Evolutionary Properties of Neural Networks: Exploration of Robustness and Evolvability in the Genotype-Phenotype Map

Abstract

Evolution is a fundamental and crucial part of life that hinges on two central properties: robustness and evolvability. Robustness is required to maintain essential traits despite mutations while evolvability produces novel traits that might prove beneficial in survival. While both robustness and evolvability are necessary, embracing them simultaneously seemingly leads to an inherent conflict due to their antagonistic goals driven by a shared mechanism, mutation. This conflict has been partially resolved by distinguishing phenotypic robustness and genotypic robustness. However, the general relationship between robustness and evolvability across various biological systems and how they are simultaneously facilitated are still unknown. Here, we study the relationship between robustness and evolvability of neural circuits using a genotype-phenotype map (GP map) which describes interactions between genes that ultimately determine the phenotype. We focus on canalization, the buffering of phenotypes against internal and external variations, to explain the relationship between robustness and evolvability. The pyloric circuit of the crustacean stomatogastric ganglion is used as the model system to study the relationship between robustness and evolvability. We use information theory to quantify robustness and evolvability and a novel simulation-based inference technique to examine the GP map which is used as an analog to a phenotype landscape. While robustness and evolvability are not directly correlated globally in the pyloric circuit, they are found to be linear with respect to collective traits of the pyloric rhythm as well as to individual traits. Robustness and evolvability are seen to be compromising around a region in the GP map, and we hypothesize that the pyloric rhythm is canalized, leading to local robustness and global evolvability. Multiple indicators of canalization and explicit calculation confirms that the pyloric rhythm is canalized in this neural circuit. Investigating the topology of the GP map reveals that the local landscape around the pyloric rhythm is flatter, facilitating robustness. Deviating from the canalized region leads to highly variable and steep landscape of the GP map, promoting evolvability. Gradient directions of the GP map at pyloric points are seen to be correlated, hinting at implicit mechanisms to preserve the biologically relevant behavior. In conclusion, this work demonstrates that the structure of the genotype-phenotype map in the stomatogastric ganglion facilitates robustness and evolvability through canalization. This study also establishes a scalable and generally applicable method to examine robustness and evolvability in any system involving mechanistic models.

Popular Abstract

Evolution is a vital part of life where organisms better suited to their environment survive to pass on their genes. Genes encode and largely determine how an organism ultimately develops, and thus they are considered to be the main factors of evolution. To have a better chance at survival, organisms must withstand genetic mutations to consistently produce important characteristics (robustness) such as functional legs for mobility, the correct set of organs, or the ability to communicate to fellow members in the same species. At the same time, to be able to adapt to environmental changes, they must innovate their physique or behavior (evolvability) that gives them a better chance at survival. Both robustness and evolvability stem from mutations, but their aims are contradicting. Robustness tries to maintain traits while evolvability changes them, and how biological systems promote both robustness and evolvability in general is still unknown.

This thesis studies the relationship between robustness and evolvability and the underlying mechanisms that promote these two evolutionary properties. We examine their relationship in neural circuits which are core components of many living beings that prompt behavior among many other crucial functions. As neural circuits largely determine organisms’ behaviors, evolution of behavior means evolution of the neural circuit responsible for those behaviors. We use a model of a neural network responsible for controlling stomach movements in crustaceans, a group of animals that includes crabs and lobsters. This network performs its function by generating regular and stable rhythm called the pyloric rhythm and has shown robustness and potential for evolvability in previous works. Robustness and evolvability are not found to have a direct relationship in this system: they are neither opposing or supporting each other. However, we see that the pyloric rhythm produced by
the circuit is canalized, meaning that this behavior is protected against small mutations.

To understand how the neural circuit canalizes the pyloric rhythm, we study how model parameters link to behavior. The structure of this transformation confirms that the underlying mechanisms of the circuit indeed canalize the pyloric rhythm. We find that this neural circuit supports robustness around the pyloric rhythm, the key biological behavior, through canalization while deviating from this behavior leads to increased evolvability. Ultimately, we show that the relationship between robustness and evolvability can be understood by studying the transformation from genetic information to the resulting traits. Additionally, this work establishes a generally applicable method that can be used to study evolutionary properties in any model at any biological scale.

• author: Cho, Daesung
• supervisor: Jan Clemens; Victor Olariu
• organization: Computational Biology and Biological Physics - Undergoing reorganization
• course: FYTM04 20212
• year: 2022
• type: H2 - Master's Degree (Two Years)
• subject: Physics and Astronomy
• keywords: robustness, evolvability, genotype-phenotype map, canalization, Bayesian inference, simulation-based inference, information theory, neural networks, pyloric rhythm
• report number: LU-TP 22-08
• language: English
• id: 9076414
• date added to LUP: 2022-03-14 16:28:33
• date last changed: 2022-03-14 16:28:33