ADT CLR space). It then learns modality weights that maximize agreement between these graphs, fusing them into a single weighted multimodal graph without requiring labeled training data.

Once the WNN graph is constructed, standard Seurat workflows apply seamlessly. The unified graph is used for Leiden clustering to identify cell populations and as input to UMAP for visualization. This produces a single clustering and embedding that reflects both transcriptomic and epitope-defined cellular heterogeneity.

WNN provides tools to inspect the learned modality weights across cell populations. Users can visualize the relative contribution of RNA vs. protein data to the definition of each cluster. This reveals which modality drives the identification of specific cell types—for example, protein markers may dominate for immune subsets where surface markers are definitive.

WNN is the core analytical engine for CITE-seq data within Seurat, simultaneously processing gene expression counts and antibody-derived tag (ADT) counts. It also supports other multimodal pairings, such as scRNA-seq with scATAC-seq, by applying the same weighted nearest neighbor framework to any set of modality-specific distance matrices.

WNN is implemented as a native workflow within the Seurat v4+ ecosystem. After creating a Seurat object with multiple assays, the analysis proceeds through a standard pipeline: FindMultiModalNeighbors() computes the weighted graph, and FindClusters() and RunUMAP() operate on the resulting wknn graph. All downstream differential expression and visualization tools remain accessible.

WNN constructs a single nearest-neighbor graph that integrates information from multiple assays (e.g., RNA-seq and ADT). By learning cell-specific modality weights, it down-weights noisy modalities for individual cells, resulting in more robust and biologically meaningful clusters than those derived from a single data type.

A primary application is the analysis of CITE-seq data, which simultaneously measures transcriptomes and surface proteins. WNN resolves cellular heterogeneity by leveraging protein markers for definitive cell-type identification while using mRNA for deep phenotyping, overcoming the sparsity of single-cell RNA data.

WNN enables the construction of multimodal reference atlases. New query cells can be projected onto a WNN-defined reference, facilitating automated cell-type annotation and the identification of novel states by comparing across multiple data modalities simultaneously.

The learned modality weights provide direct insight into the information content of each assay for every cell. This allows researchers to identify cell populations where RNA and protein measurements are discordant, revealing post-transcriptional regulation or other biological phenomena.

Beyond protein data, WNN can integrate scRNA-seq and scATAC-seq to link gene expression with chromatin accessibility. This joint analysis defines cell states by both their regulatory potential and transcriptional output, enabling the identification of distal regulators driving cell identity.

WNN can be extended to a supervised framework to identify multimodal biomarkers. By learning weights that best separate predefined classes, it pinpoints the exact features—whether transcriptomic or proteomic—that define a disease state or treatment response.