A unified multi-scale deep learning framework for molecular property prediction that bridges molecular structures and fingerprinting.

Accurate prediction of molecular properties such as bioactivity, solubility and toxicity is essential for computer-aided drug discovery, but current graph-based and fingerprint-based models often struggle to capture both local molecular structure and global molecular context. Here we show that a unified multi-scale graph-fingerprint network (UMSGFNet) that combines atom-level and substructure-level information with fingerprint-derived descriptors can improve predictive performance and robustness. The model uses a memory mechanism to represent long-range dependencies and a flexible nonlinear function to learn complex relationships between molecular structure and properties. By adaptively weighting graph-derived and fingerprint-derived features, UMSGFNet provides a single molecular representation that supports both classification of molecular activities and prediction of continuous properties. Across eight benchmark datasets, this framework achieves strong predictive performance and stable generalization compared with representative existing approaches, indicating that jointly modeling molecular structure and fingerprint information can support more reliable screening and design of candidate molecules.