Research Of Reconstructing Gene Regulatory Network Of Yeast Based On Delayed Neuron-fuzzy Network Model

People mostly research on microarray data of genes when they study genes. They hope to find useful information through various data mining knowledge. However, we could only get little microarray data from experiments. So it's a very challenging work that how to mine information effectively under the condition of limited samples. Many researchers have tried to do this work by some of methods, they also get good results. What regulation describes is how the regulators control genes, and the regulators are also products of genes, so the problem finally is regulatory relations of genes. What gene regulatory networks include is just timing regulatory information that we infer from microarray data. We get gene expression data from biologists and reconstruct gene regulatory networks to try to reveal complex mechanism of gene regulatory activities, which we send to biologists to get verification. And the information can also give biologists some suggest that help them do latter experiment research. We need regulatory relations of genes to reconstruct gene regulatory networks and the information is hard to obtain and verify exactly. So the work reconstructing gene regulatory networks is a long time job, it is also one of the central problems in functional genomics.This paper includes two works. One is the research of cluster methods that could contain regulatory information; the other is reconstructing gene regulatory networks of yeast based on the delayed neuron-fuzzy network model, it is also the main work of this paper.Firstly I talk about the research of cluster methods which could contain regulatory information. The cluster method of this paper is Affinity Propogation based on Pearson Correlation Coefficient. Pearson correlation coefficient could describe properties of positive correlation and negative correlation, so it is very suitable for expressing regulatory relations of genes. Affinity propogation cluster method has many advantages. Such as: (1) generally the cluster methods of k-centers,k-means and the expectation maximization(EM) algorithm depend on random sampling, they select cluster numbers initially and then do prune, but prune is difficult to judge and recover. AP cluster could get cluster numbers automatically, it does not need set numbers of cluster at the beginning; (2) Markov chain Monte Carlo method could find good solutions randomly, but it can't consider more possible solutions at once. AP could take into account all points as candidate centers at same time and then select centers stage by stage. It can avoid many poor solutions. AP could also consider several possible solutions; (3) hierarchical agglomerative clustering and spectral clustering generally solve completely different problems, they recursively compare pairs of data points to find their partitioning zones. The two ways don't need all elements in a class to be similar with their centers. Thus, they are not suitable for many problems. But AP does not have these disadvantages. Therefore, AP cluster could make classes contain more regulatory information by taking Pearson correlation coefficient as similarity degree of genes. To compare AP with K-means clustering, making experiments with 20 genes and taking a standard of containing more regulatory information, the accuracy of AP and K-means are 90% and 75%, the results of AP are obviously better than K-means. This could provide a new thought for cluster analysis researching regulatory information in future.Then let's talk with emphasis on reconstructing gene regulatory networks of yeast based on the delayed neuron-fuzzy network model. Here the gene regulatory network of yeast just contains 104 genes of yeast cell cycle. There are 3 meanings about the delayed neuron-fuzzy network. The time delay character shows influence of gene X at time t-1 to gene Y at time t. The fuzziness character shows that regulation relation is product of fuzziness rules. And neural networks describe relation degree of genes. Concrete neural network structure is a neuron-fuzzy network with 6 layers. The inputs are sample data of gene X at t-1 time and gene X at t time. The output is sample data of regulated gene Y at t time. The neuron-fuzzy network describes relation of gene X and gene Y with time delay character. The weights of neural network iterate until they conform to end conditions. At this time objective function would reach minimum or extreme minimum value. Then we would find the least from these extreme minimum values. The gene X that the least value corresponds to is the strongest relation with regulated gene Y. Because of time delay character gene X may be regulatory gene (a strong regulator) of gene Y. Depending on this method we get strong regulators of 104 genes and then reconstruct their gene regulatory network. Their 3 concrete works of experiments: (1) calculating strong regulators of 97 genes, the specificity is 91.75%. My first work is to calculate strong regulators of 97 genes, only 8 results do not belong to their own regulator set, the accuracy is 91.75%; (2) calculating the strong regulators of 7 genes which do not have regulator records in the SGD database. The result does not be clearly illustrated in related databases and references, so it needs to be validated by biological experiments; (3) using strong regulators of 104 genes and fuzzy rules to reconstruct a gene regulatory network of yeast. Every gene we research has about 6 regulators in the SGD database. The method can compute regulators'degree of each gene, as well as extract the strong regulator whose regulation phenomenon is most obvious. Because of properties of fuzziness rules it could describe regulatory situation of genes. We compare obtained regulatory relations with regulatory modules in SGD database. There are 70 edges and 59 edges of them are consistent with records of database. The accuracy is 84.28%. Thus, summing up the method and its experiment results we could get conclusion that this method is suitable for reconstructing gene regulatory networks. And this is a new thought. It is a contribution of gene regulatory networks research. The content of this paper includes data merging,cluster analysis,gene regulatory networks construction and complex network. It introduces some knowledge of data merging,cluster analysis and complex network and relation of the 3 aspects with gene regulatory network. It emphatically tells how to reconstruct gene regulatory network of yeast based on delayed neuron-fuzzy network. Organization structure of this paper is as follows:Chapter I: Exordium. This chapter simply introduces gene regulatory network and its research situation.Chapter II: Related theories and research contents of gene regulatory network reconstruction. This chapter introduces knowledge of data merging and cluster analysis with their function in gene regulatory network construction. It in detail tells Affinity Propogation cluster method based on Pearson Correlation Coefficient. It also introduces biology connotation of gene regulatory network and some methods for reconstructing gene regulatory network.Chapter III: Neuron-fuzzy network and its function in gene regulatory network reconstruction. This chapter introduces basic knowledge of neuron-fuzzy network with its ability and advantage for reconstructing gene regulatory network.Chapter IV: Reconstructing gene regulatory network of yeast based on delayed neuron-fuzzy network. This chapter shows that why I use delayed neuron-fuzzy network to reconstruct gene regulatory network. It tells model,algorithm,results and analysis in detail.Chapter V: Summary and Prospect. This chapter sums up my work and introduces some outlook for development of research of gene regulatory network.