(497h) Doubly Penalized Approach for Reconstructing Biological Networks

Data-driven network reconstruction of biological systems is an essential step towards extracting information from large volumes of biological data. There are several methods developed recently to reconstruct biological networks. However, to the best of our knowledge, few systematic application-specific algorithms and methods have been developed to reconstruct dynamic biological systems. Different data properties such as level and types of the noise, level of correlation/collinearity, size of the dataset, and portion of missing data on one hand, and existing knowledge such as pre-known signaling pathways, data requisition pattern, and known time-scales of dynamic datasets on the other hand provide a good motivation to develop a new method for reconstruction of dynamic biological networks based on both data and system properties. In this work, we have developed a new method called Doubly Panelized Linear Absolute Shrinkage and Selection Operator (DP-LASSO) for reconstruction of dynamic biological networks. In this method, we have implemented principal component analysis as a supervisory level filter to extract the most informative components of the network from the dataset. In the lower level reconstruction engine, we apply LASSO with extra weights on small parameters derived via partial least squares from the first layer filter to maintain the principal components and nullify the remaining small coefficients. Simulation results show fair improvements in accuracy and sensitivity of the algorithm in reconstruction of networks. Application of DP-LASSO algorithm to experimental datasets for cell division cycle of fission yeast also shows the fidelity of the reconstructed network.