The combination of multiple drugs is a significant therapeutic strategy that can enhance treatment effectiveness and reduce medication side effects. However, identifying effective synergistic drug combinations in a vast search space remains challenging. Current methods for predicting synergistic drug combinations primarily rely on calculating drug similarity based on the drug heterogeneous network or drug information, enabling the prediction of pairwise synergistic drug combinations. However, these methods not only fail to fully study the rich information in drug heterogeneous networks, but also can only predict pairwise drug combinations. To address these limitations, we present a novel Synergistic Multi-drug Combination prediction method of Western medicine based on Heterogeneous Network representation learning with Contrastive Learning, called SMC-HNCL. Specifically, two drug features are learnt from different perspectives using the drug heterogeneous network and anatomical therapeutic chemical (ATC) codes, and fused by attention mechanism. Furthermore, a group representation method based on multi-head self-attention is employed to learn representations of drug combinations, innovatively realizing the prediction of synergistic multi-drug combinations. Experimental results demonstrate that SMC-HNCL outperforms the state-of-the-art baseline methods in predicting synergistic drug pairs on two synergistic drug combination datasets and can also effectively predict synergistic multi-drug combinations.

Network embedding is a very important task to represent the high-dimensional network in a low-dimensional vector space, which aims to capture and preserve the network structure. Most existing network embedding methods are based on shallow models. However, actual network structures are complicated which means shallow models cannot obtain the high-dimensional nonlinear features of the network well. The recently proposed unsupervised deep learning models ignore the labels information. To address these challenges, in this paper, we propose an effective network embedding method of Structural Labeled Locally Deep Nonlinear Embedding (SLLDNE). SLLDNE is designed to obtain highly nonlinear features through utilizing deep neural network while preserving the label information of the nodes by using a semi-supervised classifier component to improve the ability of discriminations. Moreover, we exploit linear reconstruction of neighborhood nodes to enable the model to get more structural information. The experimental results of vertex classification on two real-world network datasets demonstrate that SLLDNE outperforms the other state-of-the-art methods.

Due to the widespread availability of implicit feedback (e.g., clicks and purchases), some researchers have endeavored to design recommender systems based on implicit feedback. However, unlike explicit feedback, implicit feedback cannot directly reflect user preferences. Therefore, although more challenging, it is also more practical to use implicit feedback for recommender systems. Traditional collaborative filtering methods such as matrix factorization, which regards user preferences as a linear combination of user and item latent vectors, have limited learning capacities and suffer from data sparsity and the cold-start problem. To tackle these problems, some authors have considered the integration of a deep neural network to learn user and item features with traditional collaborative filtering. However, there is as yet no research combining collaborative filtering and content-based recommendation with deep learning. In this paper, we propose a novel deep hybrid recommender system framework based on auto-encoders (DHA-RS) by integrating user and item side information to construct a hybrid recommender system and enhance performance. DHA-RS combines stacked denoising auto-encoders with neural collaborative filtering, which corresponds to the process of learning user and item features from auxiliary information to predict user preferences. Experiments performed on the real-world dataset reveal that DHA-RS performs better than state-of-the-art methods.

Biological network alignment is an important research topic in the field of bioinformatics. Nowadays almost every existing alignment method is designed to solve the deterministic biological network alignment problem. However, it is worth noting that interactions in biological networks, like many other processes in the biological realm, are probabilistic events. Therefore, more accurate and better results can be obtained if biological networks are characterized by probabilistic graphs. This probabilistic information, however, increases difficulties in analyzing networks and only few methods can handle the probabilistic information. Therefore, in this paper, an improved Probabilistic Biological Network Alignment (PBNA) is proposed. Based on IsoRank, PBNA is able to use the probabilistic information. Furthermore, PBNA takes advantages of Contributor and Probability Generating Function (PGF) to improve the accuracy of node similarity value and reduce the computational complexity of random variables in similarity matrix. Experimental results on dataset of the Protein-Protein Interaction (PPI) networks provided by Todor demonstrate that PBNA can produce some alignment results that ignored by the deterministic methods, and produce more biologically meaningful alignment results than IsoRank does in most of the cases based on the Gene Ontology Consistency (GOC) measure. Compared with Prob method, which is designed exactly to solve the probabilistic alignment problem, PBNA can obtain more biologically meaningful mappings in less time.