Towards large scale genetic network modeling

Abstract

Reverse Engineering Gene Regulatory Networks (GRNs) is an important and challenging problem of Systems Biology. For its superiority in both structure and parameter learning, the S-system model framework is often chosen for GRN reconstruction. The biggest challenge in reconstructing GRNs is the data having large number of genes and only a small number of samples. This “curse of dimensionality”, along with the large number of model parameters to be learnt, makes it extremely difficult to reverse engineer even a small network. For a medium or large network, the complexity becomes enormous. In this paper, we propose a method for managing large scale GRN modeling. As first step, we propose an Affinity Propagation Based Clustering to identify appropriate clusters by grouping the genes based on their time expression profiles. In the second step, the largest cluster consisting of majority of the relevant genes is considered in full detail to act as the core of the network while the other remaining clusters, which are not so significant, are each represented by their single representative gene to obtain a reduced order GRN. In the third step, we optimize the entire network by initializing the model parameters of the genes of the largest cluster with the values obtained in the second step (which are near optimal) and proceed to optimize the entire network. The initial investigations are carried out using previously reported 20-gene synthetic network. The superiority of performance is evaluated not only using the standard metrics, namely, sensitivity, specificity, precision and F-score, but also by average mean error and by comparing the time responses with those of the actual network parameters. The results obtained are promising.