GrainQuest - using Artifical Intelligence and high resolution multimodal imaging to dissect the developmental and genetic basis of seed composition

组成（蛋白质，脂质，碳水化合物等）决定了质量和最终用途，因此这些性状在育种中很重要。因此，大麦和小麦的蛋白质含量很重要。高油燕麦可能会导致新的食品产品，如非乳制品酸奶。然而，籽粒产生于双受精事件，产生了一个复合结构，具有三个遗传上不同的组织或室-母体果皮，胚胎和胚乳。它们的相对大小和组成对整体颗粒组成有不同的贡献。我们开发了最先进的成像方法（Aber的微ct扫描和多/高光谱成像，QUB的激光辅助快速蒸发电离质谱（LA-REIMS）），提供不同颗粒组织的空间分辨信息。我们的大麦研究表明，增大的胚胎可以解释高蛋白系（Cook et al, 2017），但这需要艰苦的人工分析。为了消除这一瓶颈，与计算机科学公司合作开发了人工智能例程，以快速提取和测量可以揭示成分变化发展基础的特征。了解组成变异的遗传和发育“成分”不仅在智力上很重要，而且因为它可以加速育种进程。在这些技术进步的基础上，该项目将使用IBERS实验育种种群项目中的谷物，这些项目已经使用传统方法测量了全谷物成分。这些广泛和可变的数据集的可用性将使学生能够专注于中心假设，即不同基因型的不同颗粒室的相对大小和组成确实不同，影响整体组成。具体目标是：Obj1。机器学习工具为了开发强大的深度学习协议，将使用一组特征测试颗粒来训练神经网络，以识别和量化可能与成分变化相关的特征。我们最近使用uCT扫描来测量谷物的花器官体积（Adamski et al 2021），最近使用深度学习（DL）来完成人类几乎不可能完成的任务（即测量胚胎-胚乳相对体积）。DL将被扩展到量化更广泛的性状，包括作为生化成分代表的高光谱指数（Feng等人，2017），同时保留穗（小麦）或穗（燕麦）内的位置信息。obj2。为了验证人工智能的输出，必须在不同的隔间中直接测量成分。因此，完整的颗粒将被CT扫描并使用高光谱相机成像，然后转移到QUB进行LA-REIMS，这是一种新兴的技术，用于内源性复杂脂质和代谢物的空间分析，具有高水平的化学鉴定。LA-REIMS结果将在空间上与其他成像方式共同注册，直接测试预测的成分。为了正式解决这些新性状的变异是否与遗传变异有关（因此对育种者有用），学生将使用来自遗传非结构化（绘图）群体的谷物，这些谷物以前是通过杂交选择具有独特谷物组成的燕麦对而产生的。学生将建立定量模型，以解释基因型对产量和组成质量变化的贡献，并与使用传统全谷物数据产生的模型进行比较。Obj3将在一个有意义的遗传背景下对这一假设进行严格的检验。

Composition (protein, lipid, carbohydrate, etc) determines quality and end-use so these traits are important in breeding. Thus, protein content is important for barley and wheat. High oil oat lines may lead to novel food products such as non-dairy yogurts. However, grains arise from a double fertilization event that produces a composite structure with three genetically distinct tissues or compartments - the maternal pericarp, the embryo and the endosperm. Their relative sizes and compositions contribute different constituents to overall grain composition.We have developed state-of-the-art imaging methods (microCT scanning and multi/hyperspectral imaging at Aber and Laser-assisted rapid evaporative ionization mass spectrometry (LA-REIMS) at QUB) that provide spatially resolved information on the different grain tissues. Our barley study demonstrated that enlarged embryos could explain high protein lines (Cook et al 2017) but this required painstaking manual analyses. To remove this bottleneck, a collaboration with Computer Science developed AI routines to quickly extract and measure features that can reveal the developmental basis of compositional variation. Understanding the genetic and developmental "components" of compositional variation is important, not only intellectually, but because it can accelerate breeding progress.Building on these technical advances, this project will use grain from IBERS' experimental breeding populations projects, where whole-grain composition has already been measured using traditional methods. The availability of these extensive and variable datasets will enable the student to focus on the central hypothesis that the relative sizes and compositions of different grain compartments do vary between genotypes, affecting overall composition. The specific objectives are:Obj1. Machine learning toolsTo develop robust deep learning protocols, a set of characterized test grains will be used to train neural networks to recognize and quantify features that may be related to compositional variation. We have recently used uCT scanning to measure floral organ volume in cereals (Adamski et al 2021) and recently used deep learning (DL) to undertake tasks that humans find almost impossible (i.e. measuring embryo-endosperm relative volumes). DL will be extended to quantify a wider variety of traits including hyperspectral indicies as proxies for biochemical composition (Feng et al., 2017), while retaining positional information within the spike (wheat) or panicle (oats).Obj2. Co-registration across imaging modalities To validate the outputs from AI, composition must be directly measured in the different compartments. Therefore, intact grains will be CT scanned and imaged using a hyperspectral camera before transferring to QUB for LA-REIMS, an emerging technique for spatial analysis of endogenous complex lipids and metabolites at a high level of chemical identification. LA-REIMS results will be spatially co-registered with the other imaging modalities, directly testing the predicted composition.Obj3. Verification of concept To formally address whether variation in these novel traits is associated with genetic variation (and therefore be useful to breeders), the student will use grain from genetically unstructured (mapping) populations that have been previously created by crossing selected pairs of oats with distinctive grain composition. The student will build quantitative models to account for the genotypic contribution to variation in features contributing to yield and compositional quality and compare to models produced using conventional whole grain data.Obj3 will provide a rigorous test of the hypothesis in a meaningful genetic context.