Accurately predicting molecular properties is essential for drug discovery and material design, yet it remains challenging due to the complex relationships between molecular structures and their properties. Existing methods typically rely on either SMILES strings or molecular graphs, but neither representation alone can fully capture the diverse chemical and structural information required for robust predictions. To address this limitation, we propose SMGA (SMILES and Graph-Aligned contrastive learning), a novel framework that combines the strengths of both SMILES strings and molecular graphs. SMGA utilizes a Transformer encoder to extract 1D features from SMILES sequences and a GNN encoder to capture 2D features from molecular graphs. The representations are aligned using a Latent Bootstrap Cross Consistency (LBCC) loss, which ensures cross-modal consistency without the need for data augmentation or motif extraction. Unlike many pre-training-based approaches, SMGA achieves competitive performance without requiring a pre-training phase, offering both efficiency and scalability. Extensive experiments on multiple datasets from MolecularNet demonstrate the effectiveness of SMGA across classification and regression tasks, highlighting its potential for real-world molecular property prediction.