Machine learning for tracing pathogens in the food chain

肠道沙门氏菌是全球人类胃肠炎的主要原因，据估计，非伤寒沙门氏菌每年造成约10亿人感染和约15万人死亡。因此，这种胃肠道病原体是一个重大的公共卫生问题，需要进行实时流行病学监测和后续行动。然而，疫情调查经常被国际食品贸易网络的复杂性所混淆，这些网络在全球范围内传播人畜共患病的食源性病原体。该项目旨在通过利用机器学习(ML)直接从基因组监测数据预测胃肠道病原体的地理来源来解决这一差距，从而改进公共卫生反应并更快地解决疫情。公共卫生机构，如美国疾病控制中心(CDC)、英国健康安全局(UKHSA)和加拿大公共卫生局(PHAC)，在收集相关元数据的同时，通常会对临床确诊的沙门氏菌病例应用全基因组测序(WGS)。WGS包含关于地理来源的背景遗传信息，由于它们存在于由多种来源的成分组成的复杂食品中，因此与食源性病原体特别相关。然而，从WGS推断地理来源的传统方法需要广泛的专业知识和高昂的计算成本，而且不能有效地进行扩展。我们最近确定ML是一种有效的工具来确定肠炎沙门氏菌(PMID：37042517)的地理来源，肠炎沙门氏菌是英国食源性疾病的最主要原因。该项目将建立在这一方法学基础上，与来自英国卫生与公众服务局、PHAC和美国疾病控制与预防中心的公共卫生专家合作，为排名前三位的沙门氏菌物种合成国际基因组数据库。结果将产生快速和准确的地理来源归因模型，适合立即与公共卫生机构整合，以加强现有的疾病管理对策。这将通过三个主要目标来实现：a)共同产生有效的源头归属所需的知识。为了向流行病学家提供强有力的决策支持工具，自动预测应既准确又可被最终用户理解。学生将定期与UKHSA、CDC和PHAC的专家利益相关者会面，并确保目标(B)和(C)中制定的预测性ML框架包含与流行病学后续行动和最终用户解释相关的信息。B)研究沙门氏菌基因组中的系统地理信号学生将整理由英国卫生与公众服务局、疾病控制与预防中心和PHAC向项目提供的肠炎沙门氏菌、鼠伤寒沙门氏菌和Newport的基因组监测数据集，并将进行系统地理信号分析(即基因组数据按地理来源聚集的程度)。这将提供对沙门氏菌物种的可行见解，允许标记高度适合ML分类的区域限制克隆以及需要(C)中改进分类方法的有问题的国际克隆。C)使用(B)中收集的样本优化沙门氏菌物种的来源归因模型，学生将使用分层ML和深度学习框架建立来源归因模型，以预测暴发的地理来源。这些模型将使用最先进的可解释ML方法来促进快速、有针对性的疫情应对。“

"Salmonella enterica is a leading cause of human gastroenteritis worldwide, with non-typhoidal Salmonella estimated to account for ~1 billion infections and ~150,000 deaths annually. This gastrointestinal pathogen therefore represents a major public health concern, necessitating real-time epidemiological monitoring and follow-up. Outbreak investigations, however, are often confounded by the complexity of international food-trade networks which distribute zoonotic food-borne pathogens across the globe. This project aims to address this gap by utilising machine learning (ML) to predict the geographical source of gastrointestinal pathogens directly from genomic surveillance data, allowing for improved public health response and more rapid outbreak resolution. Public health agencies, such as the US Centers for Disease Control (CDC), UK Health Security Agency (UKHSA) and Public Health Agency of Canada (PHAC), routinely apply whole genome sequencing (WGS) to clinically identified cases of Salmonella alongside collecting relevant metadata. WGS contains contextual genetic information on geographical origin, of particular relevance to foodborne pathogens due to their presence in complex foodstuffs consisting of ingredients from multiple sources. However, traditional methods for inferring geographical origin from WGS demand extensive expertise and high computational costs while not scaling effectively. We recently established that ML is an effective tool for the geographical source attribution of Salmonella Enteritidis (PMID: 37042517), the most prominent cause of foodborne illness in the UK. This studentship will build upon this methodological foundation, working alongside public health specialists from the UKHSA, PHAC and CDC to synthesis international genomic datasets for the top three Salmonella species. Outcomes will result in rapid and accurate geographical source attribution models suitable for immediate integration with public health agencies to enhance existing disease management responses. This will be achieved via three primary objectives: a) Co-produce knowledge required for effective source attribution In order to deliver a robust decision-support tool for epidemiologists, automated predictions should be both accurate and understandable by end-users. The student will regularly meet with expert stakeholders in UKHSA, CDC and PHAC and ensure the predictive ML frameworks developed in objectives (b) and (c) contain information relevant to epidemiological follow-up and end-user interpretation. b) Investigate the phylogeographical signal in Salmonella genomes The student will collate genomic surveillance datasets of Salmonella enterica serovars Enteritidis, Typhimurium and Newport provided to the project by the UKHSA, CDC and PHAC, and will perform an analysis of phylogeographical signal (i.e. how clustered the genomic data are by geographical origin). This will provide actionable insights into Salmonella species, allowing for the flagging of regionally restricted clones highly-suitable for ML classification as well as problematic international clones which require enhanced classification methods in (c). c) Optimise source attribution models for Salmonella species Using the samples collected in (b), the student will build source attribution models using hierarchical ML and deep-learning frameworks to predict the geographical sources of outbreaks. These models will use state-of-the-art explainable ML approaches to facilitate rapid, targeted outbreak responses. "