Claw4Science

Explainable AI for scRNA-seq Analysis

Mar 26, 2026

Key Takeaways

  • Explainable AI for scRNA-seq means making every analysis step transparent: why cells were clustered a certain way, why a gene was called differentially expressed, why a trajectory was inferred
  • AI agents (MedgeClaw, OmicsClaw, BioClaw) can automate scRNA-seq pipelines while providing step-by-step reasoning
  • K-Dense scRNA-seq skills encode best practices: normalization method selection, clustering parameter justification, marker gene validation
  • Key explainability techniques: SHAP values for feature importance, attention visualization for cell embeddings, provenance tracking for pipeline decisions
  • The gap: most tools automate execution but don't explain decisions — explainability is the next frontier

Why Explainability Matters in scRNA-seq

Single-cell RNA sequencing generates complex, high-dimensional data. Standard analysis involves quality control, normalization, dimensionality reduction, clustering, and differential expression — dozens of decisions that affect final results.

The problem: Most AI agents automate these steps but don't explain why they chose specific parameters. A reviewer asking "why did you use Leiden clustering with resolution 0.8 instead of 1.2?" gets no answer.

Core attributes:

  • Data type: Single-cell RNA sequencing (10x Genomics, Smart-seq2, etc.)
  • Analysis steps: QC → normalization → HVG selection → PCA → clustering → DE → trajectory
  • Explainability gap: Automation without justification

Tools for Explainable scRNA-seq Analysis

ToolExplainability LevelHow It Explains
MedgeClawHighK-Dense skills encode parameter justification; research dashboard shows reasoning
OmicsClawMediumMulti-omics pipeline with step logging; less parameter explanation
BioClawLowExecutes analysis but minimal explanation of decisions
Scanpy (manual)NoneUser must justify all decisions manually

K-Dense scRNA-seq Skills: Built-in Explainability

The K-Dense scientific skills library includes specialized scRNA-seq skills that encode expert reasoning:

Normalization skill:

  • Evaluates library size distribution before choosing method
  • Explains why SCTransform vs log-normalization was selected
  • Documents batch effect assessment and correction decision

Clustering skill:

  • Tests multiple resolutions and reports silhouette scores
  • Justifies final resolution choice with biological validation (marker genes)
  • Flags potential over-clustering or under-clustering

Differential expression skill:

  • Selects statistical test based on data distribution (Wilcoxon vs t-test vs MAST)
  • Reports effect sizes alongside p-values
  • Validates results against known marker databases

Explainability Techniques

TechniqueApplicationTool Support
SHAP valuesFeature importance for cell type classificationMedgeClaw, custom skills
Attention visualizationCell embedding interpretation in transformer modelsResearch tools
Provenance trackingRecord every pipeline decision with justificationMedgeClaw dashboard
Parameter sensitivityShow how results change with different parametersK-Dense skills
Marker gene validationCross-reference clusters with known cell type markersAll scRNA-seq tools

Practical Workflow

A fully explainable scRNA-seq analysis with AI agents:

1. Data loading → Agent documents sample metadata, sequencing platform
2. QC → Agent explains filtering thresholds (why >200 genes, <5% mito)
3. Normalization → Agent tests methods, justifies choice with data statistics
4. HVG selection → Agent reports number selected, biological rationale
5. Dimensionality reduction → Agent explains PCA components retained
6. Clustering → Agent tests resolutions, validates with marker genes
7. DE analysis → Agent selects test, reports effect sizes, validates markers
8. Trajectory → Agent justifies root cell selection, documents branch points

Each step generates a reasoning trace that can be included in a methods section.

FAQ

Q1: Which tool provides the best explainability for scRNA-seq?

MedgeClaw with K-Dense skills currently offers the most built-in explainability — parameter justification, research dashboard, and provenance tracking. OmicsClaw provides good pipeline logging but less parameter explanation.

Q2: Can I add explainability to existing Scanpy workflows?

Yes. Install K-Dense scRNA-seq skills and use them with Claude Code or any compatible agent. The skills wrap Scanpy functions with decision-justification logic.

Q3: Is explainability required for publication?

Increasingly, yes. Reviewers expect justification of analysis parameters. Explainable AI workflows generate methods sections that answer "why" questions automatically.

Q4: What about explainability for deep learning cell type classifiers?

SHAP values and attention visualization can explain model predictions. However, most scRNA-seq AI agents use traditional analysis pipelines (Scanpy/Seurat) rather than deep learning, where parameter justification matters more than model interpretation.

Summary

Explainable AI for scRNA-seq is the next frontier beyond automation. While tools like MedgeClaw and OmicsClaw can automate the full analysis pipeline, the real value is in explaining why each decision was made — normalization method, clustering resolution, statistical test. K-Dense scRNA-seq skills lead in this area by encoding expert reasoning directly into the analysis workflow.

Explainable AI for scRNA-seq Analysis | Blog