Separating drug main effects from molecular personalization in cancer drug-sensitivity prediction: a harmonized GDSC∪CTRP and CCLE benchmark with permutation-controlled biomarker discovery

Technical Summary: Separating Drug Main Effects from Molecular Personalization in Cancer Drug-Sensitivity Prediction

Problem Statement
Large-scale pharmacogenomic screens (GDSC and CTRP) paired with multi-omic data (CCLE/DepMap) are the standard substrate for predicting tumor drug response from molecular profiles. However, a persistent gap exists between the aspiration of "molecular personalization" (ranking drugs based on a specific tumor's unique profile) and the reality of data structure. The paper identifies two primary confounds:

1. Drug Main Effects: Many compounds are broadly potent across most cell lines, and different tissues have inherently different baseline sensitivities. Models often achieve high correlation by simply learning these general trends rather than true personalization.
2. Data Inconsistency: Cross-study concordance between GDSC and CTRP is only moderate, limiting the reproducibility of biomarkers learned from one screen when tested on another.

The central question is how much of the apparent predictive skill in existing models reflects genuine molecular personalization versus simpler data structures (drug potency and tissue-of-origin effects), and how to distinguish true biomarkers from artifacts of over-optimistic evaluation.

Methodology
The study constructs a harmonized resource and a rigorous evaluation framework to disentangle these signals:

• Data Harmonization: The author integrates GDSC2 and CTRP v2.2 drug-response data with CCLE multi-omic profiles (RNA-seq, mutations, methylation, etc.).
  • Identity Resolution: Cell lines are mapped to DepMap identifiers. Crucially, drugs are resolved by chemical structure (InChIKey connectivity block) rather than name, preventing the pooling of distinct isomers and increasing the number of confidently shared compounds from 71 to 86.
  • Metric Standardization: A cross-cohort comparable sensitivity metric (sens_z) is created by standardizing raw AUC or IC50 values within each (cohort, drug) group, removing drug-specific potency offsets and cohort-specific scale differences.
• Evaluation Discipline:
  • Grouped Leave-Cell-Line-Out Nested Cross-Validation: Models are evaluated such that no cell line appears in both training and test sets, preventing leakage of general chemosensitivity levels.
  • Baseline Anchoring: All models are scored against a drug-mean baseline (predicting the average sensitivity of a drug within a tissue, ignoring molecular features). If a molecular model cannot beat this baseline, it is re-learning drug main effects, not personalizing.
  • Decomposition: The prediction problem is split into:
    1. Broad-spectrum component: Predicting which drugs are generically potent in a tissue (drug-mean baseline).
    2. Mechanism-specific component: Predicting the residual sensitivity after removing drug means (and, in pan-cancer settings, tissue-of-origin means).
• Significance Control:
  • Two-Stage Protocol: Instead of relying on cross-validation confidence intervals (which measure split-stability, not sampling noise), the study employs a permutation null (shuffling sensitivity labels within the panel) followed by independent corroboration (checking against established pharmacology and cross-tissue recurrence).
  • Per-Drug Signatures: For each of the 7,836 eligible (disease, drug) panels, a supervised gene signature is fitted and tested against the permutation null.
• Tools: The analysis utilizes Elastic Net regression, ssGSEA for pathway activity, and a semantic matching engine (Semblance) for transfer-robust recommendation.

Key Results

1. Broad-Spectrum Signal is Tissue-by-Drug Structure:

   • Within individual tissues, a drug-mean model (ignoring molecular features) is strongly predictive (Spearman $\rho$ up to 0.54 in hematologic lineages).
   • When tissues are pooled, this signal vanishes (pooled $\rho \approx 0$), confirming that the "predictive power" is simply the known fact that certain tissues are generally sensitive to certain drugs, not individual tumor personalization.

2. Molecular Personalization is Minimal and Mostly Lineage-Dependent:

   • Per-Disease: In individual tissue panels (20–70 cell lines), molecular features (expression, multi-omics, pathways) add virtually no predictive power beyond the drug-mean baseline. The mechanism-specific arm clusters at zero.
   • Pan-Cancer: Pooling tissues increases the effective sample size, revealing a small non-zero correlation ($\rho \approx 0.10$). However, when the target is residualized to remove tissue-of-origin structure, this signal drops to $\rho \approx 0.05$.
   • Lineage Dominance: A "no-feature" tissue baseline (predicting based solely on the in-fold disease-drug mean) outperforms the expression-based model ($\rho = 0.214$ vs. $0.087$). This indicates that lineage identity alone predicts residual sensitivity better than 50 expression principal components. The remaining signal ($\rho \approx 0.05$) is permutation-validated but likely represents unresolved within-tissue line properties (e.g., proliferation rate) rather than established molecular personalization.

3. Per-Drug Discovery Yields a Concentrated Minority of Validated Biomarkers:

   • Applying the two-stage significance protocol (permutation null + corroboration) across 7,836 panels reduces the set of "significant" findings from 3,284 (standard CI) to 971 panels (12%).
   • Validation: The retained set is enriched 4.3-fold for established drug-target pharmacology (e.g., BRAF inhibitors in melanoma/colorectal, MEK inhibitors, BCL2 inhibitors).
   • Novelty: Beyond canonical targets, the study identifies recurrent, permutation-validated vulnerabilities in non-canonical agents. Notably, three independent GPX4 inhibitors (RSL3, ML162, ML210) all show strong, validated signatures in melanoma, recovering the known ferroptosis vulnerability of dedifferentiated melanoma.
   • Recurrence Threshold: Recurrence of a drug's signature across tissues is only informative above a threshold of three tissues; two-tissue recurrence is fully explained by chance given the number of tissues tested.

4. Pathway Re-representation:

   • Converting expression to Reactome pathway activity did not unlock additional mechanism signal; the results remained consistent with the gene-level analysis.

Significance and Claims
The paper claims its primary contribution is an honest map of where predictive signal genuinely lives in pharmacogenomic data, rather than a new state-of-the-art predictor.

• Reframing Personalization: The study argues that "personalization" in current datasets is not a distributed property across all drugs and tissues but is concentrated in a well-powered minority of drug-tissue pairs. Most apparent signal is actually drug main effects or tissue-of-origin structure.
• Evaluation Discipline: It establishes that conventional cross-validation intervals are insufficient for discovery in this domain. A permutation-controlled null is necessary to filter out artifacts, reducing the false discovery rate significantly.
• Actionable Output: The work provides a tiered recommendation framework:
  1. High-Confidence Tier: Broad-spectrum rankings based on tissue-specific mean sensitivity (drug-mean baseline).
  2. Exploratory Tier: Mechanism-specific recommendations restricted to the ~12% of drug-tissue pairs with permutation-validated signatures.
• Resource Release: The study releases a harmonized GDSC$\cup$CTRP–CCLE resource, a distributed analysis pipeline, and an open web application (Semblance 2.0) that operationalizes these confidence tiers for cell-line lookup and semantic signature matching.

The author concludes that while a small, permutation-validated personalization signal exists, it is fragile and dwarfed by tissue structure. Future progress requires focusing on per-drug, within-tissue discovery with rigorous null control, rather than seeking a universal pan-cancer predictor.