Drug-Target Affinity (DTA) prediction is a critical step in drug discovery, significantly accelerating the identification of potential therapeutic candidates. Existing methods often struggle with effectively integrating diverse biological information or suffer from limitations in feature representation. In this study, we propose a deep learning framework, MMSA-DTA (Multi-Modal Synergistic Attention for Drug-Target Affinity Prediction), designed to predict DTA by synergistically leveraging multi-modal features derived from both biomolecular sequences and interaction graph structures. Our model initiates sequence feature extraction by using pre-trained models (ESM-2 and Chemformer). These embeddings are further processed through three-layer Convolutional Neural Networks (CNNs) with residual connections to refine local and global patterns. Concurrently, we construct a drug-target interaction graph where nodes are characterized by initial physicochemical property-derived features and type embeddings, and edges are weighted by normalized binding affinities. A multi-layer Edge-aware Graph Attention Network (EGAT) is employed to learn intricate relational features from this graph. To create a comprehensive embedding for DTA prediction, the extracted sequence- and graph-modal representations undergo a hierarchical fusion process. This involves intra-modal fusion for drugs and targets, followed by inter-modal fusion of the combined representations, all implemented via a Highway network with 1D convolutional operations. Finally, MMSA-DTA uses a contrastive learning approach to refine the embedding space, enforcing similarity between representations of biochemically similar drug-drug and target-target pairs, and dissimilarity for disparate pairs. Extensive evaluations on benchmark datasets, Davis and KIBA, demonstrate that our proposed model significantly outperforms baseline methods across multiple evaluation metrics, including MSE, CI, and \(\hbox {R}^2\) .