SimH: a novel representation learning model with activation and projection mechanisms for COVID-19 knowledge bases.

The emergence of coronavirus disease 2019 (COVID-19) has had a significant impact on healthcare and the economy. To understand the COVID-19 disease mechanism and the related biological functions in the short term, both clinicians and scientists are making every effort to find an efficient way to collect and explore the vast amount of COVID-19-related knowledge. Representation learning has been highlighted as a promising method to construct a COVID-19 knowledge graph. However, most existing representation learning models do not perform very well when dealing with the COVID-19 knowledge graph because of its low-connected star-like structure and various nonlinear relationships. In this study, we propose a novel representation learning model called translation on hyperplanes with an activation operation and similar semantic sampling (SimH) for COVID-19 knowledge graphs. Specifically, the activation operation is designed to provide additional interaction features for low-in-degree entities by interaction feature permutation and share relation-specific partitions of pairwise interactions by an activation vector. As a result, problems that fewer features are captured from low-in-degree entities are alleviated. Moreover, hyperplane projection is introduced to the distance-based scoring function so that nonlinear relationships can be modeled while the lower complexity is maintained, as compared to other nonlinear models. To consider that negative sampling can improve the embedding quality of fact triples, a negative triplet sampling method that adaptively replaces entities with similar semantics is introduced to generate reliable negative triplets. Extensive experiments are conducted on the COVID-19-Concepts dataset. The experimental results show that our SimH model achieves significant improvements in prediction and classification accuracy over existing knowledge representation learning models.

DRPADC: A novel drug repositioning algorithm predicting adaptive drugs for COVID-19.

Given that the usual process of developing a new vaccine or drug for COVID-19 demands significant time and funds, drug repositioning has emerged as a promising therapeutic strategy. We propose a method named DRPADC to predict novel drug-disease associations effectively from the original sparse drug-disease association adjacency matrix. Specifically, DRPADC processes the original association matrix with the WKNKN algorithm to reduce its sparsity. Furthermore, multiple types of similarity information are fused by a CKA-MKL algorithm. Finally, a compressed sensing algorithm is used to predict the potential drug-disease (virus) association scores. Experimental results show that DRPADC has superior performance than several competitive methods in terms of AUC values and case studies. DRPADC achieved the AUC value of 0.941, 0.955 and 0.876 in Fdataset, Cdataset and HDVD dataset, respectively. In addition, the conducted case studies of COVID-19 show that DRPADC can predict drug candidates accurately.