| Doubled haploid technology has been widely applied in maize breeding.Large scale haploid production was guaranteed by effective haploid induction and rapidly accurate haploid identification.Haploid induction rate not only controlled by pollen inducer,but also affected with mother material.In this study,we performed experiments on maternal effects of haploid induction,paternal effects of qhir8,the relationship of genetic distance with HIR and discussed the application of machine learning on haploid identification,haploid male fertility prediction and DH lines selection.These results would make a foundation to the engineering of doubled haploid technology.The main results of our study are as follows:1.According to the screen of maternal haploid inducibility among 20 inbreds,we finally found Qi319 and Chang7-2 with quite difference.The two were then crossed and induced to generate F1 haploids,after chromosome doubling we got 135 doubled haploid lines.The DH population were then planted across Beijing,Jinan,Shijiazhuang and Sanya,and induced by high oil inducer CHOI3.After harvest,haploids in each ear of the DH lines were carefully identified and HIR were calculated.Combined with linkage genetic map constructed by 6K SNP,we performed QTL mapping of maternal haploid inducibility.One major QTL(qMHI9-3)explained up to 22.14%phenotype variation was detected in Sanya on chromosome9,and could also be detected in Shunyi and Jinan with different LOD peak.Another QTL(qMHI10-3)on chromsome10 detected in SJZ with PVE of 17.44%,could also be found in Jinan.These two major QTLs proved to be candidate for fine mapping.2.A series of chromosome segment substitution lines of was constructed in B73 background to analysis the effects of qhir8 and interaction with qhirl.We found that CSSL with only qhir8 locus has no haploid induction rate and embryo abortion rate.But it can improve the HIR effectively when accompanied with qhirl.The correlationship between HIR and EmbR in qhir8 locus was extremely low,but in qhirl showed a significantly high coefficient of 0.443.3.The correlationship of parental interaction and haploid induction rate was studied by genetic distance analysis.There is no certain relation between genetic distance and HIR when whole genome SNP applied to analysis.However,an obvious negative correlation was found when the genotype narrowed to the maternal inducibility QTL.The coefficient in inducer CHOI3 and CAU5 was-0.53 and-0.51 separately,which reached a significant level of 0.01.4.Including DH lines,DH cross with Mo 17,inbred cross and 5 hybrids were used in NIR machine learning to identify haploids from crossed seeds after haploid induction.Different models were built based on the algorithm of naive bayes,decision tree,k nearest neighbours,random forest,support vector machine,partial least square regression and neural network.After comparison,partial least square regression and neural network model were proved to be highly effective,whose accuracy were 93.26%and 95.42 separately.In addition,dataset of crossed seeds and haploids which can not be discriminated were used in machine learning,and the accuracy was 93.39%.Machine learning had an advantage of haploid identification independent of R1-nj with high accuracy.5.The application of machine learning in spontaneous haploid doubling prediction and DH line selection was discussed.We investigate haploid male fertility with 0-5 levels,and the machine learning predicted accuracy was only 50.8%.When divided these haploid into two levels,fertile and sterile,the accuracy could be improved to 61.06%.It would be possible to use this method in spontaneous haploid doubling in future.Seed weight of DH lines were used for model construction,a partial least square regression method was finally used and the predicted seed weight was highly correlated with true value,the correlation coefficient was 0.859.It was promising to gain insights in DH selection from yield realted traints machine learning. |
PDF Full Text Request