Modeling interactions between drugs, microbes, and diseases is essential for advancing drug discovery and precision medicine. Although most existing computational approaches focus on pairwise association prediction, such as drug–microbe or microbe–disease associations, they often overlook the interdependencies among all three entities. In real-world biomedical systems, drug–microbiome interactions can modulate treatment efficacy, influence toxicity, and shape disease trajectories. Moving beyond binary relationships to explore triplet-level associations is essential for uncovering drugs’ mechanisms of action (MoAs). Recent advances in Graph Neural Networks (GNNs) have significantly improved the modeling of complex molecular structures, enabling more accurate property prediction. However, molecular and biomedical data extend beyond graph structures, encompassing diverse modalities such as molecular sequences, taxonomic hierarchies, and ontological descriptors— features that GNNs cannot fully capture. To address these limitations, we propose MCL-DMD (Multi-modal Contrastive Learning for Drug-Microbe-Disease Association Prediction), a novel framework that combines heterogeneous graph modeling with domain-specific biomedical knowledge. MCL-DMD employs HeTAN (Heterogeneous Triple Attention Network) to model the interconnectedness of all entities in a heterogeneous graph and augments it with a biomedical knowledge encoder that leverages SMILES representations, microbial taxonomies, and MeSH disease descriptors. Through multi-modal contrastive learning, MCL-DMD aligns cross-modal representations to improve semantic consistency and predictive robustness. Experimental results demonstrate that MCL-DMD significantly outperforms state-of-the-art baselines in both classification and ranking metrics, offering a powerful framework to uncover novel drug–microbe–disease associations.