Monitoring the gut microbiome via AI and omics: a new approach to detect infection and AMR and to support novel therapeutics in broiler precision farm

供肉类消费的家禽(肉鸡)的产量在全球范围内正在上升，英国是产量最高的国家之一。英国人均禽肉消费量是猪肉的两倍，几乎是牛肉的三倍，而且还在不断增长。由细菌、病毒和寄生虫引起的家禽地方性疾病不受欢迎，因为它们可能造成相当大的经济损失。为了节省生产，广谱抗生素的使用在任何早期疾病的迹象都很普遍，即使在疾病的源头还没有确定的时候(更不用说细菌的起源了)。使用抗生素的行为增加了病原体产生耐药性的风险(抗菌素耐药性-AMR)，使未来与病原体作斗争变得更加困难。为了减少广谱抗生素的使用，养殖场迫切需要解决方案，以有效地监测牲畜，尽快识别感染和感染源，并实施更有针对性的治疗。该项目旨在开发专门为肉鸡业设计的新监测解决方案。这些解决方案被设计为交钥匙：运营商将定期将在农场内获得的数据上传到基于云的服务，在该服务中将自动评估生产状态。警告和建议将通过智能手机/平板电脑上的应用程序发送回农民，以防检测到感染、合并感染或AMR可能性增加。该项目将涵盖影响英国肉鸡养殖业的细菌、病毒和寄生虫来源的主要病原体，以及该国经常使用的主要抗生素类别的AMR。监测解决方案将如何实现他们的预测，以及我们将如何决定上传哪些数据？该项目的核心是一种由机器学习驱动的数据挖掘方法，该方法最近由申请者完善。该方法允许考虑从农场收集的大量异类信息，包括先前感染/AMR事件的历史数据，并允许开发数学模型，该数学模型基于对收集的信息中的特定模式的观察，估计感染或耐药性表现的可能性。该方法还可以分离出哪些农场变量对每种类型的预测最重要(例如，特定的感染或AMR特征)：这些变量被称为“生物标记物”。最初，我们将考虑许多变量：关于谷仓内温度、湿度、光照和空气成分的传感器数据，来自羽毛、土壤、谷仓地板、水、饲料和操作员靴子的样本的微生物分析。来自肠道微生物组分析的数据具有重要作用，即生活在肉鸡肠道中的微生物物种，其丰度、遗传特征和代谢功能已被证明与感染和抵抗力的许多方面有关。病毒和寄生虫的共存将被考虑在内。多亏了机器学习，第一次有可能剔除如此众多的变量，分离出最终预测模型中使用的最相关的(生物标记物)。这些模型将被转变为远程运行的软件应用程序，作为云服务。用户(农民)将根据需要定期上传信息(生物标记值)，使模型能够在任何时候准确地复制真实生产的状态(模型将成为“数字双胞胎”，成为真实系统的虚拟复制品)。农民随后将通过基于网络的应用程序接收消息，报告警告、警报或建议的治疗方法。确定重要变量和开发预测模型的方法在试点研究中取得了成功，从而识别了出版物中记载的有前途的生物标记物。预计该项目对肉鸡养殖监测的影响将是史无前例的。

The production of poultry for meat consumption (broilers) is rising globally, the UK being amongst the countries with the highest production. Poultry meat consumption pro capita in the UK is twice more than pork and almost three times more than beef, and growing. Poultry endemic diseases due to bacteria, viruses and parasites are frowned upon, as they can cause considerable economic losses. To save production, the use of broad-spectrum antibiotics at any sign of incipient disease is widespread, even when the source of the disease has not been pinpointed yet (let alone the bacterial origin). The act of administering antibiotics increases the risk of the pathogen developing resistance (antimicrobial resistance - AMR), making it more difficult to fight that pathogen in the future. To reduce the use of broad-spectrum antibiotics, solutions are urgently needed for farms to efficiently monitor livestock, identify infections and the source of infection as soon as possible, and administer more targeted therapeutics.The project aims at developing new surveillance solutions specifically designed for use by the broiler industry. These solutions are designed to be turn-key: operators will periodically upload data acquired within the farm to a cloud-based service where the state of production will be assessed automatically. Warnings and advice will be sent back to the farmers via apps on smartphones/tablets, in case infections, co-infection or increased likelihood of AMR are detected. The project will cover the main pathogens of bacterial, viral and parasitic origin affecting UK broiler farming, as well as AMR to the main classes of antibiotics routinely administered in the country.How will surveillance solutions achieve their predictions, and how will we decide what data to upload? At the core of the project there is a data mining method powered by machine learning, recently perfected by the applicants. The method allows to consider a large amount of heterogeneous information collected from the farm, including historical data of previous infections/AMR events, and allows the development of mathematical models that, based on observing specific patterns in the collected information, estimate the likelihood of infection or resistance manifestations. The method also allows to isolate what farm variables are the most important for each type of prediction (e.g. a specific infection, or AMR trait): these variables are called "biomarkers". Initially, we will consider many variables: sensor data on temperature, humidity, illumination and air composition in the barn, microbiological analysis of samples from feathers, soil, barn floors, water, feed, and operator boots. An important role is reserved to data originating from the analysis of the gut microbiome, i.e. the microbial species living in the broiler gut, whose abundances, genetic traits and metabolic functions, have been proven implicated in numerous aspects of infection and resistance. Co-presence of viruses and parasites will be considered. Thanks to machine learning, for the first time it will be possible to prune such a multitude of variables, isolating the most relevant (biomarkers) to be used in the final prediction models. These models will be turned into software applications running remotely as cloud services. Users (farmers) will periodically upload information (biomarker values) as required, allowing for the models to replicate exactly at any time the state of the real production (models will become "digital twins", being virtual replicas of the real system). Farmers will then receive messages via web-based apps, reporting warnings, alarms, or suggested therapies. The methods for identifying the important variables and developing prediction models have been successful in pilot studies, leading to the identification of promising biomarkers documented in publications. The projected impact of the project on surveillance in broiler farming is expected to be unprecedented.