15 April 2026 in Engineering Applications of Artificial Intelligence
12 February 2026 in SSRN Electronic Journal
To achieve accurate medication recommendations, recent studies have focused on extracting appropriate molecular structure embeddings to capture pharmacological mechanisms. Building upon single 2D modality representations, existing works further incorporate 3D structural information and employ contrastive learning to derive modality-invariant embeddings. In this paper, we identify two key limitations of existing multimodal molecular representation methods. First, different modalities of the same molecule contain both shared features that convey common semantics and modality-specific features that provide complementary structural information. Undifferentiated alignment strategies lead to the loss of modality-specific information. Second, naive concatenation of embedding vectors may cause modality collapse, where the richer 3D modality dominates the fused representation and suppresses the contribution of the 2D modality. To overcome these issues, we propose SCAMed, a molecular representation learning framework designed to achieve precise alignment of shared information, effective extraction of structurally complementary modality-specific information, and balanced multimodal fusion. Specifically, during the representation extraction stage, we introduce an orthogonal decomposition module that separates the 2D and 3D molecular encodings into three disentangled components: 2D-specific, 3D-specific, and shared representations. In the fusion stage, we design a reconstruction-based learning strategy that enforces the fused representation to accurately reconstruct the original 2D and 3D embeddings, effectively mitigating modality collapse. Finally, the fused molecular representations are integrated with patient Electronic Health Record (EHR) data for medication prediction. Comprehensive experiments on the MIMIC-III and MIMIC-IV datasets demonstrate that our framework achieves substantial improvements over state-of-the-art baselines.
6 August 2026 in Information Fusion