Multi-modal memetic framework for locating saddle points, with application to biomolecular systems : water clusters and ring-deficient covalently-bonded small molecules

• Main Author: Mostafa Ellabaan
• Other Authors: Ong Yew Soon
• Format: Theses and Dissertations
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/54820
• Institution: Nanyang Technological University

Saddle points have important applications in different fields of science and engineering, including transition states identification, path-planning and robotics navigation. Identifying such points using memetic approaches is a field, remained yet to be well-investigated. In this dissertation, we propose a multi-modal memetic framework, composing of 1) a fitness function that transforms the problem landscape to favor saddle points instead of local or global optima, 2) evolutionary operators of crossover and mutation for efficient exploration of the problem landscape, 3) niching operator based on the valley-adaptive clearing scheme for maintaining multiple saddle points during the evolutionary process, and 4) memetic operator consisting of saddle point local searcher for attaining precise saddle points. Experimental study on synthetic benchmark functions of varying complexity demonstrates the efficacy of the proposed framework for locating multiple first-order saddle points of higher precision, when compared to its evolutionary compeers. From synthetic benchmark problems to real-world problems, the multi-modal memetic framework is applied to systems representing the two main streams of the molecular systems, namely, water clusters and small ring-deficient molecules. In particular, the framework has been extended to efficiently explore the landscape of biomolecular systems for transition states, especially when using first-principle calculations which describe mathematically the interaction among atoms and their subatomic structures in biomolecular systems. For water cluster, the proposed framework is extended with water-specially-designed reproduction operators. The application of the proposed framework on water cluster has shown not only to uncover those previously reported but also to establish newly discovered transition states of size 2-4 water molecule, with the minimal computational cost among to its compeers. For ring-deficient covalently-bonded small molecules, the multi-modal memetic framework is extended with tree-based evolutionary operators, to induce configurational changes on the molecules while maintaining both the tree-structure and the molecular properties unchanged. Comparisons to recent advancements in related state-of-the-art algorithms on several small, yet very important, molecules showed that the multi-modal memetic framework not only uncovered the largest set of transition states, but also attained the achievements at significantly lower computational costs. With further analysis on the transition states uncovered by the framework, it has been found out that the number of oxygen atoms and rotatable bonds play a key role in the flexibility of the molecules, allowing larger number of transition states to be uncovered while carbon atoms have been observed to limit the flexibility of the molecule, decreasing the number of transition states to be uncovered. The number of hydrogen atoms as well as total number of atoms, in contrast, has not shown any significant impact. The framework, extended for small ring-deficient molecules, has been, equally important, considered for locating glutamic acid stereoisomers, which later uncovered to have a modulation effect on the glutamatergic systems. Such findings open new door for neuroscientist to study the impact of the glutamic acid stereoisomers on the function and dysfunction of the glutamatergic systems. From this study, we conclude that the memetic computing paradigm is the most efficient evolutionary paradigm ever applied for saddle point identification, and incorporation of domain knowledge into the evolutionary operators improves search performance significantly, constraining search process to explore only the feasible region of the landscape.