SHAP Values

The sum of all feature attributions equals the difference between the model's output for a specific instance and the average model output. This guarantees that the explanation is faithful to the original model prediction.

• For a drug-disease pair scored at 0.85 with a baseline of 0.5, the SHAP values for chemical descriptors, target proximity, and gene expression features will sum to exactly 0.35.
• This property ensures no attribution is lost or fabricated during the explanation process.

A feature that is already missing or unobserved in the input is assigned an attribution of zero. This prevents the explanation from hallucinating importance for data that was never provided to the model.

• If a drug's protein binding data is unavailable, its SHAP value is exactly 0, ensuring the explanation reflects only the information the model actually used.
• Critical for pharmacovigilance applications where incomplete patient records are common.

If a model changes so that a feature's marginal contribution increases or stays the same regardless of other features, the SHAP value for that feature will never decrease. This ensures explanations are logically coherent across model iterations.

• When retraining a polypharmacology model and a specific molecular fingerprint becomes more predictive of binding, its SHAP value will reflect this increased reliance.
• Prevents counterintuitive scenarios where a more important feature receives a lower attribution score.

SHAP values express the prediction as a linear sum of feature contributions plus a constant baseline. This additive structure makes explanations intuitive and directly comparable across different drug-disease pairs.

• A prediction = baseline + SHAP(chemical_similarity) + SHAP(target_proximity) + SHAP(gene_expression_reversal)
• Enables rapid identification of which biological mechanism—structural similarity, target binding, or transcriptomic reversal—is driving a specific repurposing hypothesis.

Aggregating absolute SHAP values across all predictions produces a global feature importance ranking, revealing which molecular descriptors or biological features the model relies on most heavily across the entire drug repurposing screen.

• A mean SHAP plot can show that gene expression signature reversal is the dominant factor in the model's decisions, followed by chemical similarity.
• This global view helps validate whether the model aligns with domain knowledge or has learned spurious correlations from the training data.

A precursor to SHAP that generates local explanations by perturbing inputs and fitting a simple, interpretable surrogate model around the prediction of interest. While model-agnostic, LIME lacks the game-theoretic guarantees of SHAP.

• Uses random perturbation sampling
• Explanations are local and linear approximations
• Faster to compute but less theoretically rigorous
• Often used as a lightweight alternative for quick sanity checks

A gradient-based attribution method designed specifically for deep neural networks. It computes feature importance by accumulating gradients along a straight-line path from a baseline input to the actual input.

• Satisfies the completeness axiom (attributions sum to prediction difference)
• Requires a meaningful baseline (e.g., zero embedding)
• Particularly effective for graph neural networks operating on molecular structures
• Common in toxicity prediction models where input gradients map to atom-level contributions

A critical validation error where information from the test set contaminates training, producing over-optimistic SHAP values. In drug repurposing, this often occurs when structurally similar compounds appear in both splits.

• Scaffold split leakage: Same molecular core in train and test
• Temporal leakage: Using future clinical data to explain past predictions
• Target leakage: Including features derived from the outcome variable
• Always validate SHAP explanations on a truly held-out set with no structural overlap

The process of corroborating SHAP-identified important features against known biochemical pathways and binding assays. A high SHAP value for a molecular substructure should correspond to a plausible interaction with the predicted target.

• Cross-reference with DrugBank target annotations
• Validate against crystallography data where available
• High SHAP importance without mechanistic plausibility may indicate spurious correlations
• Essential for regulatory acceptance of AI-driven repurposing hypotheses

SHAP values measure associational importance, not causal effects. A feature with high SHAP importance may be correlated with the outcome through a confounder rather than a direct mechanism.

• Use Mendelian Randomization to validate causal drug-target-disease relationships
• Instrumental variable analysis can disentangle causation from correlation
• SHAP is a hypothesis generator, not a causal proof
• Combine with do-calculus frameworks for rigorous causal claims in repurposing

A complementary approach where SHAP explanations on gene expression models identify which specific genes drive a drug-disease reversal prediction. This connects feature attribution directly to biological mechanism.

• Uses Connectivity Map (CMap) reference data
• SHAP values highlight the most influential differentially expressed genes
• Enables validation against pathway enrichment analysis
• Bridges the gap between model explanation and wet-lab validation