| According to the central dogma,many biological traits are related to genes.However,genes in living organisms usually interact with other genes,forming gene networks to collaboratively perform specific biological functions,rather than perform functions alone.For example,intracellular gene co-expression networks and intercellular signaling networks play crucial roles in life processes.Besides,gene networks often change by the biological environment or conditions.Thus,constructing gene networks and inferring patterns that change with the environment can shed light on relevant biological mechanisms.It is difficult to dynamically observe gene networks on a large scale only by experimental methods.How to infer relevant networks from high-throughput transcriptional data with computational methods is an important issue of common concern in the field of biomedical,statistical,machine learning and so on.Based on the theory of probabilistic graphical model and structured sparse learning,from transcriptional data,mathematical models and calculation methods for inferring intra-cellular and inter-cellular gene networks,and discovering the pattern of network difference between different states were put forward in this thesis,which provide computational support for the analysis of related biological mechanisms from the perspective of gene networks.Firstly,based on Gaussian graphical model and structured sparse learning,we proposed a novel method to infer the co-expression networks of cancer subtypes from bulk transcriptional sequencing data.The innovation of this method lies in the following two aspects.By combining with Gaussian mixture model,this method can not only infer the co-expression networks of cancer subtypes,but also identify the subtypes of cancer samples.By integrating normal samples and cancer samples,the problem of high dimensionality and small size samples in gene network inference can be effectively alleviated,because of the increased sample size.Simulation results show that the proposed method is better than the competing methods in subtype recognition and gene network estimation.The application on breast cancer data sets shows that this method can better cluster breast cancer samples,and the key genes in the inferred gene networks play important roles in the development and characterization of breast cancer subtypes.Secondly,the interactions between genes often change by the change of disease states,and the analysis of single gene co-expression network can not clarify the change pattern of gene interaction between different disease states.Therefore,based on Gaussian graphical model and sparse learning theory,a novel differential co-expression network inference method was proposed from bulk transcriptional sequencing data,to explore the changes of gene networks in different disease states.On the one hand,by defining the differential co-expression network as the difference between partial correlation coefficient matrices rather than the difference of precision matrices,the model effectively eliminates the pseudo-differential edges caused by the change of conditional variance.On the other hand,the prior information obtained from multiple hypothesis testing allows to impose different penalties on the interactions between different genes,which improves the performance of the model.The model is better than competing methods in simulating data.For the differential co-expression networks inferred by the proposed method,key genes in the differential network between Luminal A and basal-like breast cancers,as well as the differential network between acute myeloid leukemia and normal samples,are essential for characterizing different disease states.Furthermore,although gene co-expression networks can describe gene interactions inside cells,inferring signaling networks for intercellular signaling is also very important for studying the life activities of multicellular organisms.Thus,we proposed a new method for simultaneously inferring intracellular and intercellular signal networks from single-cell spatial transcriptional data based on Bayesian networks.Different from the previous signal network inference method which only considers the abundance of signal molecules,our proposed method can infer a more accurate signal network by integrating cell location information.The prior transmission network of transcription factor-ligand-receptor-transcription factor-target gene were collected and integrated from multiple data sources.Then a Bayesian network model based on the prior network was established,and we applied a neural network to solve the Bayesian network model.By combining with neural network,the solution of a series of conditional probability in Bayesian network model is avoided.By combining with biological network,neural network can be interpreted.Finally,the method was applied to the mouse embryo data.The biomolecules that play important roles in the signal transduction process were identified through key node analysis based on the learned signaling networks.The difference of the mechanism between the signal sending cells and the signal receiver cells were explored by differential signaling network analysis.In conclusion,based on probabilistic graphical model,from transcriptional data,three gene network inference methods were proposed in this thesis,which are of great significance for exploring the occurrence and development of diseases and the development of organs. |
