<p>Molecular property prediction is pivotal for drug discovery, offering significant potential to accelerate development and reduce costs. With the rapid development of artificial intelligence, molecular representation methods have become increasingly diversified. However, existing methods still have obvious deficiencies in the comprehensiveness of molecular representation and the effectiveness of feature fusion: single representation methods often can only capture part of a molecule’s features, while multi-representation methods focus on limited combinations and use simple fusion strategies. To address these issues, we propose Mol-SGGI, a comprehensive multi-representation learning framework that integrates four molecular representations: sequences, 2D graph structures, 3D geometric structures, and images. For each representation, we design specialized modules for extracting features and introduce appropriate attention mechanisms in each module to effectively capture the structural and chemical information of the molecule. Additionally, we propose an attention-guided adaptive weighted fusion module, which achieves multimodal feature alignment through contrastive learning and dynamically adjusts fusion weights. Experimental results on eight molecular property prediction tasks show that our model significantly outperforms the majority of existing methods.</p>