Local structure-activity landscape roughness detects the activity-cliff failure mode that standard QSAR reliability methods miss

Technical Summary: Local Structure–Activity Landscape Roughness for QSAR Reliability

Problem Statement
Quantitative Structure–Activity Relationship (QSAR) models and deep learning approaches rely fundamentally on the molecular similarity principle: structurally similar molecules possess similar properties. However, "activity cliffs"—pairs of structurally similar molecules with large differences in potency—represent a principled exception where this assumption fails. Recent benchmarks (e.g., MoleculeACE) confirm that machine learning models perform significantly worse on activity cliffs than on other compounds.

The critical issue is that standard per-compound reliability estimators fail to detect these specific failure modes. Existing tools, including random-forest ensemble variance, distance-based applicability domain (AD) analyses, Sheridan's three-measure error model, and reliability-density-neighbourhood (RDN) methods, are all organized around the concept of proximity to the training data. Because activity cliffs occur in dense, well-sampled regions among close analogs, these proximity-based methods often report high confidence for cliff compounds, effectively missing the failure.

Methodology
The study benchmarks standard reliability estimators against a new metric: local structure–activity landscape roughness.

• Data: Analyses were conducted on the MoleculeACE benchmark (30 macromolecular targets, ~48.7K measurements) and external physicochemical benchmarks (ESOL solubility and lipophilicity).
• Models: Primary analysis used Random Forests (200 trees); Gradient Boosted Trees, Support Vector Regressors (RBF), and Graph Isomorphism Networks (GIN) were used for generality checks.
• Descriptors:
  • Standard Baselines: Ensemble variance, distance-based AD (1 - Tanimoto similarity to nearest neighbor), Sheridan's error model, and a regression adaptation of RDN.
  • Proposed Metric (Local Roughness): A family of functionals computed from a compound's nearest training neighbors ($N_k$) without using the query's own activity ($y_i$). Key descriptors include:
    • Neighborhood Activity Dispersion ($D$): The standard deviation of activities among neighbors.
    • Pairwise-SALI Density ($S_{mean}$): The mean of $|y_a - y_b| / \max(1-s_{ab}, \epsilon)$ over neighbor pairs.
    • Landscape Descriptors: Variants like Dirichlet energy and Lipschitz constants that use the query's own activity (used as theoretical bounds rather than deployable signals).
• Evaluation: Performance was measured by ROC-AUC for ranking high-error compounds and flagging labeled activity cliffs. Statistical significance was assessed via Spearman correlations, partial correlations (controlling for AD), and linear mixed-effects models.

Key Results

1. Detection of Activity Cliffs: Standard reliability methods perform at chance or worse when identifying activity cliffs. For example, the distance-based AD yielded an ROC-AUC of 0.35 (inverted), and ensemble variance was at chance (0.51). In contrast, local roughness achieved a median ROC-AUC of 0.69 for flagging cliffs.
2. Distinct Reliability Axis: Local roughness is statistically distinct from the applicability domain. Even after controlling for nearest-neighbor similarity and local density via partial correlation and mixed-effects modeling, roughness remained a significant predictor of error (e.g., neighborhood dispersion correlation $\rho \approx 0.23$ after control). This confirms roughness captures a failure mode (cliffs in dense regions) that proximity-based metrics miss.
3. Generalizability: The relationship between roughness and error held across different model classes (Random Forest, Gradient Boosting, SVR) and different molecular representations (ECFP4, ECFP6, physicochemical descriptors). The signal also reproduced on external physicochemical benchmarks (solubility and lipophilicity).
4. Operational Triage: While ensemble variance remains the strongest single flag for general high-error predictions (recovering 38% of top-quartile errors), it fails to identify cliffs. Roughness is the superior flag for cliffs, recovering 33% of labeled cliffs in the top 20% of flagged compounds, whereas variance performs at chance (1.0-fold enrichment).
5. Conformal Prediction Calibration: Conditioning conformal predictors on local roughness restores valid coverage on activity cliffs. Standard conditioning on variance or AD leaves cliff coverage below nominal (e.g., 86% vs. 90% target). A roughness-stratified Mondrian predictor maintained flat coverage across the roughness range and lifted cliff coverage to 89.5%.

Significance and Claims
The paper claims that local landscape roughness provides a complementary, per-compound reliability axis that is distinct from distance to the training data. Its primary contribution is not replacing uncertainty quantification but adding a specific sensitivity to the "cliff failure mode" that standard methods systematically miss.

By identifying regions where the structure–activity landscape is "rough" (sharply changing) based solely on training neighbors, the method allows for the construction of cliff-aware conformal predictors. This ensures that prediction intervals remain valid precisely where QSAR models are most likely to fail and most confidently wrong. The authors emphasize that this approach is practical, reproducible, and deployable without requiring the assay of the query compound, addressing a critical gap in the trustworthiness of computational drug discovery tools.