Function Module Identification And Parameter Estimation Of Biological Network

It is shown that, for biopharmaceutical industry, the common difficultyis that the limited number of clinical medicine has not kept pace with the enormous increase in pharmacy R&D spending. Moreover, many medicines have less effective or toxicology when they are used to treat the complex diseases, for example the cancer, melancholia and cardiovascular diseases. As a result, many researchers believe that, the "lock-and-key’ model, i.e., one drug acting on a single target, faces many challenges. Therefore, it is urgently demanding to develop an effective model to identifymulti-target drug.Pharmacologyprovides the possibility of identification of multi-target by means of the biological network. With the availability of post-genomics experimental data for complete genome sequences and high-throughput, many biological processescan be characterized within the concept of network. And these networks provide different perspectives for identifying drug targets, especially for protein interaction networks and metabolic networks. Protein-protein interactions form the basis of cellular functions and operate many life processes. Abnormal regulation triggered by anomalous disruption of these interactions is the main cause of a significant number of human diseases. It is widely known that proteins seldom act alone; rather, they must interact with other proteins and aggregateinto protein modules to execute their function. The diversity of the function for proteins results in the overlapping of modules, and detection of the overlapping modules contributes to the identifying of drug multi-targets. Concentrations of metabolites are the biomarkers for diseases. In metabolic network, parameters control concentrations of metabolites. Therefore, estimating parameters in metabolic networks becomes one of the research hotspots in identification of drug targets based on networks.The paper has done some researches on the following three aspects within the framework of identifying drug target based on biological networks:Since the prior knowledge is difficult to obtain in recogniting protein overlapping modules.a soft clustering algorithm developed from the combination of random walk hard clustering and cliques’identification algorithm is proposed. Then, the algorithm is validated on a protein-protein interaction networks. The protein interaction network containing3528proteins and13475interactions is constructed by extracting the interaction data from budding yeast complete protein-protein interaction networks in DIP dataset and the interaction reliability is verified by MIPs dataset. And the results of clustering show that this algorithm can recognize overlapping modules which are validated by MIPs. By comparing with other algorithms, the modules recognized by this algorithm have high precision, especially, the multiple-role proteins shared by overlapping protein modules relate to many existing drugs in low f-measure. Furthermore, it is proved that the proteins in same module have high relationship by Semantic density measurement.A two-stage Bregman homotopy regularization inversion algorithm is developed to solve four problems with parameter estimation in metabolic network:(ⅰ) high nonlinearity;(ⅱ) largequality of parameter;(ⅲ) large parametervariations;(ⅳ) limited experimental data.In the algorithm, estimating parameters of metabolic network is regarded as an inversion.And the regular homotopy terms imported is not only used to overcome the ill-conditioned, but also to suppress noise.Two-stage regular homotopy factor regulations and combined perturbation strategydeveloped in the framework of homotopy are used topromote the convergence rate. The application of the algorithm is illustrated onArachidonic Acid metablic network, in which all unknown parameters are estimated. It is demonstrated thatthe method can get better optimization values than other popular algorithms in slightly more computational time.In real project, high-dimensional parameter estimationoften manifests as a resource constrained optimization problem. The algorithm of two-stage Bregman homotopy regularization has two shortcomings:ⅰ) solving ordinarydifferentialequations is time-consuming. ⅱ) Point estimation can not provide information about uncertainty of estimate parameters. Therefore, an algorithm which combinesthe Kriging surrogate model, the single-parameter optimization and the dynamic coordinate perturbation isformulated. The method adopts maximum of general expected improvement and mutual information as objective function and searches the multiple new points in multi-area of ’expected improvement’ function simultaneously. To avoid trapping in local optimum, single parameter optimization strategy and dynamic coordinate perturbation strategy are adopted in the procedure of searching new infill samples. Meanwhile, the domain decomposition strategy, which is based on principalcomponentanalysis, is introduced to save modeling time. Through the tests by examples, the algorithm can get the comparable objective function value to other algorithms in relatively less time. Additionally, the parameters estimated are close to their experimental values with small estimation errors in the vicinity of optimal values.