A Semi-supervised Molecular Learning Framework for Activity Cliff Estimation

TL;DR

Activity cliffs undermine the SSP principle in molecular property prediction, particularly for graph-based models. The paper introduces SemiMol, an instructor-guided semi-supervised framework that treats pseudo-labels with learned confidence via a separate instructor model and uses a self-adaptive curriculum to gradually incorporate harder unlabeled samples. Across 30 activity-cliff datasets in MoleculeACE, SemiMol consistently outperforms state-of-the-art pretrained GNNs, descriptor-based approaches, and other SSL baselines, demonstrating substantial improvements in regression tasks and strong results on a classification benchmark. The findings highlight the importance of backbone architecture and SSL design in leveraging unlabeled molecular data for robust activity-cliff estimation in low-data regimes.

Abstract

Machine learning (ML) enables accurate and fast molecular property predictions, which are of interest in drug discovery and material design. Their success is based on the principle of similarity at its heart, assuming that similar molecules exhibit close properties. However, activity cliffs challenge this principle, and their presence leads to a sharp decline in the performance of existing ML algorithms, particularly graph-based methods. To overcome this obstacle under a low-data scenario, we propose a novel semi-supervised learning (SSL) method dubbed SemiMol, which employs predictions on numerous unannotated data as pseudo-signals for subsequent training. Specifically, we introduce an additional instructor model to evaluate the accuracy and trustworthiness of proxy labels because existing pseudo-labeling approaches require probabilistic outputs to reveal the model's confidence and fail to be applied in regression tasks. Moreover, we design a self-adaptive curriculum learning algorithm to progressively move the target model toward hard samples at a controllable pace. Extensive experiments on 30 activity cliff datasets demonstrate that SemiMol significantly enhances graph-based ML architectures and outpasses state-of-the-art pretraining and SSL baselines.

Figures (4)

• Figure 1: For non-activity cliffs, GNNs successfully discover the motif of COC(=O)c1ccccc1 and group molecules with this pattern as local anesthetics. For activity cliffs, GNNs are misguided by the shared motif pattern of COCCNCc1ccccc1 and wrongly predict the inactivity of relevant molecules.
• Figure 2: Illustration of SemiMol. The target model first assigns predictions for unlabeled molecular data. Then the instructor model analyzes the confidence of those proxy labels. After that, a self-adaptive curriculum learning schema is adopted to move the target model towards hard samples progressively. By iteratively repeating these processes, the target model can extract task-specific information from vast unannotated molecules to the greatest extent.
• Figure 3: Comparison of different GNNs with and without pretraining on 30 activity cliff datasets. The light blue corresponds to the performance without pretraining, while the antique white represents the performance with pretraining.
• Figure 4: Performance of various ML models measured by the overall RMSE on 30 activity cliff datasets.