Prompting Meta-Learned Hierarchical Graph Network for Molecular Property Prediction

In the realm of molecular research, the observation of molecular properties often comes with a limited number of molecular samples. Consequently, researchers have framed the task of molecular property prediction as a few-shot problem. Existing methods for few-shot molecular learning treat molecules as graphs where atoms serve as nodes and bonds serve as edges. However, these approaches commonly overlook the inherent significance of functional groups, also known as motifs. Moreover, it is crucial to recognize that diverse molecular properties may exhibit intricate associations with different motifs. When predicting properties in new tasks, it becomes imperative to adaptively focus on the relevant motifs. To tackle this limitation, we investigate prompting a meta-learned hierarchical graph neural network (HGNN). The HGNN extracts motifs from the original molecular graph and obtains two distinct views of the molecule: the lower atom-level and the higher motif-level. Then it employs a cascading message passing paradigm to propagate information from atoms to motifs within the molecule. To enable the HGNN to adapt quickly to new tasks with rarely supported samples, we propose to learn a novel task prompt in meta-learning. The task prompt fills the gaps between tasks by discerning and assigning higher importance to the motifs that are specifically relevant to each task. Specifically, the task prompt is learned from a comprehensive task context, which incorporates two types of information: textual semantic features and class prototypes. Extensive experimental results on four benchmarks show the superior performance of our method. Our code is available at https://github.com/HICAI-ZJU/PH-Mol.