TMolNet: A Task-Aware Multimodal Neural Network for Molecular Property Prediction

Abstract Molecular property prediction plays a vital role in drug discovery, materials science, and chemical biology. Although molecular data are intrinsically multi-modal—comprising 1D sequences or fingerprints, 2D topological graphs, and 3D geometric conformations—conventional approaches often rely on single-modal inputs, thereby failing to leverage cross-modal complementarities and limiting predictive accuracy. To overcome this limitation, we propose TMolNet, a task-aware deep learning framework for adaptive multi-modal fusion. The architecture integrates modality-specific feature extractors to learn distinct representations from 1D, 2D, and 3D inputs, reducing the bias caused by incomplete or under-represented modalities. A contrastive learning scheme aligns the representations across modalities within a shared latent space, enhancing semantic consistency. Furthermore, a novel task-aware gating module dynamically modulates the contribution of each modality based on both data characteristics and task requirements. To promote balanced modality usage during training, we introduce a modality entropy regularization loss, which encourages diversity and stability in learned representations. Extensive evaluations on multiple benchmark datasets demonstrate that TMolNet consistently outperforms existing state-of-the-art methods in terms of both predictive accuracy and generalization. These findings validate the effectiveness of our task-aware fusion strategy and establish a new direction for multi-modal molecular property prediction.