Seurat Unpaired Cross-Modality Integration

Overview

When you have separate scRNA-seq and scATAC-seq experiments from the same biological system (but not the same cells), Seurat can transfer labels between them by treating the gene-activity matrix from ATAC as a proxy for RNA. CCA is used to find shared structure in this synthetic shared feature space.

When to reach for this skill:

• No paired multiome data, no bridge dataset.
• Gene-activity matrix already computed in the ATAC object (via Signac::GeneActivity()).
• You want cell-type labels transferred from an RNA reference onto the ATAC cells, or RNA expression imputed for the ATAC cells.

When NOT to reach for this skill:

• Paired multiome — use signac-chromatin-analysis (WNN is strictly better when both modalities are on the same cell).
• A multiome bridge dataset is available — use seurat-bridge-integration.
• The Python ecosystem is preferred — use scglue-unpaired-multiomics-integration.

---

Part 1 — Find CCA transfer anchors

# Gene activity must already exist in atac_obj's "RNA" slot
# (i.e. assigned via atac_obj[["RNA"]] <- GeneActivity-based assay)

DefaultAssay(rna_ref) <- "RNA"
DefaultAssay(atac_obj) <- "RNA"   # gene activity scores

rna_ref <- FindVariableFeatures(rna_ref, nfeatures = 3000)

transfer.anchors <- FindTransferAnchors(
  reference        = rna_ref,
  query            = atac_obj,
  features         = VariableFeatures(rna_ref),
  reference.assay  = "RNA",
  query.assay      = "RNA",       # gene-activity scores
  reduction        = "cca",       # CCA for cross-modality (NOT rpca)
  dims             = 1:30,
)

Critical: use reduction = "cca". rpca does not handle the modality gap and gives near-random anchors.

---

Part 2 — Transfer labels

predictions <- TransferData(
  anchorset         = transfer.anchors,
  refdata           = rna_ref$cell_annotations,
  weight.reduction  = atac_obj[["lsi"]],
  dims              = 2:30,       # skip LSI 1 (depth-correlated)
)

atac_obj <- AddMetaData(atac_obj, predictions)

# QC: well-anchored cells have prediction.score.max > 0.5
sum(atac_obj$prediction.score.max > 0.5) / ncol(atac_obj)

The weight.reduction = atac_obj[["lsi"]] argument tells Seurat to use the ATAC native embedding (not gene activity) when weighting anchors — this preserves ATAC-resolved structure.

---

Part 3 — Expression imputation (exploratory)

refdata <- GetAssayData(rna_ref, assay = "RNA", layer = "data")[
              VariableFeatures(rna_ref), ]

imputation <- TransferData(
  anchorset         = transfer.anchors,
  refdata           = refdata,
  weight.reduction  = atac_obj[["lsi"]],
  dims              = 2:30,
)

atac_obj[["imputed_RNA"]] <- imputation

Imputed expression is suitable for population-level TF-gene hypotheses, not single-cell regulatory dynamics.

---

Part 4 — Quality and limitations

What works well	What does NOT
Label transfer (~90 % accuracy on Ma et al. benchmark)	Single-cell-resolution regulatory inference
Co-embedding in a shared UMAP	Rare cell states (too few anchors)
Population-level gene signatures	Enhancer-driven expression (gene activity misses distal regulation)
Validation layer for ChromVAR TF activity	Dynamic range — variance is compressed

For GRN inference: Imputed expression is suitable for population-level TF-gene relationships but NOT single-cell regulatory dynamics. For publication-quality GRN, use the SCENIC+ metacell approach (scenic-grn-inference). Seurat imputation can serve as a validation layer for SCENIC+ eRegulons.

Benchmark reference: Ma et al., Genome Biology 2023 — Seurat and scGLUE were the top two methods for unpaired RNA+ATAC integration.

---

Common pitfalls

Pitfall	Resolution
Random-looking label transfer	Confirm reduction = "cca", not "rpca".
All prediction.score.max near 0	Variable features in reference not present in gene-activity matrix; recompute gene activity with a matching annotation.
Imputed expression is smooth-looking	Expected — gene activity is already smoothed; do not over-interpret single-cell patterns.
LSI component 1 sneaks back in	Always pass dims = 2:30 to TransferData when using weight.reduction = atac_obj[["lsi"]].

---

API quick reference

# Anchor finding
FindTransferAnchors(reference, query, features, reduction = "cca",
                    reference.assay, query.assay, dims = 1:30)

# Label / expression transfer
TransferData(anchorset, refdata, weight.reduction = obj[["lsi"]],
             dims = 2:30)

---

Resources

• Cross-modality integration vignette: https://satijalab.org/seurat/articles/atacseq_integration_vignette
• Ma et al., Genome Biology 2023 (benchmark).
• Stuart et al., Cell 2019 (CCA anchoring methodology).

---

When not to use

• Do not use for paired multiome data — use signac-chromatin-analysis.
• Do not use for single-cell-resolution GRN inference — gene-activity smoothing is too coarse; use scenic-grn-inference.

---

See also

• signac-chromatin-analysis
• seurat-bridge-integration
• scglue-unpaired-multiomics-integration
• scenic-grn-inference