ABSTRACT Forecasting drug‐target interactions (DTI) plays a critical role in accelerating drug discovery. Substantial cost savings arise when accurate predictions streamline repurposing and flag side effects. While knowledge graph completion remains popular, bottlenecks regarding computational efficiency and inefficient data integration plague current multimodal strategies. High costs are also incurred by traditional models during retraining. To surmount these obstacles, we present AE‐ multi‐modal drug‐target interactions (MMDTI) (confidence‐aware ensemble multimodal DTI), a framework merging efficient fine‐tuning with dynamic multimodal data learning and fusion. By utilizing parameterized hypercomplex multiplication in lightweight adapter layers, we drastically lower complexity. Distinct from methods treating predictions equally, our ensemble mechanism weights outputs based on confidence. Benchmarking confirms AE‐MMDTI surpasses current techniques, improving Recall and F1 score by 7.4% and 10.1% against MultiDTI, respectively. This approach not only enhances performance but also minimizes resource usage, promoting sustainability. Supporting data are available upon request.
Xinyuan et al. (2026) studied this question.