Seurat Bridge Integration & Azimuth Mapping

Overview

Two related workflows for transferring annotations from a well-curated reference onto a query dataset:

Workflow	Use when
Azimuth (RunAzimuth)	Query is RNA-only (or has been WNN-integrated) and matches a SeuratData reference.
Bridge integration (FindBridgeTransferAnchors)	Query is ATAC-only, no matched RNA; a multiome bridge dataset connects ATAC space → RNA space.

Pre-reqs:

install.packages(c("Seurat", "Signac"))
install.packages("SeuratData")
install.packages("Azimuth")

---

Part 1 — Azimuth: RNA query → RNA reference

library(Azimuth)
library(SeuratData)

# InstallData("pbmcref")
reference <- LoadData("pbmcref")

# Map a query Seurat object (RNA assay must be present)
query <- RunAzimuth(query = seurat_obj, reference = "pbmcref")

head(query$predicted.celltype.l2)
head(query$predicted.celltype.l2.score)

DimPlot(query, group.by = "predicted.celltype.l2", reduction = "wnn.umap")

Azimuth runs an internal SCTransform on the query, finds anchors against the reference's pre-computed PCA, and transfers both labels and the reference UMAP coordinates.

Quality check:

• predicted.celltype.l2.score > 0.5 is a conventional minimum for "well-anchored" cells.
• mapping.score > 0.5 likewise — Azimuth provides two distinct scores.

---

Part 2 — Bridge integration: ATAC query → RNA reference

The bridge is a paired multiome dataset (RNA + ATAC on the same cells) that shares cell types with both the ATAC query and the RNA reference. Anchors are found by projecting the ATAC query into the bridge's LSI, then finding nearest neighbours in the bridge's RNA space.

library(Seurat); library(Signac); library(SeuratData)

# Load the three datasets
atac_query <- LoadData("pbmc_multiome_atac")
bridge     <- LoadData("pbmc_multiome")        # has RNA + ATAC
reference  <- LoadData("pbmcref")              # RNA only

# Preprocess the bridge — must process BOTH modalities
DefaultAssay(bridge) <- "RNA"
bridge <- SCTransform(bridge, verbose = FALSE)
bridge <- RunPCA(bridge, verbose = FALSE)

DefaultAssay(bridge) <- "ATAC"
bridge <- RunTFIDF(bridge)
bridge <- FindTopFeatures(bridge, min.cutoff = "q0")
bridge <- RunSVD(bridge)

# Find anchors via the bridge
anchors <- FindBridgeTransferAnchors(
  reference  = reference,
  bridge     = bridge,
  query      = atac_query,
  reduction  = "lsiproject",   # project ATAC query into bridge LSI space
  dims       = 2:30,            # skip LSI 1
)

# Transfer labels (and project query onto reference UMAP)
atac_query <- MapQuery(
  anchorset       = anchors,
  reference       = reference,
  query           = atac_query,
  refdata         = list(l1 = "celltype.l1", l2 = "celltype.l2"),
  reduction.model = "wnn.umap",
)

DimPlot(atac_query, group.by = "predicted.l2")

---

Part 3 — Bridge prerequisites

The bridge must:

1. Share cell types with both the query (ATAC) and the reference (RNA). A bridge of PBMC multiome cannot annotate a brain ATAC dataset.
2. Span the same protocol / batch domain as the query when possible. Strong batch effects between bridge and query degrade anchor quality.
3. Be preprocessed in both modalities — RNA through SCTransform + PCA, ATAC through RunTFIDF + RunSVD.

---

Common pitfalls

Pitfall	Resolution
Low mapping scores	Check that bridge cell types overlap query AND reference.
Missing features after RunTFIDF	Ensure peak set is consistent between bridge and query (or harmonise to a common peak set first).
Azimuth refuses to run	Confirm the reference name (pbmcref, tonsilref, lungref, ...) and that SeuratData::AvailableData() lists it.
Memory blow-up on bridge	The bridge can be downsampled — 10-20 k cells often suffices.

---

API quick reference

# Azimuth
RunAzimuth(query, reference)

# Bridge integration
FindBridgeTransferAnchors(reference, bridge, query,
                          reduction = "lsiproject", dims = 2:30)
MapQuery(anchorset, reference, query,
         refdata = list(...), reduction.model = "wnn.umap")

---

Resources

• Bridge integration tutorial: https://satijalab.org/seurat/articles/seurat5_integration_bridge
• Azimuth: https://azimuth.hubmapconsortium.org/
• Hao et al., Cell 2021 — Azimuth + reference-mapping paper.

---

When not to use

• Do not use without a multiome bridge dataset — see seurat-unpaired-cross-modality.
• Do not use for hierarchical label transfer with deep tree priors — see treearches-hierarchy-learning.

---

See also

• seurat-unpaired-cross-modality
• signac-chromatin-analysis
• seurat-multimodal-analysis
• treearches-hierarchy-learning