CAHAN: Drug-Disease Association Prediction Based on Cross-Attention Mechanism

In order to obtain high-quality node embeddings, a drug-disease association prediction model CAHAN based on cross-attention mechanism is proposed in this paper. After obtaining drug and disease node representations through graph neural networks using a hierarchical attention mechanism, the interactions between drugs and diseases are learned using a cross-attention mechanism, and link prediction is subsequently performed. We compare the performance of proposed model CAHAN with some state-of-the-art models. The experimental results show that the proposed model CAHAN exhibits significant improvements over those previous methods and achieves satisfactory performance, which validates the effectiveness of the model CAHAN. In order to obtain reliable negative samples, a negative sampling strategy based on drug similarity is proposed in this paper, in which the similarity between drugs is evaluated by introducing additional target information, and then the scores of drug-disease pairs are calculated to obtain reliable negative samples. Experiments are conducted on all models, and experimental results show that the performance of all models is significantly improved after applying the proposed negative sampling strategy compared with random negative sampling, which verifies the feasibility of the proposed drug similarity-based negative sampling strategy. And a case study is designed to validate the biological interpretability of CAHAN.