Improving Reconstruction Efficiency of Time-varying Gene Regulatory Networks
No Thumbnail Available
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Motivation: Flow of regulatory-information between genes are often modelled as a ‘time-varying gene regulatory network’. Here, nodes represent genes. A directed edge from gene X to gene Y at time-interval T is interpreted as ‘X regulates Y at time-interval T’. A widely-used approach is to reverse-engineer or ‘reconstruct’ the network from time-series ‘gene expression data’ using computational algorithms called ‘time-varying gene regulatory network reconstruction algorithms’. Problem Formulation: We apply the existing algorithms on three benchmark datasets. It is found that an algorithm named ‘ARTIVA’ provides state-of-the-art correctness. However, ARTIVA requires months for processing large-scale datasets with thousands of variables. Hence, the objective of the thesis is set to designing algorithms that can process large-scale datasets in acceptable time, while remaining competitive to ARTIVA in correctness. Contributions: Towards that objective, the thesis makes four contributions. In the first contribution, an algorithm named ‘TGS’ is proposed. It can process large-scale datasets in minutes. However, its correctness is not competitive to that of ARTIVA. In the second contribution, we propose an algorithm named ‘TGS-Plus’ (TGS+). It is as fast as TGS, and competitive to ARTIVA in correctness. Nonetheless, TGS+’s main-memory requirement grows exponentially at (2^n *(n + 1)), where n is the number of genes in a dataset. As a result, it tends to encounter memory-related issues for large-scale datasets. In the third contribution, an algorithm named ‘TGS-Lite+’ is proposed. It preserves the same time-complexity and correctness as those of TGS+. At the same time, TGS-Lite+’s main-memory requirement grows linearly at (2n + 2) only. Nevertheless, we observe that TGS-Lite+ is unable to capture ‘transient’ edges – the edges that remain active for a small number of time intervals, yet, can have cascading effects. For this reason, capturing transient edges is critical to understand the underlying interactions. In the fourth contribution, we propose an algorithm named ‘TGS-T-Lite+’. It is able to capture twice as many edges as that of TGS-Lite+ for larger benchmark-datasets. Conclusion: In conclusion, this thesis advances the state of ‘time-varying gene regulatory network reconstruction’ algorithms in terms of time-efficiency, memory-efficiency, and correctness. It also discusses future opportunities.
Description
Supervisor: Ashish Anand
Keywords
COMPUTER SCIENCE AND ENGINEERING