Predicting biomolecule-disease associations based on machine learning approaches

Abstract

Biomolecules, such as microRNAs (miRNAs), circRNAs, and long non-coding RNAs (lncRNAs) interact and regulate each other to sustain a rich variety of life activities. The dysregulation of one kind of biomolecule may affect the stability of the whole biomolecule regulation system, which eventually causes diseases. Identifying the biomolecule-disease associations (BDAs) can uncover the complex disease pathogenesis to greatly facilitate the diagnosis, treatment, prognosis, and prevention of diseases. Due to the low efficiency of exploring BDAs by biological experiments, prioritizing disease-related biomolecule candidates by computational approaches is imperative which can serve as an auxiliary tool to select high probability candidates.

The performance of a computational model usually relies on two factors: the biological data evidence involved in the model and the computational algorithm used in the model. The increasing amount of omics data and the development of machine learning methods provide an unprecedented opportunity for the improvement of machine learning-based BDAs prediction. This thesis combines the characteristics of omics datasets and the advantages of different machine learning methods to construct some novel machine learning-based BDA prediction models to explore the deep underlying of complex nonlinear associations.

The thesis starts with a comprehensive review of machine learning-based approaches for BDAs prediction, which includes multi-view data sources, data integrating strategies, feature representation methods, and machine learning-based prediction approaches. Then, six different machine learning-based BDA prediction models are presented, and the BDA mainly focuses on circRNA-disease association (CDA), miRNA-disease association (MDA), and lncRNA-disease association (LDA). The first model is developed for predicting CDAs by considering the issue of overlooking the global information in similarity networks. The random walk with restart gets the global relationship information to construct the feature representations, and then logistic regression is applied to predict the novel associations. The second model is a deep belief network (DBN)-based matrix factorization for MDA prediction. It factorizes the miRNA-disease adjacency matrix into two matrices by DBNs. Specifically, this model combines unsupervised DBN pre-training and supervised backpropagation training for fine-tuning the parameters. The third model further improves the performance of MDAs prediction by a variational graph auto-encoder (VGAE) model. The graph convolutional neural network naturally incorporates the node features from the graph structure while the variational autoencoder makes use of latent variables to predict associations from the perspective of data distribution. After that, considering the combination of the linear feature representations and non-linear feature representations, another MDA prediction model, variational graph auto-encoder with matrix factorization (VGAMF), is developed. VGAMF gets the linear representations by a non-negative matrix factorization from the MDA matrix, while it obtains the non-linear representations from two comprehensive similarity networks with two variational graph auto-encoders, respectively. The fifth model is a factorization machine-based deep neural network with binary pairwise encoding (DFMbpe) for both MDA prediction and LDA prediction. This proposed binary pairwise encoding can consider the interdependence of features even though they never appear in the same sample, and the architecture emphasizes both low-order and high-order feature interactions simultaneously. Finally, the sixth model is a Laplacian regularized deep factorization machine (DeepFM) model for MDA prediction. To consider the limitation of standard DeepFM that overlooks the relationships among items in the same field, this model regularizes the items embedding features via their similarity-based Laplacians. Besides, to judiciously train this model, Laplacian eigenmaps are applied to initialize the weights in the embedding layer.

In summary, this thesis proposes six novel machine learning-based BDA prediction models to either address the limitation of some specific machine learning methods, or to construct novel models by combining the advantages of several machine learning methods. Experimental results show that the proposed methods can improve the performance of BDA prediction.