Reconstructing gene regulatory networks via memetic algorithm and LASSO based on recurrent neural networks

Reconstructing gene regulatory networks (GRNs) from gene expression data is an important and challenging problem in system biology. In general, the problem of reconstructing GRNs can be modeled as an optimization problem. Recurrent neural network (RNN) has been widely used for GRNs. However, in a real GRN, the number of genes is very large and the relationships between genes are usually very sparse. In this paper, we design a memetic algorithm to learn partial parameters of RNN, and develop a framework based on the least absolute shrinkage and selection operator (LASSO) to reconstruct GRNs based on RNN, which is termed as MALASSO RNN -GRN. In the LASSO, the task of reconstructing GRNs is decomposed into a sparse signal reconstructing problem. In the experiments, MALASSO RNN -GRN is applied on synthetic data and well-known benchmark datasets DREAM3 and DREAM4. The effect of parameters on MALASSO RNN -GRN is discussed, and MALASSO RNN -GRN is compared with three other algorithms which are all state-of-the-art RNN learning algorithms. The results show that MALASSO RNN -GRN performs best and is capable of reconstructing large-scale GRNs. Graphic abstract