Evaluating Protein Structure Prediction Models with Evolutionary Algorithms

EAs are competent at solving complex, multimodal optimization problems in applications with large and badly understood search spaces. EAs are therefore among the most promising algorithms for solving the protein structure prediction problem. In this chapter, we use this insight to evaluate, and show the limitations of, simplified models for protein structure prediction. These simplified models, e.g., lattice-based models, have been proposed for their computational efficiency, and it has been proposed that simplified models will work if only a sufficiently competent optimization algorithm is developed. However, in this chapter we show that simplified models do not contain the biological information necessary to solve the protein structure prediction problem. This is demonstrated in two steps: first, we show that the EA finds the correct structure given a fitness function based on information of the known structure. This shows that the EA is sufficiently competent for accurate protein structure prediction. Second, we show that the same algorithm fails to find correct structures when any of the simplified models is used. Our main contribution is to have strengthened the hypothesis that solving the problem of protein structure prediction will require detailed models encoding information at the atomic level. We have also demonstrated that EAs indeed are promising algorithms for eventually solving the protein structure prediction problem.