Work

Evolution Strategies for Learning Sparse Matrix Representations of Gene Regulatory Networks

Year: 2020
Type: dissertation
Abstract: Currently, a massive amount of temporal gene expression data is available to researchers, which makes it possible to infer Gene Regulatory Networks (GRNs). Gene regulatory networks are theoretical models to represent excitatory and inhibitory interactions between genes. GRNs are useful in understanding how genes function, and hence they are also useful in pharmaceutical and other applications in biology and medicine. However, despite the importance of GRNs, the process of inferring GRNs from observational data is very difficult. This thesis applies evolutionary algorithms to the problem of GRN inference. We propose a novel evolutionary algorithm: hierarchical evolution strategy (HES) to target the specific difficulties in GRN inference. We propose a sparse matrix representation of GRN to account for sparse connectivity in biological gene interactions. Unlike traditional evolution strategies, we divide our optimization into two concurrent processes: connectivity construction and numerical optimization. In each generation, we first establish connectivity structure of the GRN. Inside the same generation, we apply a secondary ES to find the best numerical values with those fixed connections. We also propose a hybrid crowding method to maintain high population diversity while applying the evolutionary algorithms. High population diversity leads to broader exploration area in the search space, therefore preventing premature convergence. The results obtained show that the proposed HES outperforms other algorithms, and has the potential to scale up to realistic problems with thousands of genes.
Institution: Syracuse University
Language: English
Open Access status: green
OpenAlex ID: W3167632237
Cites: 32
Cited by: 1
Related to: 10