基于可见光光谱高效鉴别玉米单倍体籽粒

单倍体技术已发展成为玉米遗传研究及现代玉米育种的重要技术之一，单倍体籽粒的鉴别筛选是其中的重要环节。目前单倍体籽粒主要是依赖于籽粒的R1-nj遗传标记通过人工肉眼观察颜色的有或无进行鉴别，费时费工。而且部分材料由于标记颜色很难从籽粒外部观察到，导致人工筛选准确率较低。基于可见光光谱分析建立玉米单倍体籽粒鉴别方法，探索利用可见光光谱鉴别玉米单倍体籽粒的可行性。同时，由于每季用于诱导单倍体的育种材料不尽相同，模型须能够鉴别未参加建模的材料的单倍体。本研究以9个遗传背景的单倍体和杂交籽粒共284粒作为试验材料，利用便携式紫外-可见光光纤光谱仪采集单个玉米籽粒的可见光漫透射光谱。光谱数据经平滑、矢量归一化预处理和主成分分析，基于支持向量机方法建立单倍体和杂交籽粒判别模型。每次选择1个背景的样本作为测试集，其余背景的样本作为建模集对模型进行交叉验证。模型交叉验证平均正确判别率达到92.06%。其中8次测试正确判别率在85%以上。结果表明利用可见光光谱分析建立玉米单倍体籽粒鉴别方法，并使模型可鉴别未参与建模材料的单倍体具有可行性。并且基于该方法有望建立玉米单倍体籽粒的自动化快速筛选系统，提高玉米单倍体育种效率。

Discrimination of Maize Haploid Seeds from Hybrid Seeds Using Vis Spectroscopy and Support Vector Machine Method

Doubled haploid (DH) lines are routinely applied in the hybrid maize breeding programs of many institutes and companies for their advantages of complete homozygosity and short breeding cycle length. A key issue in this approach is an efficient screening system to identify haploid kernels from the hybrid kernels crossed with the inducer. At present, haploid kernel selection is carried out manually using the“red-crown” kernel trait (the haploid kernel has a non-pigmented embryo and pigmented endosperm) controlled by the R1-nj gene. Manual selection is time-consuming and unreliable. Furthermore, the color of the kernel embryo is concealed by the pericarp. Here, we establish a novel approach for identifying maize haploid kernels based on visible (Vis) spectroscopy and support vector machine (SVM) pattern recognition technology. The diffuse transmittance spectra of individual kernels (141 haploid kernels and 141 hybrid kernels from 9 genotypes) were collected using a portable UV-Vis spectrometer and integrating sphere. The raw spectral data were preprocessed using smoothing and vector normalization methods. The desired feature wavelengths were selected based on the results of the Kolmogorov-Smirnov test. The wavelengths with p values above 0.05 were eliminated because the distributions of absorbance data in these wavelengths show no significant difference between haploid and hybrid kernels. Principal component analysis was then performed to reduce the number of variables. The SVM model was evaluated by 9-fold cross-validation. In each round, samples of one genotype were used as the testing set, while those of other genotypes were used as the training set. The mean rate of correct discrimination was 92.06%. This result demonstrates the feasibility of using Vis spectroscopy to identify haploid maize kernels. The method would help develop a rapid and accurate automated screening-system for haploid kernels.