Model Interpretation for Omics Classifiers
SHAP TreeExplainer (v0.47+ API)
import shap
from sklearn.ensemble import RandomForestClassifier
model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)
explainer = shap.TreeExplainer(model)
# CORRECT (v0.47+): Call explainer directly, NOT .shap_values()
shap_values = explainer(X_test)
# shap_values is an Explanation object
# .values has shape (n_samples, n_features) for binary
# .base_values has expected value
print(f'SHAP values shape: {shap_values.values.shape}')
Summary Plot (Global Feature Importance)
import shap
import matplotlib.pyplot as plt
# Beeswarm plot: shows impact direction and magnitude
shap.plots.beeswarm(shap_values, max_display=20, show=False)
plt.tight_layout()
plt.savefig('shap_summary.png', dpi=150, bbox_inches='tight')
plt.close()
# Bar plot: mean absolute SHAP values
shap.plots.bar(shap_values, max_display=20, show=False)
plt.savefig('shap_bar.png', dpi=150, bbox_inches='tight')
Force Plot (Individual Prediction)
# Explain single prediction
sample_idx = 0
shap.plots.force(shap_values[sample_idx], matplotlib=True, show=False)
plt.savefig('shap_force_single.png', dpi=150, bbox_inches='tight')
# Waterfall plot (cleaner alternative)
shap.plots.waterfall(shap_values[sample_idx], max_display=15, show=False)
plt.savefig('shap_waterfall.png', dpi=150, bbox_inches='tight')
SHAP for XGBoost
from xgboost import XGBClassifier
import shap
xgb = XGBClassifier(n_estimators=100, random_state=42, eval_metric='logloss')
xgb.fit(X_train, y_train)
explainer = shap.TreeExplainer(xgb)
shap_values = explainer(X_test)
# For XGBoost, shap_values contains log-odds contributions
shap.plots.beeswarm(shap_values, max_display=20)
LIME (Local Interpretable Model-agnostic Explanations)
from lime.lime_tabular import LimeTabularExplainer
import numpy as np
explainer = LimeTabularExplainer(
X_train.values,
feature_names=X_train.columns.tolist(),
class_names=['control', 'disease'],
mode='classification'
)
# Explain single instance
sample_idx = 0
exp = explainer.explain_instance(
X_test.iloc[sample_idx].values,
model.predict_proba,
num_features=20
)
exp.save_to_file('lime_explanation.html')
# Or get as list: exp.as_list()
Extract Top Features from SHAP
import pandas as pd
import numpy as np
# Mean absolute SHAP value per feature
mean_shap = np.abs(shap_values.values).mean(axis=0)
feature_importance = pd.DataFrame({
'feature': X_test.columns,
'mean_shap': mean_shap
}).sort_values('mean_shap', ascending=False)
top_features = feature_importance.head(20)
top_features.to_csv('shap_top_features.csv', index=False)
Dependence Plot (Feature Interactions)
# Shows how SHAP value varies with feature value
# Automatically colors by interacting feature
shap.plots.scatter(shap_values[:, 'GENE1'], color=shap_values, show=False)
plt.savefig('shap_dependence.png', dpi=150, bbox_inches='tight')
Multi-class SHAP
explainer = shap.TreeExplainer(model)
shap_values = explainer(X_test)
# For multi-class, shap_values.values has shape (n_samples, n_features, n_classes)
# Access class-specific values:
class_idx = 1
shap.plots.beeswarm(shap_values[:, :, class_idx], max_display=20)
Related Skills
- machine-learning/omics-classifiers - Train models to interpret
- machine-learning/biomarker-discovery - Compare with selection-based importance
- data-visualization/heatmaps-clustering - Visualize SHAP values as heatmap
FridrichMethod