A Denoising Adversarial Model Based on Hyperellipsoidal Knowledge Representation Learning for DTI Prediction

The prediction of drug-target interactions (DTI) is of critical importance in the field of drug discovery. Knowledge representation learning is commonly used to embed high-dimensional, sparse drug knowledge graphs into low-dimensional, dense vector spaces. Existing methods primarily rely on node features of drug knowledge graphs, which often neglect inherent biases introduced by network characteristics, such as noise, long-tail distribution, data sparsity, and complex relations (1-N, N-1, N-N). This oversight leads to incomplete semantic representation of knowledge graphs during the embedding process, thereby resulting in suboptimal DTI prediction performance. To address these challenges, we propose a novel model termed DAH-DTI, which comprises four components: knowledge graph denoising, high-quality negative sample generation, hyperellipsoid knowledge graph embedding based on Mahalanobis distance, and DTI link prediction. Specifically, the knowledge graph is first denoised to filter out noisy triples and obtain a high-quality gold dataset. Subsequently, adversarial negative sampling is incorporated during pre-training to moderately expand the number of triples. Next, the tail entities of the triples are embedded into a predefined hyperellipsoid for model training. Finally, the trained model is applied to all potential unknown drug-target pairs through link prediction to predict potential DTIs. Experimental results demonstrate that by effectively addressing challenges including data noise, long-tail distribution, data sparsity, and complex relations. The DAH-DTI achieves significant improvements in ACCU, REC, MRR, and Hit@10 on a multi-level phenotype-drug-molecular knowledge graph. Furthermore, molecular docking simulations validate the reliability of DAH-DTI predictions for DTI involving the AR target. In summary, DAH-DTI provides a promising approach to mitigate inherent biases caused by network characteristics and accurately predict potential drug-target interactions.