Machine Learning Algorithms for Actionable Knowledge Discovery in Synthetic Biology

合成生物学应用工程原理来设计自然界中不存在的生物系统，以便在给定的生物体中实现所需的特性。这种方法对社会有很大的价值，因为它可以用来生产高价值的材料，如精细化学品，药品，生物修复，生物燃料等，然而，无法预测的行为，生物系统在很大程度上阻碍了生物工程应用的进展。虽然领域知识无法预测基因型变化对表型的影响，但机器学习技术的发展和组学技术产生的大量数据使这成为可能。因此，该项目设想了创新的计算方法，以发现可用于合成生物学实验和工业开发的可操作知识。其好处是双重的：（1）从组学信息中推导出有意义的生物学发现。(2)新的机器学习模型能够从高通量数据集中提取高层次的信息。更具体地说，主要的生物学任务是识别特定生物状态的生物标志物，通常与风险相关，并构建生物网络，其节点代表基因，蛋白质，代谢物和边缘指示复杂的关系，可以是功能或调节。例如，该项目的第一部分是研究细菌如何调整其transciptomics以适应不同的环境刺激，细菌通常被用作合成生物学中设计遗传电路的生物体。这是非常重要的，因为细菌在不同的条件下生长时几乎总是经历各种各样的压力，这可能会影响它们自身的生长以及所需的特性。据我们所知，以前的研究还没有描述过生物体在各种条件下可能表现出的不同生物状态的遗传变化，也没有探索过缓解压力的补偿性遗传回路。这项研究将学习在各种条件下生长的细菌的表型景观，识别响应一般应激条件的基因（即生物标志物），通过观察这些生物标志物的基因表达（即基因指纹）来预测细胞状态，并最终回答细菌如何调整其转录组学以适应生物网络形式的不同条件的问题。本文的工作可以扩展到任何类似的问题，用于不同的目的，而同一套例程可以复制。数据挖掘技术的自动管道将被设计为从高噪声，高维组学数据中提取所需的信息。作为一个完全数据驱动的研究，以补充详细的机械理解领域知识，统计测试和无监督学习算法，如差分表达式分析，降维和聚类方法将首先应用于有效地处理数据维度和提取有趣的数据模式，在此基础上进行监督学习。生物标志物识别将通过设计嵌入分类器的特征选择方法来实现，该方法对小样本数据具有鲁棒性。虽然生物网络可以更灵活，但研究最广泛的是关联网络，其中实体仅以某种方式在功能上连接。我们的目标是通过利用机器学习模型的结构和各种表示来描述生物过程，最好是以概率的方式，超越纯粹的关联。

Synthetic biology applies engineering principles to design biological systems that do not exist in the natural world so as to achieve desired properties within a given organism. This approach is of great value to society since it can be used to produce high-value materials, such as fine chemicals, pharmaceuticals, bio-remediation, bio-fuels, etc. However, the inability to predict the behaviour of biological systems largely hinders progress in bioengineering applications. While domain knowledge fails to predict the effect of genotypes changes on phenotype, the development of machine learning techniques and tremendous amounts of data generated by omics technologies have made this possible. This project thus envisions innovative computational methods to discover actionable knowledge that can be fed into synthetic biology experiments and exploit in industry. The benefits are two-folds: (1) Meaningful biological findings deduced from omics information. (2) Novel machine learning model capable of extracting high-level information from high throughput dataset.More specifically, the main biological tasks are identifying biomarkers for a particular biological state, typically related to the risk, and constructing the biological network whose nodes representing gene, proteins, metabolites and edges indicating complex relations which can be functional or regulatory. For example, the first part of this project is to look at how bacteria, which are often used as the organism to design genetic circuits in synthetic biology, adjust their transciptomics to adapt to different environmental stimuli. This is very important as bacteria almost always experience a diverse range of stresses while growing in different conditions which may affect their own growth as well as the desired properties. To our knowledge no previous research has characterised genetic changes underpinning different biological states an organism may exhibit in various conditions, nor compensatory genetic circuit to relieve the stresses has been explored. This research will learn the phenotypical landscape of bacteria growing in various conditions, identify the genes responding to general stress conditions (i.e. the biomarkers), predict the cell state by looking at the gene expressions of these biomarkers (i.e. the gene fingerprint) and ultimately answer the question of how bacteria adjust their transcriptomics to adapt to different conditions in the form of a biological network. This work can be extended to any similar questions for different purpose while the same set of routines may be replicated.An automatic pipeline of data mining techniques will be designed to extract desired information from heavy noise, high dimension omics data. As a totally data-driven research to complement with detailed mechanistic understanding in domain knowledge, statistical tests and unsupervised learning algorithms such as differential expression analysis, dimension reduction and clustering methods will first be applied to effectively tackling the data dimensionality and extract interesting data patterns, based on which supervised learnings are followed. Biomarker identification will be achieved by devising feature selection methods embedded with classier that are robust to small sample size data. While biological networks can be much more flexible, the most widely studied ones are association networks where entities are only known to be functionally connected in some way. We aim to go beyond the mere association to causation by exploiting the structure and various representations of machine learning models being used to describe the biological processes, preferably in a probabilistic way.