In this paper, we present a knowledge-based memetic algorithm to tackle the three-dimensional protein structure prediction problem without the explicit use of experimentally determined protein structures' templates. Our algorithm proposal was divided into two main prediction steps: (i) solutions sampling and initialization; and (ii) structural models’ optimization coming from the previous stage. The first step generates and classifies several structural models for a given target protein, through the Angle Probability List strategy, to identify distinct structural patterns and consider reasonable solutions in the memetic algorithm initialization. The Angle Probability List takes advantage of structural knowledge stored in the Protein Data Bank to reduce the size and, consequently, the conformational search space complexity. The second step of the method consists in the optimization of the structures generated in the first stage by the proposed memetic algorithm. It uses a tree-based population where each node can be seen as an independent subpopulation that interacts with each other over global search operations, aiming at knowledge sharing, population diversity, and better exploration of the multimodal search space. The method also encompasses ad hoc global search operators, whose objective is to increase the method exploration ability focusing on specific characteristics of the protein structure prediction problem, combined with the artificial bee colony algorithm used as an exploitation technique applied to each node of the tree. The proposed algorithm was tested on a set of 24 amino acid sequences, as well as compared to the reference method in the protein structure prediction area, the method of Rosetta. The obtained results show the ability of our method to predict three-dimensional protein structures with similar folding to the experimentally determined ones, regarding the structural metrics Root-Mean-Square Deviation and Global Distance Total Score Test. We also show that our method was able to reach comparable results to Rosetta, and in some cases, it outperformed Rosetta, corroborating the effectiveness of our proposal.

在本文中，我们提出了一种基于知识的模因算法，无需明确使用实验确定的蛋白质结构模板即可解决三维蛋白质结构预测问题。我们的算法建议分为两个主要的预测步骤：（i）解决方案采样和初始化；和（ii）结构模型的优化来自上一阶段。第一步，通过角度概率列表策略为给定的目标蛋白质生成并分类几个结构模型，以识别不同的结构模式并考虑模因算法初始化中的合理解决方案。角度概率列表利用蛋白质数据库中存储的结构知识来减小大小，从而减小构象搜索空间的复杂性。该方法的第二步在于通过拟议的模因算法优化第一阶段生成的结构。它使用基于树的种群，每个节点都可以看作是一个独立的子种群，可以通过全球搜索操作相互互动，以实现知识共享，种群多样性，并更好地探索多模式搜索空间。该方法还包括专门的全局搜索运算符，其目标是结合针对用作树的每个节点的开发技术的人工蜂群算法来提高针对蛋白质结构预测问题的特定特征的方法探索能力。该算法在24个氨基酸序列上进行了测试，并与蛋白质结构预测领域的参考方法Rosetta方法进行了比较。获得的结果表明，关于结构度量根均方差和整体距离总分测试，我们的方法能够预测具有与实验确定的折叠相似的折叠的三维蛋白质结构。