Multimodal molecular representation and property prediction based on structured complementary feature fusion.

In this work, we propose Molrep-SCF, a multimodal self-supervised pretraining framework for molecular representation learning using SMILES sequences, 2D molecular graphs, and 3D geometric structures. The three modalities are encoded using a Transformer, a graph neural network, and an invariant 3D graph neural network, respectively. To enrich geometric supervision beyond a single conformation, Molrep-SCF incorporates three selected and complementary 3D conformational views. We further introduce a 2D-guided conformational consistency refinement mechanism that uses molecular topology to guide the fusion and refinement of their geometric representations. Instead of relying on a single static conformation, our method enforces structural consistency across conformers under the guidance of 2D topological priors. Furthermore, we design a structure-aware cross-modal learning strategy with atomic-level alignment via position-specific masking, enabling fine-grained correspondence across modalities. In addition, masked reconstruction across modalities enhances deep semantic coupling and facilitates information exchange between heterogeneous representations. Extensive experiments demonstrate that Molrep-SCF consistently outperforms existing multi-modal frameworks on molecular property prediction and other tasks, showing stronger generalization and representation quality.