An Overview of Pre-Trained Graph Models for Molecular Structure Analysis

Molecules are naturally represented as graphs, which benefit graph models in understanding molecular structures to capture complex relationships and interactions between atoms. However, the limitations of labeled molecules can significantly impact the graph model's ability to generalize different molecular properties and structures, leading to poor performance. Lately, pretraining Graph Neural Networks (GNNs) have emerged as a powerful tool for learning molecular representations, addressing the challenge due to the lack of labeled molec-ular data. This study introduces and examines the pretraining GNNs strategies in learning molecular structure in computational chemistry and biology. We first categorize recent pretraining existing studies into three main groups: node-level strategies, contrastive learning, and graph-level strategies. Furthermore, we also highlight recent studies that significantly enhance model interpretability and prediction accuracy. This comprehensive perspective on the current state and future prospects of GNNs underscores their pivotal role in accelerating drug discovery and material innovation through advanced AI-driven analysis.