(217e) In silico Dynamic Gene Expression Models Quantify the Impact of Regulatory Logic Gates on Network Inference
- Conference: AIChE Annual Meeting
- Year: 2015
- Proceeding: 2015 AIChE Annual Meeting
- Group: Emerging Frontiers in Systems and Synthetic Biology
Monday, November 9, 2015 - 4:33pm-4:51pm
Network inference algorithms are widely used to infer biological networks, but their capacity to resolve regulatory logic gate relationships remains largely unknown. In this study, we built an in silico model of dynamic gene expression to quantify the influence of logic gate regulation on the accuracy and confidence of network inference algorithms in order to elucidate the strengths and limitations of inferring gene regulatory networks.
Network graphs are a simple and intuitive means to represent the underlying topology and information transfer inherent to gene regulatory systems. Accurate network models provide valuable insight into functional modules and information flow, highlighting targets for therapeutic intervention. With high throughput technologies, network inference has become a conventional method to uncover regulation among genes and identify strategies to modulate system behavior. However, every method has its limitations: the DREAM consortium recently demonstrated specific network motifs, and method of inference, impact accuracy of the resulting inferred network model1. A significant source of this error is likely attributed to regulatory logic gates and its associated kinetic parameters, neither of which are considered in common practice.
An in silico model was built to simulate a gene regulatory network. Six different networks–each containing five nodes—define the activation and inhibition among genes. Genes with multiple inputs are assigned an AND, OR, NAND, or NOR logic gate relationship. A thermodynamic ODE model generates dynamic gene expression data based on the network interactions and logic gate relationships. GENIE3, a well-established inference algorithm that can deal with combinatorial and non-linear interactions, was selected to infer the putative gene regulatory network2. Each in silico data set was inferred 100 times using GENIE3 and scored against a permuted data set in order to calculate each edge’s statistical significance. The network inference and scoring objective were performed over a range of 441 logic gate kinetic parameters
The GENIE3 algorithm infers AND-gates with a higher confidence than OR-gates for 94% of the tested kinetic parameter space for the bi-fan network structure. In the feed-forward network structure, the AND-gate relationship is inferred with a higher confidence for 90% of the kinetic parameter space. In the bi-fan network, GENIE3 infers OR-gates with higher confidence when both inputs into a single gene have low kinetic parameters. However, for the feed-forward loop, the OR-gate is inferred with a higher confidence when only one edge within the network motif has a low kinetic parameter value. Lastly, GENIE3 shows little inference bias between NAND and NOR gates. Further simulations will evaluate Bayesian, regression, and correlation based inference algorithms, as well as the SUM logic gate relationship.
GENIE3, one of the top performing inference algorithms in the DREAM challenge, demonstrates bias when inferring specific logic gate relationships, limiting its predictive insight. For instance, in drug design, it is critical to consider biases in an inferred mechanism because AND-gates would be targeted and perturbed differently than OR-gates. These biases would also be critical in designing combinatorial drug targets and predicting the impact of upstream perturbations on the mechanism of interest. By considering the limitations and biases of network inference algorithms, we will be able to build more informed network models for the future design of drug targets or perturbation studies from inferred network models.
(1). Marbach, Daniel et al. “Wisdom of Crowds for Robust Gene Network Inference.” Nature Methods 9,8 (2012): 796-804
(2). Huynh-thu, Van Anh et al. “Inferring Regulatory Networks from Expression Data Using Tree-Based Methods.” PLoS ONE, 5(9): e12776, 2010.