Abstract: To address a critical limitation of conventional 3D Transformer-based molecular property prediction methods that neglect angular information, a 3D molecular property prediction method based on Angle-aware Transformer. First, the method systematically calculated three fundamental angular descriptors from atomic coordinates: bond angles, torsional angles, and rotational angles. These angular features were subsequently transformed into high-dimensional embeddings through spherical Bessel functions and spherical harmonics. Second, the derived angular information was processed through local and global graph convolutional layers to generate learnable angle-aware positional encodings, which were then integrated into the Transformer architecture as bias terms. Finally, a predictor comprising two SiLUactivated linear layers with an aggregation function was designed for property-specific prediction. Comprehensive evaluations on benchmark quantum chemistry and physicochemical datasets demonstrate that this method achieves significant accuracy improvements over baselines.