Latent Diffusion Transformer for 3D Molecular Property Prediction
Qing He, Gang Wang, Guoliang Hao, Minglong Lei · Procedia Computer Science · 2025
Predicting 3D molecular properties is highly significant in both drug discovery and materials science. However, existing methods are limited to providing point estimates and fail to comprehensively model the distribution of molecular properties. Furthermore, they exhibit limitations in capturing the complex non-linear relationships between 3D structural information and molecular properties. To address these challenges, this paper proposes a latent diffusion transformer-based method for 3D molecular property prediction. First, in the 3D Transformer encoder, we explicitly model molecular 3D structures using atomic coordinates and embed them into latent representations to fully encode spatial 3D molecular information. Second, in the latent space, we cluster molecular latent representations to generate pseudo-labels, which are employed as conditional guidance for the diffusion process. This enables the model to iteratively approximate the distributional relationship between molecular 3D structures and their target properties. Finally, a task-specific decoder maps the refined latent representations to the predicted distribution of molecular properties. Experimental results demonstrate that our method significantly outperforms existing benchmark models across multiple public 3D molecular property prediction datasets.