Multistrategy Learning Model Based on Analogical Reasoning and its Application to Function Prediction of Proteins

在这个项目中，我们在逻辑编程环境中提出了一种新的机器学习技术：类比归纳预测（IPA）。IPA从目标谓词的示例中学习与源谓词相似的目标谓词的描述。IPA的一个关键特征是它使用类比来约束假设空间，使用一阶谓词逻辑表示的分类信息。IPA解决的典型问题是当知识库中没有目标的概念描述时，确定给定的基础原子是否有效。该方法包括以下步骤：（1）识别候选类比源;（2）建立源域和目标域之间的类比映射;（3）对目标谓词实例的映射和推理进行评估;（4）合并类比结果。为了验证该方法的有效性，我们将其应用于蛋白质功能预测知识库的构建。利用氨基酸序列的相似性进行预测的传统技术使我们只能将蛋白质功能分类到功能组中。它们通常不能预测特定的蛋白质功能。为了克服这一局限性，我们利用IPA进行功能特征分析。“功能特征”是指表征具有氨基酸序列的蛋白质的功能的氨基酸序列的特征。它们是对应于包含蛋白质功能的功能元件的序列的二级和/或三级结构。我们已经证明了IPA的有效性，将其应用于分类功能的细菌视紫红质样蛋白质家族。

In this project, we present a novel machine learning technique in a logic programming environment : Inductive Prediction by Analogy (IPA). IPA learns the description of a target predicate similar to a source predicate from examples of the target predicate. A key feature of IPA is that it uses analogies to constrain the hypothesis space using taxonomic information represented by first-order predicate logic.Typical problems addressed by IPA are to decide whether a given ground atom is valid or not, when no concept descriptions for the goal are available in a knowledge base. This is attained by the steps : 1) recognition of a candidate analogous source, 2) elaboration of an analogical mapping between source and target domains, 3) evaluation of mapping and inferences to given examples of the target predicate, and 4) consolidation of the outcome of the analogy.To validate the effective of the proposed method, we apply it to build a knowledge-base for protein function prediction. Conventional techniques for the prediction using similarities of amino acid sequences enable us to only classify the protein functions into function groups. They usually fail to predict specific protein functions. To overcome the limitation, we utilize IPA for functional feature analysis. By "functional feature", we mean a feature of an amino acid sequence characterizing the function of a protein with the amino acid sequence. They are secondary and/or tertiary structures of the sequences that corresponds to functional elements comprising the functions of a protein. We have shown the effectiveness of IPA by applying it to classifying functions of a bacteriorhodopsin-like family of proteins.