Ecology or genetics? Adapting machine learning approaches to understand determinants of cross-species transmission and virulence in RNA viruses

来自动物的新发传染病继续威胁着人类健康，最近埃博拉病毒、寨卡病毒和中东呼吸综合征冠状病毒爆发的传播和严重疾病就是例证。世卫组织指出，一种新出现的病原体极有可能引发公共卫生危机，并将其命名为“X病”。“X病”最有可能是由RNA病毒引起的，因为它们进化得更快，比其他病原体更有可能出现并感染人类。人畜共患病毒（即从非人类动物跨物种传播给人类的病毒）也具有较高的出现风险。虽然有些人畜共患病毒在感染人类后会引起严重和危及生命的疾病，但其他病毒似乎会引起轻微疾病或根本不会引起疾病。为了对“X疾病”的公共卫生影响做出早期预测，必须确定哪些因素导致“毒性”的这种变化（即疾病结果有多严重）。然而，我们目前对跨物种传播中驱动感染和毒力风险的因素了解甚少，部分原因是缺乏可用的风险因素信息。传统的方法是使用经典的统计模型来识别生态风险因素，尽管这些模型往往过于简化，无法捕捉到出现背后复杂的进化模式。此外，现代RNA测序的便利性已经导致病毒的大型遗传数据资源的更广泛的可用性。遗传模式或“基序”在整个病毒序列中反复出现，某些基序在某些宿主的感染中更频繁地出现，这可能有助于病毒复制或逃避免疫系统。因此，基因序列对于预测跨物种传播后新宿主的感染或毒力可能具有重要的信号。然而，由于序列数据中存在大量潜在信息，因此寻找捕获预测建模基序的实用方法具有挑战性。这项研究的中心目标是将生态学和遗传学结合起来，以改进对哪些动物病毒对人类的出现和严重疾病构成最大风险的预测。为了释放RNA病毒基因序列中的这种潜在力量，需要新的分析方法。我将运用机器学习作为最先进的建模方法。机器学习模型可以根据大量高度多样化的预测因子和复杂的相互作用来预测结果。这些模型将使我能够确定影响跨物种传播的关键遗传基序，并直接比较遗传和生态特征。为了提高预测性能，我将比较一系列机器学习算法（例如，分类和回归树，支持向量机）和方法(例如：“bagging”，将多个单独的模型集合在一起；“增强”，允许模型逐渐学习)。这项研究将利用利物浦大学开发的增强传染病数据库（EID）中异常广泛的数据，确定所有已知哺乳动物和鸟类RNA病毒的模式。EID2包含从基因记录（GenBank）和科学文献（PubMed）自动收集的29,500对宿主-病原体的感染数据。尽管毒力是病毒的一个关键特征，但没有描述疾病结果的可比资源。在EID2平台上进行扩展，我将开发自动文本挖掘工具，从描述实验性感染的科学文本中捕获不同宿主的疾病结果数据。这项研究将在多种RNA病毒中测试进化理论。提议的机器学习模型将通过提高我们预测突发事件的能力，并为预防未来跨物种传播的疾病爆发提出战略目标病毒或宿主，为公共卫生风险评估提供信息。

Emerging infectious diseases from animal sources continue to threaten human health, exemplified by the spread and severe disease of recent Ebola virus, Zika virus and MERS coronavirus outbreaks. The WHO has noted the serious possibility of a new emerging pathogen to cause a public health crises, denoting this 'Disease X'.'Disease X' is most likely to be caused by an RNA virus, as they evolve faster and are more likely to emerge and infect humans than other pathogens. Zoonotic viruses (i.e. those that transmit cross-species from non-human animals to humans) are also known to have higher emergence risks.Although some zoonotic viruses cause severe and life-threatening illness upon infecting humans, others appear to cause mild disease or no disease at all. To produce early predictions of the public health impacts of 'Disease X', it is essential to identify which factors drive this variation in 'virulence' (i.e. how severe disease outcomes are). However, we currently only have a poor understanding of which factors drive infection and virulence risk in cross-species transmissions, partly because of the lack of available risk factor information. The traditional approach is to identify ecological risk factors using classical statistical models, though these models are often too reductionist to capture the complex evolutionary patterns behind emergence.Additionally, the ease of modern RNA sequencing has led to a much wider availability of large genetic data resources for viruses. Genetic patterns or 'motifs' recur throughout virus sequences, with certain motifs recurring more often within infections of certain hosts, which may aid virus replication or evasion of the immune system. Genetic sequences could therefore hold important signals towards predicting infection or virulence within a new host after cross-species transmission. However, finding practical ways of capturing motifs for predictive modelling has proven challenging due to the large volumes of potential information within sequence data. The central goal of this research is to combine both ecology and genetics to improve predictions of which animal viruses pose the greatest risks of emergence and severe disease in humans. To unlock this potential power in RNA virus genetic sequences, new analytical approaches are needed. I will apply machine learning as a state-of-the-art modelling method. Machine learning models can predict outcomes based on large sets of highly diverse predictors and complex interactions. These models will allow me to identify key genetic motifs influencing cross-species transmission and directly compare genetic and ecological traits. To improve predictive performance, I will compare a range of machine learning algorithms (e.g., classification and regression trees, support vector machines) and approaches (e.g. 'bagging', aggregation over many individual models; and 'boosting', allowing models to gradually learn).This research will identify patterns across all known mammal and bird RNA viruses by using the exceptional breadth of data within the Enhanced Infectious Disease Database (EID), developed at the University of Liverpool. EID2 contains infection data from 29,500 host-pathogen pairs, automatically collected from genetic records (GenBank) and scientific literature texts (PubMed). Despite virulence being a key virus trait, no comparably-sized resources exist describing disease outcomes. Extending on the EID2 platform, I will develop automated text mining tools to capture data on disease outcomes in different hosts from scientific texts describing experimental infections.This research will test evolutionary theory across a large diversity of RNA viruses. The proposed machine learning models will inform public health risk assessment by improving our capacity to predict emergence and suggesting strategic target viruses or hosts for preventing future disease outbreaks from cross-species transmission.