MotifGrIm: Motif-Based Multi-Granularity Graph-Image Pretraining for Molecular Representation Learning

ICLR 2026 Conference Submission16081 Authors

Keywords: Multi-Modal Contrastive Learning, Molecular Representation Learning, Graph Neural Network

Abstract: Molecular representation learning is widely considered as a crucial task in computer-aided molecular applications and design. Recently, many studies have explored pretraining models on unlabeled data to learn molecular structures and enhance the performance of downstream tasks. However, existing methods mainly focus on graph domains, with limited attention to other modals, such as the images. In addition, most existing methods focus on the atomic or molecular level, which leads to the neglect of high-order connection information or local structure information. In this work, we propose a motif-based multi-granularity graph-image pretraining framework, MotifGrIm, for molecular representation learning.In this framework, we incorporate motifs into the image domain for the first time,by generating distinct background features for different motifs in molecular im-ages, offering a novel approach to enhancing molecular representation. Through contrastive learning within and across modules, we effectively tackle two key challenges in molecular motif pretraining with graph neural networks: (1) the over-smoothing problem, which restricts GNNs to shallow layers and hinders global molecular information capture, and (2) the aggregation of motif nodes, which leads to the loss of connectivity information between motifs. Additionally, to more effectively capture information across different molecular granularities, we propose a multi-granularity prediction pretraining strategy to optimize the model. For downstream tasks, we use only the graph encoders for prediction, reducing both time and memory consumption. We evaluate MotifGrIm on molecular property prediction and long-range benchmarks. Across eight commonly used molecular property prediction datasets, MotifGrIm outperforms state-of-the-art models with an average ROC-AUC improvement of 1.16% and achieves the best results on five of them. On long-range datasets, MotifGrIm improves the performance by at least 14.8%.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16081