Workflows

Overview

The AI Workspace offers multiple workflow types for analyzing single-cell RNA sequencing data. Each workflow serves different analytical needs and has specific requirements:

• Standard Workflow: Analyze a single dataset independently
• Co-Embed Workflow: Integrate multiple datasets into a unified embedding space
• Comparison Workflow: Compare your data against the CELLxGENE Census reference atlas

Standard Workflow

Analyze your dataset independently without reference to external data.

• Purpose: Single-dataset analysis and exploration
• Data Requirements: H5AD files with ENSEMBL gene IDs
• Available Models: PCA, scVI, TranscriptFormer
• Organism Support: Varies by model (see Data Requirements)

Learn more about Standard Workflow →

Co-Embed Workflow

Integrate multiple datasets into a unified embedding space for comparative analysis.

• Purpose: Combine and analyze multiple datasets together
• Data Requirements: Two H5AD files with compatible gene IDs
• Available Models: PCA, scVI, TranscriptFormer
• Integration: Automatic batch correction via Harmony (PCA) or de novo training (scVI)
• Classification: De novo label transfer from reference dataset to target dataset

Learn more about Co-Embed Workflow →

Comparison Workflow

Compare your data against the CELLxGENE Census reference atlas.

• Purpose: Contextualize cells against comprehensive reference data
• Data Requirements: H5AD files with human or mouse data only
• Available Models: scVI, TranscriptFormer only
• Gene Requirements: ENSEMBL IDs (ENSG or ENSMUSG prefixes)

Learn more about Comparison Workflow →

Co-Embed Workflow

The Co-Embed workflow integrates multiple datasets into a unified embedding space, enabling comparative analysis across different data sources.

Co-Embed is automatically triggered when you explicitly select the "co-embed" workflow type.

Processing Steps

The workflow combines datasets through an outer-join operation that creates a dataset_source column to track the origin of each cell. The integration approach varies by model:

PCA with Co-Embed

1. PCA Step: Performs principal component analysis on the combined dataset
   • Uses dataset_source as the batch column (automatically created from outer-join)
2. Harmony Integration: Applies Harmony batch correction to integrate the datasets
   • Corrects for batch effects between different data sources
   • Uses dataset_source as the batch column
   • Outputs integrated embeddings in X_pca

scVI with Co-Embed

• De Novo Training: When classification is also enabled, the workflow uses de novo scVI training instead of pre-trained inference
  • Trains a new model on the combined datasets
  • Uses dataset_source as the batch column
  • Runs for 400 epochs to ensure proper integration
  • Produces better integration quality for label transfer

Use Cases

• Compare cells from different experimental conditions
• Integrate data from multiple time points
• Combine datasets from different studies or labs
• Transfer labels from a reference dataset to a query dataset

Cell Type Classification

Automated cell type annotation transfers cell type labels to your data using either pre-trained Census models or de novo label transfer from a reference dataset.

Classification Methods

The system uses two different approaches depending on your workflow configuration:

Standard Classification (Pre-trained Census Model)

Available for Standard and Comparison workflows when using scVI or TranscriptFormer models.

• Method: Uses pre-trained models from the CELLxGENE Census
• Requirements:
  • scVI or TranscriptFormer models only (PCA not supported)
  • Human or mouse data
• Process:
  1. Generates embeddings using your selected model
  2. Applies pre-trained classifier trained on Census reference data
  3. Transfers cell type labels with confidence scores
• Embedding Keys:
  • scVI: Uses scvi embeddings
  • TranscriptFormer: Uses transcriptformer embeddings

Co-Embed Classification (De Novo Label Transfer)

Available when combining multiple datasets in a Co-Embed workflow.

• Method: Transfers labels from a reference dataset to a target dataset
• Requirements:
  • Two uploaded files
  • Reference dataset must contain a cell_type column in its metadata
  • Works with PCA, scVI, and TranscriptFormer models
• Process:
  1. Combines datasets via outer-join (creates dataset_source column)
  2. Generates integrated embeddings using your selected model
  3. Identifies reference dataset as dataset_2 (second file uploaded)
  4. Uses KNN classifier to transfer labels from reference to target cells
  5. Applies labels based on nearest neighbors in embedding space
• Embedding Keys (automatically selected based on model):
  • PCA: Uses X_pca (Harmony-corrected embeddings)
  • scVI: Uses scvi embeddings
  • TranscriptFormer: Uses transcriptformer embeddings
• Note: Currently requires cell_type column in reference data (future versions will support dynamic column selection)

Classification Output

Both methods produce:

• Predicted cell type labels for each cell in your dataset
• Confidence scores indicating prediction reliability
• Integration quality metrics (for co-embed workflows)

When to Use Each Method

• Pre-trained Census: Best for single-dataset analysis when you want to leverage comprehensive reference annotations
• De Novo Transfer: Best when you have a specific reference dataset with known cell types that you want to transfer to your query data

Workflow Selection