Effective molecular representation learning is essential for molecular
property prediction. Contrastive learning, a prominent self-supervised approach
for molecular representation learning, relies on establishing positive and
negative pairs. However, this binary similarity categorization oversimplifies
the nature of complex molecular relationships and overlooks the degree of
relative similarities among molecules, posing challenges to the effectiveness
and generality of representation learning. In response to this challenge, we
propose the Graph Multi-Similarity Learning for Molecular Property Prediction
(GraphMSL) framework. GraphMSL incorporates a generalized multi-similarity
metric in a continuous scale, capturing self-similarity and relative
similarities. The unimodal multi-similarity metrics are derived from various
chemical modalities, and the fusion of these metrics into a multimodal form
significantly enhances the effectiveness of GraphMSL. In addition, the
flexibility of fusion function can reshape the focus of the model to convey
different chemical semantics. GraphMSL proves effective in drug discovery
evaluations through various downstream tasks and post-hoc analysis of learnt
representations. Its notable performance suggests significant potential for the
exploration of new drug candidates.

我们提出了图形多相似学习（GraphMSL）框架，它在连续尺度上捕捉了分子的自相似性和相对相似性，从而显著提高了分子性质预测中的表示学习的效果。此框架通过不同化学模态推导出单峰多相似性度量，融合这些度量到多模态形式中，灵活的融合函数可以重塑模型的重点以传达不同的化学语义，并通过各种下游任务和事后分析显示出了显著的性能，在药物发现评估中具有重要潜力。