Skip to content
TheChatArchiveThe Chat Archive
DeepSeekdeepseek-r1-0528· DeepSeekHugging Face LMArena Human Preference 140K

For single-cell RNAseq: I have data that were mutliplexed using HTO antibodies.…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Claudeclaude-sonnet-4-20250514-thinking-32k’s answer.
  1. User

    For single-cell RNAseq: I have data that were mutliplexed using HTO antibodies. Unfortunately (as expected, though), not all cells got tagged, so what is a proper strategy: normalize using all cells and then exclude untagged, or exclude untagged and then normalize? Please discuss both approaches

  2. DeepSeek Assistant

    When dealing with multiplexed single-cell RNA-seq data with untagged cells (negative for HTOs), the recommended strategy is to exclude untagged cells first, then normalize the remaining tagged cells. Below, I discuss both approaches, their implications, and rationale:

    Approach 1: Normalize Using All Cells, Then Exclude Untagged

    • Steps:
      1. Apply normalization (e.g., log-normalization, SCTransform, or library size correction) to the entire dataset (tagged + untagged cells).
      2. Remove untagged cells (and doublets) post-normalization.
    • Pros:
      • Preserves technical/biological variation from all cells during normalization.
      • Avoids batch effects if untagged cells are systematically different (e.g., due to sample-specific biases).
    • Cons:
      • Skews normalization: Untagged cells often represent low-quality cells (e.g., damaged, empty droplets, or apoptotic cells) with aberrant RNA content. Including them:
        • Distorts library size estimates (e.g., median-of-ratios in SCTransform).
        • Biases gene expression scaling (e.g., untagged cells may inflate/deflate mean expression).
      • Amplifies noise: Untagged cells dilute signal from biologically relevant cells, reducing resolution in downstream analysis (e.g., PCA, clustering).
      • Wastes resources: Normalization computations on cells destined for removal are inefficient.

    Approach 2: Exclude Untagged Cells First, Then Normalize

    • Steps:
      1. Identify and remove untagged cells (and doublets) using HTO demultiplexing (e.g., with Seurat::HTODemux, MULTIseqDemux, or hashedDrops).
      2. Apply normalization only to tagged cells.
    • Pros:
      • Focused normalization: Normalization is based solely on high-quality, sample-specific cells.
        • Avoids contamination from low-quality/non-biological signals.
        • Preserves sample-specific differences (e.g., biological variation between multiplexed samples).
      • Downstream accuracy: Improves clustering, DE analysis, and visualization by reducing noise.
      • Standard practice: Supported by tools like Seurat and best-practice workflows (e.g., Satija Lab).
    • Cons:
      • If untagged cells include rare biological populations (e.g., unannotated cell types), they are lost.
      • Risk of over-filtering if demultiplexing is too stringent (mitigated by adjusting HTO thresholds).

    Key Considerations & Recommendations

    1. Untagged Cells Are Typically Noise:

      • Most untagged cells arise from failed HTO binding, empty droplets, or dead cells. Retaining them adds technical artifacts.
      • Exception: If you suspect untagged cells include viable populations (e.g., in rare cell-type studies), validate their quality (mitochondrial %, gene counts) before deciding.
    2. Demultiplexing First Streamlines Workflow:

      • Demultiplexing (e.g., with Seurat’s HTODemux) should precede gene expression analysis. This:
        • Separates multiplexed samples.
        • Filters low-confidence cells.
      • Only tagged singlets should proceed to normalization.
    3. Normalization Method Matters:

      • Log-normalization: Excluding untagged cells prevents distortion of scaling factors (e.g., scale.factor in NormalizeData).
      • SCTransform: More sensitive to library size biases; untagged cells with low RNA content corrupt variance stabilization.
    4. Practical Workflow Outline (using Seurat):

      # Step 1: Create Seurat object with RNA + HTO assays
      seurat_obj <- CreateSeuratObject(counts = rna_counts)
      seurat_obj[["HTO"]] <- CreateAssayObject(counts = hto_counts)
      
      # Step 2: Demultiplex (exclude negatives/doublets)
      seurat_obj <- NormalizeData(seurat_obj, assay = "HTO", normalization.method = "CLR")
      seurat_obj <- HTODemux(seurat_obj, assay = "HTO", positive.quantile = 0.99)
      seurat_obj <- subset(seurat_obj, subset = HTO_classification.global == "Singlet")
      
      # Step 3: Normalize RNA data (only singlets)
      seurat_obj <- NormalizeData(seurat_obj)  # or SCTransform()
      

    Conclusion

    Exclude untagged cells before normalization in almost all cases. This approach:

    • Targets normalization to biologically relevant cells.
    • Minimizes artifacts from low-quality/background signals.
    • Aligns with established scRNA-seq best practices.

    Exception: Only retain untagged cells if they pass rigorous QC and are biologically relevant (e.g., validated via marker genes). Even then, analyze them separately from tagged cells.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.