AIPPN: Interpretable AI enabled Molecular Identification Pipeline for Plant-parasitic Nematodes

植物寄生线虫(PPN)是温带、亚热带和热带农业系统中许多农作物的破坏性害虫。它们的取食造成的破坏导致全球农作物产量损失14%。此外，PPN可以与真菌和细菌病原体形成破坏性的疾病复合体，并作为病毒载体。关于管理的决定，如杀线虫剂的使用、生物农药的使用、抗性/耐性品种的选择和轮作长度，往往以土壤测试为基础，以确定现有属或种的密度。在土壤采样后，使用分类特征和二分检索表对提取的线虫进行评估。分子检测，如qPCR，也可用于线虫诊断，但主要用于鉴定感兴趣的个别物种。要确定土壤样本中自由生活的PPN的阵列，需要训练有素的分类学家，通常至少有3-4年的经验。不幸的是，这样的专家正在减少，培训人员也很少。囊线虫由经济上重要的球状线虫属和异形线虫属组成，在土豆和甜菜等作物中进行常规评估。在英格兰和威尔士，马铃薯胞囊线虫(Globodera_rostochiens_和_G.pallida_)出现在48%的马铃薯生产用地上，每年造成的损失高达26-5000万GB。目前对马铃薯胞囊线虫的评估主要基于称职分析人员对胞囊形状(雌性线虫)的形态特征。在BBSRC最近的一个项目中，我们通过建立基于CNN的PPN检测模型来识别土壤中重要的PPN属，从而开发了这个问题的早期解决方案(BBSRC项目BB/X01200X/1)。在这里，我们建议通过扩展可解释的人工智能模型来提高PPN识别、分类和量化的准确性和稳健性，以此为基础，通过开发、培训和测试基于可解释的AI的自动图像识别技术作为标准PPN识别系统的替代方案，为这一技术技能差距提供一种变革性的解决方案。具体地说，该技术能够：1)从混合PPN样本中提取形态PPN特征；2)获取PPN特征在PPN识别中的敏感度；3)实现准确的PPN分类和定量。它将建立在该联盟(BB/X01200X/1)现有资源的基础上，包括：2K PPN图像(SparkSoft)、世界领先的害虫检测模型(PestNet；UoS)、农学和害虫管理(ADAS)，本研究的成果将代表着线虫快速筛选的重要进展，并具有许多研究应用。

Plant parasitic nematodes (PPNs) are destructive pests of many crops in temperate, sub-tropical, and tropical agricultural systems. Damage caused by their feeding results in a 14% loss in global crop yield. Furthermore, PPNs, can form damaging disease complexes with fungal and bacterial pathogens and act as virus vectors. Decisions on management, such as nematicide application, biopesticide use, choice of resistant/tolerant varieties and rotation length is frequently based on soil testing to determine the density of genera or species present. Following soil sampling, the extracted nematodes are assessed using taxonomic characteristics and dichotomous keys. Molecular tests such as qPCR may also be used in nematode diagnosis but these are mainly used to identity individual species of interest. To determine the array of free living PPNs in a soil sample, trained taxonomists, typically with at least 3-4 years of experience, are needed. Unfortunately, such specialists are in decline and there are few training providers.Cyst nematodes, consisting of the economically important genera _Globodera_ and _Heterodera_ spp., are routinely assessed in crops such as potatoes and sugar beet. The potato cyst nematodes (_Globodera_ _rostochiensis_ and _G. pallida_) occur in 48% of the land used for potato production in England and Wales and are associated with annual losses amounting to £26-50M. Current assessment of potato cyst nematode is mainly based upon the morphological characterisation of cyst shape (female nematodes) by competent analysts.In a recent BBSRC project we developed an early-stage solution to this problem by building a CNN-based PPN detection model to recognise important PPN genera from soil (BBSRC project BB/X01200X/1). Here, we propose to build on the success of this project by expanding interpretable AI models to improve the accuracy and robustness of PPN identification, classification, and quantification.This project will provide a transformative solution for this technical skills gap by developing, training and testing an interpretable AI based automatic image recognition technique that will act as an alternative to the standard PPN identification system. Particularly, this technique is able to: 1) extract morphological PPN features from mixed PPN samples; 2) access PPN features' sensitivity in PPN recognition; 3) achieve accurate PPN classification and quantification.It will build on existing resources in the consortium (BB/X01200X/1), including: 2K PPN images (SparkSoft), a world-leading pest detection model (PestNet; UoS), agronomic and pest management (ADAS),Outputs of this study would represent an important advancement in the rapid screening of nematodes and have many research applications.