Multi-objective molecular property prediction is a critical task in drug discovery and materials science, enabling the accelerated screening and optimization of novel molecular materials. However, existing methods face challenges due to the scarcity of labeled data and interference among tasks, resulting in insufficient learning of molecular 3D structural features and limited model generalization. This study proposes MolPropNet, a multi-objective molecular property prediction model based on a pre-trained framework. By incorporating dynamic noise injection, de-regularization, and group training strategies, MolPropNet significantly enhances its ability to model molecular structural features and improves prediction accuracy across multiple tasks. Specifically, dynamic noise injection adds Gaussian noise to the 3D coordinates to boost model robustness; de-regularization disables the Dropout mechanism to reduce variance and stabilize predictions; group training optimizes the model jointly based on the distribution characteristics of different molecular properties, effectively mitigating task interference. Additionally, a model ensemble strategy further improves overall performance. Experimental results on a second-generation OLED material molecular property dataset demonstrate that the pro- posed method achieves a weighted R2 (WR2) score of 0.7649, up from the baseline of 0.7208, with dynamic noise injection, de- regularization, and group training contributing 2.19%, 0.63%, and 1.59% performance gains, respectively. This study provides a feasible solution for efficiently screening and optimizing novel molecular materials, offering broad application potential.