tools-only Model Interpretation Usage Guide
Overview
Explain machine learning predictions on omics data using SHAP values for global and local feature attribution, and LIME for model-agnostic explanations.
Prerequisites
pip install shap lime scikit-learn xgboost matplotlib pandas
Quick Start
Tell your AI agent what you want to do:
- "Explain my classifier predictions with SHAP"
- "Which genes are most important for my model's predictions?"
- "Show me a SHAP summary plot for my classifier"
- "Use LIME to explain individual predictions"
Example Prompts
Global Feature Importance
"I trained a random forest on gene expression data. Use SHAP to show me which genes are most important across all predictions."
"Create a SHAP beeswarm plot for my XGBoost classifier. Show the top 20 features."
Individual Predictions
"Use SHAP to explain why my classifier predicted 'disease' for sample X. Show a waterfall plot."
"Generate LIME explanations for the 5 samples with the highest prediction confidence."
Feature Ranking
"Extract the mean absolute SHAP values for all features and save to a CSV file."
Interactions
"Create a SHAP dependence plot for BRCA1 to see how its expression level affects the prediction."
What the Agent Will Do
- Load trained model and test data
- Create appropriate SHAP explainer (TreeExplainer for tree models)
- Calculate SHAP values for test samples
- Generate summary visualizations (beeswarm, bar)
- Save top features to file
Tips
- Use
explainer(X)not.shap_values(X)for SHAP v0.47+ - TreeExplainer is fastest for tree-based models (RF, XGBoost, LightGBM)
- LIME is model-agnostic but slower; use for non-tree models
- Beeswarm plots show both importance and direction of effect
- For multi-class, SHAP values have an extra dimension for classes
- SHAP values sum to the difference between prediction and expected value