Learning and assessing the robustness of Bayesian networks for biological data

贝叶斯网络模型是一种日益流行的生物数据分析方法。在生物学应用中，贝叶斯网络可以看作是一台由输入产生输出的机器。这里的输入是数据、超参数和分布，输出是生物学上有意义的结果，比如变量表达的可能性及其与表型的联系。虽然研究贝叶斯网络的构建已经为目标基因发现和药物开发带来了许多见解，但这些模型通常是计算昂贵的。在贝叶斯网络应用于生物数据时，计算成本造成了三个主要问题。首先，贝叶斯网络的输出对输入变化的敏感性很少被研究。其次，生物学家通常期望更快的分析方法，这限制了贝叶斯网络在第一手数据中的应用。最后但并非最不重要的是，在优化这些模型的特征以获得更有生物学意义的结果方面存在差距。优化需要在输入空间中重复实验和设计。因此，需要一种更快的方法来实现贝叶斯网络的结果。我们建议开发一个鲁棒的贝叶斯仿真器来有效地模拟贝叶斯网络模型的结果，从而降低贝叶斯网络的计算成本。我们还将探索新的生物学参考来验证我们模型的可信度。因此，我们可以在不同的输入上进行实验设计，并优化贝叶斯网络的性能。改进后的方法在一般生物分析实践中具有一定的应用潜力。

Bayesian Network models are an increasingly popular way to analyse biological data. In Biology application, Bayesian Networks could be viewed as a machine which produces outputs from inputs. Here the inputs are data, hyper-parameters and distribution, and the outputs are biological meaningful results like the possibility of variable expression and its links to phenotype. While studying the construction of Bayesian Networks has led to many insights for target gene discovery and drug development, such models are typically computing-expensive. The computation cost caused three main problems in Bayesian Network's application to biological data. Firstly, the sensitivity of Bayesian Networks' outputs to the changes in inputs is rarely accessed. Secondly, biologists normally expect a faster analysis method, which limits the application of Bayesian Network to first-hand data. Last but not least, there is a gap in optimising such models' feature to achieve a more biologically meaningful result. Optimisation requires experiment repetition and design in the inputs space. Therefore, a faster method to achieve Bayesian Networks' results is demanded. We propose to develop a robust Bayesian emulator to mimic the result from the Bayesian Network model efficiently, which could reduce the computation cost of Bayesian Networks. We will also explore novel biological references to validate the credibility of our model. Thus we could do experiment design over different inputs and optimise the performance of Bayesian Networks. The improved method would have potential in general biology analysis practice.