Machine-learning to predict and understand the zoonotic threat of E. coli O157 isolates

肠出血性大肠杆菌（EHEC） O157是一种主要储存在食品生产动物（主要是牛）中的细菌，可导致人类严重和危及生命的感染。确定肠出血性大肠杆菌O157的具体因素包括微注射（3型分泌）系统和特定志贺毒素的产生。然而，近二十年来我们已经知道，并不是所有的亚型对人类健康都构成同样的威胁，我们已经付出了巨大的努力来理解为什么会这样。一个关键原因是有不同的志贺毒素类型，有些可能比其他的毒性更大，它们的产生水平在不同的分离株之间不同。这种可变性来自这样一个事实，即志贺毒素是通过被称为噬菌体的细菌病毒感染而被引入细菌的。这些噬菌体以“噬菌体”的状态将它们的DNA整合到细菌基因组中。当细菌细胞受到威胁时，它可以激活噬菌体，产生自身的副本和新的噬菌体。从大肠杆菌的全基因组测序中，我们现在知道大肠杆菌基因组中存在多个噬菌体，其中一些处于不同的衰变状态，但它们可以相互影响并重新组合以产生新的变体。大肠杆菌O157分离株之间的差异很大程度上取决于它们的噬菌体含量，但序列鉴定方法通常只使用“核心”基因进行流行病学研究。我们最近应用机器学习方法来检测来自牛和人类的大肠杆菌O157的全基因组序列。我们使用这些作为训练集，然后要求它预测其他大肠杆菌O157分离株应该分配到哪一组。令人惊讶的是，它只将一小部分（<10%）来自牛的分离株分配给人类群体，这表明只有这一小部分可能对人类健康构成更大的威胁。这项拨款是为了调查这一选择过程的生物学基础。我们知道，机器学习分配是基于预测主要由前噬菌体基因表达的歧视性蛋白质变体，因此这符合我们对这些分离株中存在的变异的理解。拟议的工作将是生物信息学研究和“湿”感染生物学研究的结合。在生物信息学方面，我们可以使用主观和客观的方法在分离序列之间交换基因变异，包括整个噬菌体，并重新计算它们的宿主预测分数。这将使我们能够确定用于预测人畜共患病可能性的最重要的基因组合。它还可以突出特定的基因，以简化识别过程。在实验室中，我们将首先比较在核心基因组水平上非常相似但在预测分数上显着不同的分离株。我们将在小鼠模型中检查它们的基因表达谱，代谢谱和关键表型，如志贺毒素产生，细胞相互作用和病理。然后我们将交换或突变生物信息学鉴定的基因，并在相同的实验室分析中测试这些菌株变异。这项研究应该有助于验证这种令人兴奋的新方法，以了解细菌的毒力，并确定与这种危险病原体的人畜共患威胁有关的基因。然后，我们应该能够开发更简单的方法，在农场上识别这些特定的变异，并进行干预，例如疫苗，以减少对人类健康的威胁。该方法还可能用于预测人类分离株之间的毒力差异，这可能对如何管理特定的疫情产生影响。这项研究是及时的，因为它建立在我们最近和独特的机器学习应用的基础上，以预测人畜共患的可能性，并获得与James Bono博士（美国农业部，内布拉斯加州）合作产生的英国牛和人类大肠杆菌O157分离物的完全注释PacBio序列。

Enterohemorrhagic Escherichia coli (EHEC) O157 are bacteria that have their main reservoir in food production animals, predominately cattle, and can be responsible for serious and life-threatening infections in humans. There are specific factors that define EHEC O157, including a micro-injection (type 3 secretion) system and production of specific Shiga toxins. However, we have known for nearly twenty years that not all subtypes represent the same threat to human health and significant effort has gone into understanding why this is the case. On key reason is that there are different Shiga toxin types, some potentially more toxic than others, and their production levels differ between isolates. This variability comes from the fact that Shiga toxins are introduced into the bacteria by infection with bacterial viruses, known as bacteriophages. These integrate their DNA into the bacterial genome in a 'prophage' state. When the bacterial cell is threatened this can activate the prophage to produce copies of itself and new bacteriophages. From whole genome sequencing of E. coli we are now aware that multiple prophages are present in E. coli genomes, some in different states of decay, but they can impact on each other and recombine to produce new variants. Much of the differences between E. coli O157 isolates are down to their prophage content yet sequence identification methods generally use only 'core' genes for epidemiological studies. We have recently applied machine-learning approaches to examine whole genome sequences of E. coli O157 from cattle and humans. We use these as training sets and then ask it to predict which group other E. coli O157 isolates should be assigned to. Surprisingly it only assigns a small proportion (<10%) of isolates from cattle to the human grouping, indicating that only this small subset may be more of a threat to human health. This grant is to investigate the biological basis of this selection process. We know that the machine-learning assignment is based on discriminatory protein variants predicted to be expressed from mainly prophage genes, so this fits with our understanding of the variation present in these isolates. The proposed work will be a combination of bioinformatics research and 'wet' infection biology research. For the bioinformatics we can use subjective and objective approaches to swap gene variants, including whole prophage, between isolate sequences and re-calculate their host prediction scores. This will allow us to define the most important combinations of genes being used for the prediction of zoonotic potential. It may also highlight specific genes to simplify the identification process. In the laboratory we will initially compare isolates that are very similar at the core genome level but differ markedly in their prediction scores. We will examine their gene expression profiles, metabolic profiles and key phenotypes such as Shiga toxin production, cellular interactions and pathology in a mouse model. Then we will swap or mutate genes identified by the bioinformatics and test these strain variants in the same laboratory assays. The research should help validate this exciting new approach to understanding bacterial virulence and identify genes involved in the zoonotic threat of this dangerous pathogen. We should then be able to develop simpler approaches to identifying these specific variants on farms and intervene with, for example a vaccine, to reduce the threat to human health. The approach may also work to predict differences in virulence between human isolates and this could have repercussions for how specific outbreaks are managed. This research is timely as it builds on our recent and unique application of machine learning to predict zoonotic potential and access to fully annotated PacBio sequences of UK cattle and human E. coli O157 isolates generated in partnership with Dr James Bono (USDA, Nebraska).