Robust and Adaptive Deep Model Ensemble Framework Fine-tuned by Structural Information for Drug-Target Interactions

Abstract

In the fields of new drug development and drug repositioning, drug-target interactions (DTI) play a pivotal role. Although deep learning models have already made significant contributions in this domain, the state-of-the-art models still exhibit shortcomings in predictive performance and issues of false-negative errors. Based on these observations, we constructed a streamlined yet effective base learner model. With our designed adaptive feature weight network, the model can capture key features within drugs (targets). Furthermore, by cross-partitioning the training data, multiple base learners are integrated into a powerful ensemble model named EADTN. The performance of the model is further enhanced as the number of base learners increases. Additionally, we employed a single-linkage clustering algorithm to cluster drugs and proteins and leveraged this clustering information to fine-tune the base learners, which elevates the value of EADTN in real-world applications like drug repositioning and targeted drug development. Our designed substructure importance ranking method also demonstrates the model's exceptional capability to recognize key substructures. Benefiting from the model's low generalization error capability, we successfully identified false-negative samples within the dataset, revealing new interaction relationships. Experimental results indicate that EADTN consistently outperforms existing state-of-the-art models across multiple datasets. More importantly, the ensemble learning and clustering fine-tuning approaches adopted by our model offer a fresh perspective for related fields.