Batch Processing

Overview

Batch processing allows you to annotate all clusters in your single-cell dataset in one run. CASSIA processes multiple clusters in parallel, generating cell type predictions with detailed reasoning for each cluster.

Quick Start

CASSIA.runCASSIA_batch(
    marker = markers,
    output_name = "my_annotation",
    model = "anthropic/claude-sonnet-4.5",
    tissue = "brain",
    species = "human",
    provider = "openrouter"
)
Python

Input

Marker Data Formats

CASSIA accepts three input formats:

1. Seurat FindAllMarkers Output

Standard output from Seurat's FindAllMarkers function with differential expression statistics:

2. Scanpy rank_genes_groups Output (Recommended for Python)

Output from Scanpy's sc.tl.rank_genes_groups() function, typically exported using sc.get.rank_genes_groups_df():

3. Simplified Format

A two-column DataFrame with cluster ID and comma-separated marker genes:

Loading Marker Data

import CASSIA
import scanpy as sc
import pandas as pd

# Option 1: Load Seurat FindAllMarkers output (exported as CSV)
markers = pd.read_csv("seurat_markers.csv")

# Option 2: Use Scanpy rank_genes_groups output directly (Recommended for Python)
# (assuming you already have an AnnData object with rank_genes_groups computed)
markers = sc.get.rank_genes_groups_df(adata, group=None)  # Get all groups

# Option 3: Load your own simplified marker data
markers = pd.read_csv("path/to/your/markers.csv")

# Load example marker data for testing
markers = CASSIA.loadmarker(marker_type="unprocessed")

# Preview the data
print(markers.head())
Python

Parameters

Required

Optional

Parameter Details

Model Selection

• Default is anthropic/claude-sonnet-4.5 for best performance
• Use google/gemini-2.5-flash for faster, preliminary analysis
• For detailed model recommendations, see How to Select Models and Providers

Marker Gene Selection

• Default: top 50 genes per cluster
• ranking_method controls how marker genes are ranked and selected:
  • "avg_log2FC" (default): Rank by average log2 fold change
  • "p_val_adj": Rank by adjusted p-value
  • "pct_diff": Rank by difference in percentage expression
  • "Score": Rank by custom score
• Filtering criteria: adjusted p-value < 0.05, avg_log2FC > 0.25, min percentage > 0.1
• If fewer than 50 genes pass filters, all passing genes are used

Additional Context

• Use additional_info to provide experimental context
• Examples:
  • Treatment conditions: "Samples were antibody-treated"
  • Analysis focus: "Please carefully distinguish between cancer and non-cancer cells"
• Tip: Compare results with and without additional context

Output

Files Generated