| name | batch_integration_advisor |
| description | Recommend a batch-integration strategy and explain the tradeoffs for preserving biology versus removing technical effects. |
| category | bio/single_cell_rna |
| version | 1 |
| requires_tools | ["search_knowledge_base","python_repl"] |
| requires_network | false |
| user_invocable | true |
| tags | ["batch","integration","harmony","scanorama","seurat"] |
| aliases | ["integration_strategy_helper"] |
| species | any |
| modality | single_cell_rna |
| stage | preprocess |
| stability | stable |
| safety_level | low |
Batch Integration Advisor
Purpose
Help the user choose an integration strategy for single-cell datasets while preserving the biology they care about.
When to use
Use this skill when the user asks how to integrate batches, donors, or experiments, or whether they should integrate at all.
Required inputs
- datasets or batches: a short description
- main goal: clustering, annotation transfer, visualization, or DE
- known biology to preserve: condition, treatment, donor, timepoint, lineage, or unknown
Steps
- Search
knowledge/batch-integration-playbook.md using search_knowledge_base.
- Identify the likely technical sources of variation versus the biological signal the user wants to keep.
- Recommend a strategy such as Harmony, Scanorama, Seurat integration, or no integration.
- Use
python_repl if it helps to structure options or decision criteria.
- Include diagnostics the user should check after integration.
Output format
- Recommended strategy
- Why it fits the stated goal
- What could go wrong
- Suggested diagnostics
Failure modes
- Too little information: ask what the user wants to preserve.
- DE-focused question: warn that integrated embeddings are not always the right input for DE.
- Strongly non-overlapping cell states: warn that integration may over-correct.
Examples
- "Should I use Harmony or Seurat integration for three donor batches?"
- "I want to integrate two conditions but preserve treatment response. What should I do?"
