Improving hit discovery by integrating activity cliff sensitivity into active learning

Active learning has emerged as an effective strategy for accelerating molecular discovery under limited labeling budgets. However, existing methods primarily focus on global information, often overlooking activity cliff-sharp changes in bioactivity caused by small structural perturbations-leading to suboptimal sample selection. In this work, we propose a model-agnostic, activity cliff-aware active learning framework designed to improve hit discovery efficiency without imposing constraints on the underlying molecular representations. Our framework introduces an auxiliary activity cliff scoring module trained on pairwise molecular relationships to explicitly capture local structure-activity sensitivity. The outputs of this module are integrated into a cliff-aware acquisition function that prioritizes structurally informative molecules whose labels are expected to be most beneficial for model improvement. Notably, the proposed strategy is agnostic to backbone architectures and molecular feature, enabling seamless integration with a wide range of existing active learning pipelines.

We evaluate our approach on multiple benchmark datasets under a fixed labeling budget. Across all targets, the proposed method consistently identifies more active compounds than baseline acquisition strategies, demonstrating improved robustness in early-stage, data-scarce learning scenarios. Ablation studies further confirm the contribution of activity cliff awareness to the observed performance gains. Overall, our results underscore the importance of explicitly modeling activity cliffs within active learning frameworks and highlight the effectiveness of a model-agnostic design for accelerating hit discovery in data-limited drug discovery settings.

The source code is accessible online at https://github.com/wnsgk/AC-Active.