A population-based evolutionary search approach to the multiple minima problem in de novo protein structure prediction

Abstract

BackgroundElucidating the native structure of a protein molecule from its sequence of amino acids, a problem known as de novo structure prediction, is a long standing challenge in computational structural biology. Difficulties in silico arise due to the high dimensionality of the protein conformational space and the ruggedness of the associated energy surface. The issue of multiple minima is a particularly troublesome hallmark of energy surfaces probed with current energy functions. In contrast to the true energy surface, these surfaces are weakly-funneled and rich in comparably deep minima populated by non-native structures. For this reason, many algorithms seek to be inclusive and obtain a broad view of the low-energy regions through an ensemble of low-energy (decoy) conformations. Conformational diversity in this ensemble is key to increasing the likelihood that the native structure has been captured.MethodsWe propose an evolutionary search approach to address the multiple-minima problem in decoy sampling for de novo structure prediction. Two population-based evolutionary search algorithms are presented that follow the basic approach of treating conformations as individuals in an evolving population. Coarse graining and molecular fragment replacement are used to efficiently obtain protein-like child conformations from parents. Potential energy is used both to bias parent selection and determine which subset of parents and children will be retained in the evolving population. The effect on the decoy ensemble of sampling minima directly is measured by additionally mapping a conformation to its nearest local minimum before considering it for retainment. The resulting memetic algorithm thus evolves not just a population of conformations but a population of local minima.Results and conclusionsResults show that both algorithms are effective in terms of sampling conformations in proximity of the known native structure. The additional minimization is shown to be key to enhancing sampling capability and obtaining a diverse ensemble of decoy conformations, circumventing premature convergence to sub-optimal regions in the conformational space, and approaching the native structure with proximity that is comparable to state-of-the-art decoy sampling methods. The results are shown to be robust and valid when using two representative state-of-the-art coarse-grained energy functions.

Generating, Maintaining, and Exploiting Diversity in a Memetic Algorithm for Protein Structure Prediction

“…An evolutionary algorithm employing crossover and mutation operators in a fragment-assembly context has also been described (Olson et al, 2013). Saleh et al (2013) also describe the use of evolutionary and memetic algorithms involving fragment assembly, and show that the memetic algorithm can reduce the likelihood of deep descent into local optima on the energy landscape, which is known to be detrimental to predictive accuracy (Molloy et al, 2013), and has been an issue for some of the above approaches. Owing to inaccuracies in typical low-resolution scoring functions, some first methods have now started to consider structural diversity at strategic points during the search, e.g.…”

Multi-Objective Stochastic Search for Sampling Local Minima in the Protein Energy Surface

“…The incorporation of domain-specific insight even to basic EAs such as basin hopping (BH) has been shown to result in competitive decoy sampling algorithms [23,28] (BH has been shown effective even in other structure modeling settings beyond template-free structure prediction [25]). Our initial comparison of BH with population-based EAs in [31,32] indicated that population-based EAs have higher exploration capability and prompted the further pursuit of population-based EAs in our work.…”

“…Studies applying EAs for decoy sampling are thus limited to very small proteins or toy models; currently, EAs are not competitive against state-of-the-art methods in template-free structure prediction. In recent work we have provided a pathway for employing EAs in template-free structure prediction by incorporating coarse-grained representations and molecular fragment replacement, improving the exploration capability of single-trajectory or population-based EAs [24,28,29,32] as well as tree-based search algorithms [26,33]. In this paper we build upon this foundation and propose a more powerful novel evolutionary search algorithm.…”

Informatics-driven Protein-protein Docking

“…Our ongoing work is focusing on obtaining more information about the predictive capability of the method in a larger setting of dimeric assemblies covering diverse functional classes. Future work will additionally consider enhancing the sampling capability of the search algorithm with evolutionary search techniques investigated and tested by us in the related problem of tertiary structure prediction in proteins [47,48,51,52]. The model score may also be combined with terms of a physics-based energy function to conduct the search on a combined potentially less rugged surface.…”