Predicting molecular properties is a cornerstone of AI-driven drug discovery, yet capturing the hierarchical complexity of chemical structures remains a significant challenge. This paper proposes a novel Hierarchical Graph Neural Network (HiGNN) framework that integrates four distinct graph representations: Atom, Bond, Fragment, and Fragment-Connection graphs. Our architecture employs a multi-scale feature extraction mechanism combined with a Probabilistic Latent Encoder inspired by Variational Autoencoders (VAEs) to model structural uncertainty. The model was pre-trained on a large-scale dataset of 28,995 molecules for molecular energy prediction, achieving a high degree of structural encoding with an \(R^2\) score of 0.9144 and an MAE of 15.66 on the test set. We evaluated the learned transferable embeddings across six MoleculeNet benchmarks. Experimental results demonstrate that the framework achieves an accuracy of 92.85% on the ClinTox dataset and a ROC-AUC of 0.6065 on the BACE benchmark. While performance on complex regression tasks like LIPO suggests further needs for data scaling, our findings highlight the potential of hierarchical representations in capturing essential chemical semantics for robust molecular property inference.