Research Of Multi-level Biomolecular Network Construction And Its Visualization Platform Development

Biological systems are composed of complex interactions between various types of biomolecules.Various existing high-throughput technologies can identify molecular interactions and express them through different network models.As an important method of computational biology,biological network modeling can effectively integrate scattered experimental data to study complex diseases from a system level.However,existing network biology researches still generally focus on molecular networks at the genome,transcriptome,metabolome or proteome level separately.In fact,a complex disease is not caused by a single gene mutation or lack of interactions between a single pair of genes.Its pathogenic mechanism is extremely complicated,usually caused by abnormal interactions between multiple genes or biological molecules of different levels.In recent years,the continuous accumulation of biological knowledge and the emergence of multi-omics data have provided opportunities for the establishment of multi-level biomolecular networks and the exploration of complex disease pathologies.In the first part of the thesis,we proposed a method called DeepIDA,which is based on data fusion and deep neural networks,to predict isoform-disease associations.Alternative splicing produces different isoforms from the same gene locus,it is an important mechanism for regulating gene expression and proteome diversity.The abnormality of alternative splicing is closely related to many complex diseases.However,the existing isoform-disease association research generally stays at the level of wet experiments.This is mainly due to the lack of disease label data at the isoform level.To overcome this challenge,we utilized isoform-gene association to transfer disease labels at the genetic level to the isoform level.On this basis,in order to accurately predict the isoform-disease association,DeepIDA first established a multilevel biomolecular network by collecting and aligning multi-level biomolecule association data.Secondly,in order to improve the prediction accuracy of the model,DeepIDA also fused isoform basic feature data: isoform sequence data and isoform expression data.DeepIDA uses two parallel DNN sub-networks with different structure to synchronously extract heterogeneous features from multi-level biomolecular networks and isoform basic feature data and finally fuse the above features for isoformdisease association prediction.In addition,isoform-disease associations prediction is a typical class imbalance problem.In deep learning,the majority class will control the direction of the gradient loss and thus dominate the overall model learning direction.In this paper,focal loss function is utilized instead of the traditional cross-entropy loss function to address this challenge.Focal loss can reduce the weight of a large number of well-classified negative samples after balancing the weights of the positive and negative classes,thereby making the model pay more attention to samples that are more difficult to classify.The experimental results on public datasets show that compared with traditional machine learning methods,DeepIDA can significantly improve the prediction effect in different evaluation metrics.Because isoform-disease association prediction research is still in its infancy,there are currently no relevant online platforms.However,since the online platforms for gene-disease association prediction have become very popular,we borrowed a lot of design ideas from the visualization analysis platform for gene-disease association and analyzed the common needs in the disease association prediction problem at the isoform level and the gene level.In addition,we also analyzed the individual needs of the multilevel biomolecular network visualization display.In the second part of the thesis,we carried out detailed feasibility study and requirement analysis of the isoform-disease association visualization analysis platform.Then we described the system architecture design and database design.Finally,based on the node.js express framework,we developed online search module,user module,calculation module and result display module.Among them,the online search module supports isoform ID search and isoform sequence search.The user module includes user upload data function and task management function.The calculation module includes pre-processing function for users' uploaded data and online calculation for isoform-disease association function.The result display module includes text result display function and visualization result display function.