Multi Omics Integration

Multi-Omics Integration — Reasoning Skill

Overview

Guide the agent through principled selection of multi-omics integration strategies, batch correction approaches, and enrichment methods. This skill encodes decision frameworks — not code — so the agent reasons correctly before generating any pipeline.

Usage

• Invoke when choosing between early, intermediate, or late integration strategies
• Use for batch correction decisions across omics layers or partial sample overlap
• Activate for MOFA+/iCluster/SNF method selection or multi-omic enrichment planning

---

Core Concepts

Response Format

• Lead with the direct recommendation or classification (≤3 sentences)
• Structure as: recommendation → justification (citing specific criteria/thresholds) → caveats
• Use tables for comparisons; bullet points for criteria lists
• Omit background the user already knows — they asked the question
• Target: 200-400 words unless the user requests exhaustive detail

The decision trees and parameter selection tables in this skill are for internal reasoning only. Apply them to select the correct integration strategy, but do not reproduce the full trees in your response. Never narrate the tree traversal or show deliberation between options. Present only the recommended approach with justification.

1. Integration Strategy Decision Tree

Follow this tree top-down. The first matching leaf is the recommended strategy.

START
 ├─ Sample size per omic layer?
 │   ├─ n ≥ 200 AND full overlap across layers
 │   │   └─ → EARLY integration (Section 2)
 │   ├─ 50 ≤ n 10x between layers with no overlap strategy

## When to Escalate to a Human Expert

- When integrated signatures will inform clinical trial stratification
- When batch effects cannot be distinguished from biological signal
- When results contradict established biology for the disease area

## 9. Anti-Patterns to Flag

1. **Wrong:** Concatenating omic layers without independent scaling
   **Right:** Scale each layer independently (e.g., z-score or variance-stabilize) before concatenation
   **Why:** The high-variance layer dominates all downstream analyses, masking signal from other layers

2. **Wrong:** Running ComBat without including biological covariates in the model
   **Right:** Always include biological variables of interest (disease status, sex) as covariates to protect
   **Why:** Without protected covariates, ComBat removes real biological signal along with batch effects

3. **Wrong:** Imputing more than 30% of missing values in an omic layer
   **Right:** Use late integration instead of imputation when missingness exceeds 30%
   **Why:** Heavy imputation introduces artificial structure that does not reflect biology and distorts factor analysis

4. **Wrong:** Running MOFA+ with fewer than 20 samples
   **Right:** Use late integration or simple correlation analysis for very small sample sizes
   **Why:** MOFA+ requires sufficient samples for stable factor estimation; with 5% variance in at least one view
   **Why:** Reporting only the top factor is cherry-picking; biologically important signals may appear in lower-ranked factors

6. **Wrong:** Not comparing multi-omic integration results to single-omic baselines
   **Right:** Always run single-omic analyses as baselines and demonstrate that integration adds value
   **Why:** Without a baseline comparison, you cannot assess whether integration actually improves over the best single-omic result

7. **Wrong:** Using gene symbols directly across species without ortholog mapping
   **Right:** Use proper ortholog databases (HGNC Comparison of Orthology Predictions, Ensembl Compara) for cross-species mapping
   **Why:** Ortholog mapping is non-trivial — many genes have no 1:1 ortholog, and symbol overlap does not guarantee functional equivalence

---

## 10. Glossary

| Term | Definition |
|---|---|
| View | A single omic layer (e.g., transcriptomics, proteomics) in MOFA+ terminology |
| Factor | A latent variable capturing shared or omic-specific variation |
| Block | A feature matrix from one omic layer in mixOmics/DIABLO terminology |
| Kernel | A similarity matrix computed from one omic layer |
| RAF | Risk Adjustment Factor (unrelated — see risk-adjustment skill) |
| ELBO | Evidence Lower Bound; MOFA+ convergence metric |
| BIC | Bayesian Information Criterion; used for selecting k in iCluster |

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [awslabs](https://github.com/awslabs)
- **Source:** [awslabs/hcls-agent-skills](https://github.com/awslabs/hcls-agent-skills)
- **License:** MIT-0

Install and usage instructions live in the source repository linked above.