Explainable AI for Cancer Drug Response Prediction: Beyond Univariate Feature Attributions

Technical Summary: Explainable AI for Cancer Drug Response Prediction: Beyond Univariate Feature Attributions

Problem Statement
Predicting cancer drug response from transcriptomic profiles is a critical component of precision oncology. While machine learning models have achieved strong predictive performance in modeling high-dimensional, nonlinear dependencies between tumor molecular profiles and drug response, their scientific utility is often limited by a lack of interpretability. Current Explainable AI (XAI) approaches in this domain predominantly rely on univariate feature attribution methods, such as SHAP. These methods suffer from two primary limitations:

1. Robustness and Scalability: SHAP-based explanations exhibit limited robustness and struggle to scale effectively to high-dimensional transcriptomic data (thousands of genes).
2. Biological Oversimplification: Univariate methods assess genes independently, failing to capture the coordinated activity of multiple genes and pathways that drive cellular sensitivity and resistance. Consequently, higher-order biological mechanisms and gene-gene interactions remain hidden.

Methodology
The authors propose ILLUME+, a scalable post-hoc explainability framework designed to overcome the limitations of existing methods and move beyond single-gene assessments. The methodology is structured as a fully data-driven, end-to-end pipeline:

1. Data Preprocessing and Discretization: Continuous drug response data (e.g., $IC_{50}$) is discretized into three ordered classes (sensitive, intermediate, resistant) using tertile-based binning on training data to facilitate the extraction of interpretable decision rules.
2. Feature Selection: A fully data-driven approach using the Boruta algorithm identifies an informative subset of genes relevant to each drug without imposing arbitrary thresholds or user-defined biases.
3. Black-Box Prediction: A gradient-boosting ensemble of decision trees is trained on the selected features to model the nonlinear relationships between gene expression and drug sensitivity.
4. ILLUME+ Framework:
   • Scalable Architecture: ILLUME+ extends the ILLUME meta-explainer by addressing computational bottlenecks in high-dimensional settings. It employs a 1-layer low-rank hypernetwork to compress the weights of the mapping function $W(x)$, reducing the parameter count from $O(v \times m^2)$ to $O(v \times m \times r)$ (where $r \ll m$).
   • Efficient Regularization: It replaces the computationally expensive full Jacobian penalty with an efficient stochastic estimator based on Jacobian-vector products, enforcing local linearity without explicit Jacobian construction.
   • Explanation Generation: The framework learns a locally-linear latent representation of the black-box model. From this space, it decodes multiple explanatory objects:
     • Univariate Attributions: Gene importance scores derived from surrogate models (logistic regression or decision trees).
     • Decision Rules: Factual and counterfactual rules capturing joint expression patterns (e.g., "if Gene A > $c$ AND Gene B < $d$, then Sensitive").
     • Gene-Gene Interactions: Pairwise interactions are identified by analyzing the co-occurrence of genes within decision rules using the lift measure. These interactions are visualized as a weighted gene-gene interaction graph.

Key Contributions

• ILLUME+: A scalable extension of the ILLUME framework capable of extracting robust, structured explanations from high-dimensional transcriptomic data, addressing the scalability limitations of existing post-hoc XAI methods.
• End-to-End Pipeline: A novel pipeline that moves beyond univariate feature attributions to recover explanations capturing gene-gene mechanisms underlying drug sensitivity and resistance.
• Hypothesis Generation: A demonstration that the method can confirm known biological associations and reveal novel gene interactions to generate testable hypotheses without relying on predefined biological priors.

Results
The authors evaluated ILLUME+ against baselines (SHAP and LIME) using 20 drug datasets derived from the GDSC and CCLE resources.

• Univariate Attribution Performance:

  • Target Recovery: ILLUME+ more effectively ranks putative drug targets among the top genes compared to SHAP and LIME, as measured by NDCG@k.
  • Pathway Enrichment: ILLUME+ prioritizes Mechanism-of-Action (MoA) related genes more effectively than SHAP (statistically significant) and yields results comparable to LIME.
  • Robustness: ILLUME+ demonstrates significantly higher robustness (measured by cosine similarity of importance vectors in local neighborhoods) across all neighborhood sizes compared to both SHAP and LIME.

• Gene-Gene Interaction Analysis:

  • Pathway Overlap: Top-ranked gene pairs identified by ILLUME+ exhibit significantly higher pathway overlap (functional proximity) than bottom-ranked pairs. This separation is consistent across all drugs, whereas SHAP-based interaction rankings showed significance for only 40% of drugs.
  • Biological Coherence: In specific case studies (e.g., Venetoclax), the method identified distinct interaction patterns: sensitive cells showed interactions related to immune response and neutrophil chemotaxis, while resistant cells showed interactions related to angiogenesis and cell adhesion.
  • Network Centrality: Putative drug targets were found to occupy central positions in the interaction graphs (e.g., high strength for chromatin remodeling targets in sensitive lines; high betweenness for signaling targets like EGFR in sensitive lines), consistent with their biological roles.
  • Rule Extraction: The framework extracted concise, interpretable decision rules involving known prognostic genes and MoA-related pathways, as well as novel combinations suggesting new hypotheses (e.g., specific gene combinations for Erlotinib resistance).

Significance and Claims
The paper claims that ILLUME+ shifts the paradigm of drug response modeling from purely performance-driven exercises to hypothesis-generating tools for AI-assisted scientific discovery. By providing robust, scalable, and pairwise gene explanations, the framework allows for:

1. Faithful Auditing: A more accurate understanding of the decision logic learned by black-box models in high-dimensional settings.
2. Unbiased Discovery: The ability to uncover biologically meaningful structures and gene interactions without relying on predefined assumptions or biological priors, thereby avoiding the bias inherent in models that embed prior knowledge.
3. Actionable Insights: The generation of testable biological hypotheses regarding novel interaction-driven molecular signals in cancer biology.

The authors emphasize that while the extracted interactions are model-derived signals, they serve as a foundation for generating hypotheses that require independent experimental validation. The work demonstrates that principled explainability can effectively bridge the gap between predictive accuracy and biological insight in precision oncology.