Activity Cliffs and Machine Learning Generalization in QSAR: Why Similar Molecules Remain Difficult to Predict

Abstract

Machine learning has substantially expanded the predictive capability of quantitative structure-activity relationship (QSAR) modeling, yet reliable generalization remains difficult in chemically challenging regions. Activity cliffs-pairs of highly similar molecules that exhibit large differences in biological activity-represent a particularly stringent test because they violate the local smoothness assumed by many similarity-based and representation-learning approaches. This narrative review examines why activity cliffs remain difficult for contemporary QSAR models and how they expose limitations that may be concealed by conventional global performance metrics. We discuss the definition, detection, and chemical origins of activity cliffs; their relationship to discontinuities in structure-activity landscapes; and the influence of molecular representation, training-test partitioning, structural redundancy, and data leakage on apparent model performance. Evidence from fingerprint-based machine learning, deep neural networks, graph neural networks, and recent molecular foundation models indicates that increasing model complexity does not consistently resolve activity-cliff prediction. Models with strong overall QSAR performance may still exhibit substantially larger errors or poor sensitivity for cliff-forming compounds. Recent approaches that explicitly incorporate activity-cliff information through pairwise, contrastive, or activity-aware representation learning provide a promising alternative to conventional compound-level regression. We therefore propose an activity-cliff-aware QSAR perspective integrating rigorous bioactivity curation, activity-landscape characterization, leakage-resistant validation, global and cliff-specific evaluation, applicability-domain analysis, uncertainty quantification, and cliff-aware learning objectives. Activity cliffs should ultimately be regarded not merely as problematic outliers, but as information-rich stress tests of molecular machine-learning generalization and local structure-activity understanding.