Drug-induced liver injury (DILI) continues to pose a major challenge in pharmaceutical research, leading to frequent clinical trial failures and drug withdrawals from the market. Traditional computational models often emphasize molecular fingerprints to describe structural information, yet they overlook the sequential and contextual nuances encoded within SMILES representations. To address this limitation, we present a deep learning-based multimodal architecture that jointly leverages structural and sequential molecular representations for improved DILI prediction. Our framework employs multiple fingerprint descriptors—MACCS, Avalon, and RDKit—to represent substructural diversity, while a convolutional neural network (CNN) encoder learns expressive sequential features directly from SMILES strings. The two modalities are adaptively fused using an attention-based mechanism that enhances cross-modal interactions beyond simple feature concatenation. Experimental evaluations demonstrate that the proposed model offers a superior predictive accuracy of 0.83 surpassing the baseline direct fusion approach by 2%. Ablation studies further confirm that the combination of structural and sequential features provides complementary insights into molecular mechanisms associated with hepatotoxicity, highlighting the promise of multimodal learning for toxicity assessment.