Feng Shou, Fu Ping, Zheng Wenbin
Department of Automatic Test and Control, Harbin Institute of Technology, Harbin 150080, China.
Rev Sci Instrum. 2018 Mar;89(3):034302. doi: 10.1063/1.5010353.
Predicting gene function based on biological instrumental data is a complicated and challenging hierarchical multi-label classification (HMC) problem. When using local approach methods to solve this problem, a preliminary results processing method is usually needed. This paper proposed a novel preliminary results processing method called the nodes interaction method. The nodes interaction method revises the preliminary results and guarantees that the predictions are consistent with the hierarchy constraint. This method exploits the label dependency and considers the hierarchical interaction between nodes when making decisions based on the Bayesian network in its first phase. In the second phase, this method further adjusts the results according to the hierarchy constraint. Implementing the nodes interaction method in the HMC framework also enhances the HMC performance for solving the gene function prediction problem based on the Gene Ontology (GO), the hierarchy of which is a directed acyclic graph that is more difficult to tackle. The experimental results validate the promising performance of the proposed method compared to state-of-the-art methods on eight benchmark yeast data sets annotated by the GO.
基于生物仪器数据预测基因功能是一个复杂且具有挑战性的层次多标签分类（HMC）问题。当使用局部方法来解决这个问题时，通常需要一种初步结果处理方法。本文提出了一种名为节点交互方法的新型初步结果处理方法。节点交互方法对初步结果进行修正，并确保预测结果符合层次约束。该方法利用标签依赖性，并在第一阶段基于贝叶斯网络进行决策时考虑节点之间的层次交互。在第二阶段，该方法根据层次约束进一步调整结果。在HMC框架中实现节点交互方法，也提高了基于基因本体（GO）解决基因功能预测问题的HMC性能，其中GO的层次结构是一个更难处理的有向无环图。实验结果验证了与在由GO注释的八个基准酵母数据集上的现有方法相比，所提出方法具有良好的性能。
