Genomic selection, prediction models, GEBV values, genomic selection in plant breeding

Editor's Notes

• #2: Good afternoon everyone welcome u all to my 1st doctoral seminar on genomic selection for crop improvement
• #3: My flow of seminar goes like this; At first I am introducing about the concept of GS Then the process of Genomic selection; it includes training population, its genotyping and phenotyping, estimation of GEBV’s afterwards I will explain the Insights into the GS. Then I will tell about the applications of GS i.e., where and when we need to apply the GS in the plant breeding scenario. And I will cover about the research studies on GS and Finally I will conclude my seminar.
• #4: Coming to the introduction. We the plant breeders our major goal is to breed for novel traits and genotypes. We all know Plant breeding is an art and science of improving the genetic make up of crop plants. During crop breeding we need to carry out different activities. i.e. we need to create the variability its by natural or by artificial means. Naturally through domestication i.e. bringing wild species under human management. Germplasm collection from different countries or locations and introduction of cultivar from a new area where it is not grown earlier. Through artificially we will hybridize between the plants, we will do mutation and polyploidy and we will induce variation in clonally propagated crops means it is somaclonal variation and if we use recombinant DNA technology means it is genetic engineering. After creating the variability we need to select the right variability which we need for the improvement of particular trait. Here selection is a key step. It plays a crucial role for the plant breeders. There are two types of selection 1. Natural selection i.e selection by the nature here nature selects based on the survival of the fittest principle. It was proposed by Charles Darwin the other one is artificial selection here selection is by human i.e we select based on our experience and phenotypic observations of a particular trait. Over the time Hazel and lush given a concept called selection index in this linear combination of characters associated with a particular trait we need to select. It is more reliable compared to single trait selection and increases aggregate genetic gain. But mere selecting based on phenotype is not precise it may mislead. There may be a chances of selection of not desirable traits or individuals. So, markers were developed. Conventional selection is based on phenotype so it is called phenotypic selection, here the environment having drastic impacts Breeders choose good offspring using their experience and the observed phenotypes of crops, so as to achieve genetic improvement of the target traits. There he considering one trait at a time. So, in 1942 Hazel and Lush proposed the selection index method, which uses a total score to select for multiple traits simultaneously. It improve the aggregate genetic gain. With the development of computer science, genetic evaluation methods for analysis of multiple traits combinely also developed.
• #5: In 1990’s markers were came to the rescue of plant breeders. As we all known that these molecular markers are the land marks on the chromosome which is use to track the dynamic trait of our interest. MM are not Crop stage specific They have simple inheritance and They are environmentally neutral so they are very effective for selection instead based only on the phenotype Markers are surrogates for the trait of our interest. If we select any genotype based on the markers data its called as MAS And this MAS is suitable for traits controlled by small number of major genes. But most economic traits of crops are complex and affected by a large number of genes, which is having small effect. SO MAS has also some limitations i.e. it is effective for only major gene/QTL Success achieved with only qualitative traits MAS does not identify minor QTL effects But, most traits are quantitative in nature and contain both large and small effect QTLs So, MAS for QTs and small effect QTL has resulted in less genetic gain So, there is a need for other method which overcomes all these limitations The method which rectifies all the limitations of MAS is Genomic selection
• #6: GS is a specialized form of MAS, The concept was introduced by Haley and Visscher at 6th world congress on Genetics Applied to Livestock production at Armidale, Australia in the year 1998. The term GS was coined by Meuwissen in the year 2001 in the seminal paper entitled Prediction of total genetic value using genome-wide dense marker maps. Published in Genetics journal. Seminar paper: is a work of original research that presents a specific thesis and is presented to a group of interested people, usually in an academic setting
• #7: GS is specialized form of MAS because In conventional MAS we will indirectly select the desired allele based on the molecular markers linked to the trait interest. In GS, all the markers or dense markers covering the entire genome is used for the selection, weather they are significant or not
• #8: This slide shows the schematic representation of the Genomic selection process The GS consists of two types of population i.e., Training population and Breeding population The training population is genotyped and phenotyped to 'train' the genomic selection (GS) prediction model. Genotyping is done with a large number of markers evenly distributed over entire genome The phenotyping should be accurate, replicated and multi-location data should be there. Based on the Genotypic and phenotypic data the GS model was trained. Coming to the Breeding population; it is only genotyped based on the same markers of the training population but no phenotypic evaluation is done. The genotypic values of the BP was put on the GS model to estimate the GEBV’s for individuals or lines from the marker data Based on the GEB Values we will select the individuals In GS main role of phenotyping is to calculate effect of markers & cross validation.  
• #9: This slide indicates the difference between the GS and MAS. Both consists of Training phase and the breeding phase. In the training phase of GS phenotyping and genotyping was done to train the GS model. In MAS training phase mapping population is phenotyped and genotyped to identify QTLs. In the breeding phase of GS. Based in genotyping and GS model the GEBVs were calculated. The individuals with highest GEBVs were selected in the GS. In MAS the plants which were having QTLs are selected.
• #10: GS is a specialized form of MAS, in which data on marker alleles covering the entire genome forms the basis of selection Information from genotypic data on all markers covering the entire genome form the basis of selection Irrespective of weather the effects are significant or not, covering the entire genome are estimated These markers effects are used to estimate the GEBV’s
• #11: The basic steps involved in GS are Development of training population with complete genotypic and phenotypic data Statistical model development Estimation of Genomic Estimated Breeding Values of new breeding lines with genotypic data Cross validation Selection of individuals
• #12: Training population: TP is the population with both phenotypic and genotypic data It must be representative of the BP because it should maximize the proportion of trait variance with the markers. It should be achieved by including the lines with divergent GEBV’s: Eg: HF TP did not produce accurate GEBV in a jersy popl. Larger the training population size increases the accuracy of GEBV estimates It may be Germpalsm lines, biparental derived population i.e., F2, RIL or DH’s. Which ever the lines or population may be used as TP but make in to a note that it should capture maximum allelic diversity of a trait under study. For example, Heffner et al. (2011) reported that the average ratio of GS accuracy to PS accuracy for grain quality traits in biparental wheat populations containing were 0·66, 0·54 and 0·42 for training population sizes of 96, 48 and 24, r espectively. higher training :breeding population ratio is required with greater genetic diversity, smaller-sized breeding populations, lower heritability of traits and larger numbers of existing QTLs to obtain GEBVs with high accuracy. Exhibit low collinearity between markers
• #13: TP can be genotyped by Single nucleotide Polymorphism, Diversity Array Technology, Simple Sequence Repeats, Restriction site Associated DNA and Genotyping by sequencing makers are widely used. Because they can scan genome wide polymorphisms Codominant markers should be used because it distinguishes the polymorphisms RAD (Restriction site Associated DNA) SNPs are RAD and GBS marker systems that can scan GW polymorphisms in de novo would bypass the need for prior marker development and rather allow direct genotyping of the training and breeding populations. thereby also maximizing the number of QTL whose effects will be captured by markers.
• #14: Breeding population: BP is the population with only genotypic data Genotyping done for the same markers as in the training population Ideally the breeding population should be derived from the parental lines that are present in the training population.
• #15: These are the models to predict the GEBVs The GEBVs are calculated based on the sum of the effects of markers across the genome Based on all the markers covering the entire genome we will calculate the GEBVs: these GEBVs forms the selection criterion in the GS. There are different models of GEBV prediction: They are divided into Shrinkage models: Makes the data smaller in size. They are SR, RR-BLUP and G-BLUP 2. Dimension reduction methods: Reduce the no. of random variables under study, which makes analyzing data much easier and faster. They are Partial least square regression, Principal component regression and LASSO-Least absolute shrinkage and selection operator- CAPTURE SMALL NO OF QTL WITH LARGER EFFECTS. 3. Variable selection models: Selection of subset of relevant features, i.e., variables or predictors for use in model construction: they are Bays A & B, Bays Cpi and Bays Dpi. 4. Kernel regression: It is a non parametric technique in statistics used to estimate the conditional expectation of a random variable. Objective is to find out a non-linear relation b/w a pair of random variable X and Y. 5. Machine learning methods: Construction of algorithms; that can learn from and make predictions. They are SVM and RF: Support vector machine regression and Random forest. which shrinks the variance towards zero. The models which shrinks the variance towards zero are: SR, RR-BLUP and G-BLUP …and in the day of high-density markers, this means we probably have many more markers than observations, resulting in the well-known large p, small n problem. This means ordintary least squares cannot be used for estimation, but a variety of other more sophisticated models can be used. The most population is RR-BLUP, where markers are treated as random effects to be sampled from a common distribution. That’s all I’ll say about that.
• #16: Most widely used statistical models are SR – Stepwise regression, RR-BLUP – Ridge regression Best Linear unbiased predictors and Bayesian regression. Genomic Prediction: basic idea Choice of statistical methods for estimating marker Effects also can affect model accuracy. A variety of methods for genomic prediction is currently available. For brevity, we highlight three statistical methods available to train the GS model: ridge regression best linear unbiased prediction (RR-BLUP), Bayes-A, and Bayes-B.
• #17: Select most significant markers on the basis of arbitrary significant thresholds and non significant markers effect equals to zero. Estimate the effect of significant markers using multiple regression and only a portion of the genetic variance will be captured.
• #19: Marker variance treated more realistically by assuming specified prior distribution. assume a prior mass at zero, thereby allowing for markers with no effects. Some marker effects can be = 0 Least demanding in terms of computation
• #20: GS uses all markers as predictors to achieve assessment and selection in early generations, it reduce the time cost per cycle and shorten the generation interval. Population design plays a vital role for GS Germplasm pool can be sampled as the training set for GS, but it may limit scope of prediction In hybrid breeding breeders evaluate an inbred line not by its phenotype but by its potential to create superior hybrids, thus for selection of desirable hybrids needs field trail. GS facilitate hybrid breeding by obtaining better hybrids with fewer crosses.
• #21: Xing wang suggested to look 2 aspects before implementing the GS. 1. is population design and 2. model design. Population design should be like that as I told earlier While designing the model we need to select a model based on the population of selection. If we are using a model which is having dominance and additive effect we need to select the model which explains those effects effectively And we need to use the model which incorporate the Multi-trait and multi-environment data in to it. All this accelerate the breeding and increase the genetic gain.
• #22: These are the GS methods for GEBV estimation into Parametric and Nonparametric methods. In parametric methods there RR-BLUP, BayesA, BayesB, BayesC, LASSO, Bayesian LASSO, GBLUP, Elastic net. The non parametric methods are SVM (Vector method), RKHS(Reproducing Kernel Hilbert Spaces), Random Forest, RBFNN.
• #23: Where to apply GS in the breeding cycle(which generations) and how many lines to select for genotyping. Testing and development of models to implement GS in existing breeding schemes. In this case training and breeding population will be used in the same generations
• #25: To assess the results of statistical model It is a natural way of assessing model performance from breeder’s perspective a training set and validation set, models are fitted in the training set, fitted models are used to predict outcomes in the validation set or to verify the accuracy of GEBV’s 2. Verification of accuracy of GEBV’s on selection candidates : We will be having predicted GEBV’s based on Genomic selection model, simultaneously we will be conventionally calculating breeding values. Both will be compared by Pearson's correlation.
• #28: An exotic inbred can be crossed with an adapted inbred and the F2 and subsequent generations from this cross can be handled by the two-step procedure described above. Incase it is desired to continue GS beyond 7-8 cycles. The plants in cycle 7 or 8 should be genotyped as well as evaluated for testcross performance, marker effects should be estimated afresh, and the new estimates should be used for further GS.
• #31: The time efficiency over PS could come from the second cycle of selection, which uses the TP from the previous cycle to predict the new DH lines, thus excluding testcross formation and first-stage multi-location evaluation trials. Reducing cost upto50% Saving time by selecting lines directly for stage II instead going for stage I (used in PS) This significantly reduces the cost of testcross formation and evaluation at each stage of multi-location evaluations.
• #33: This is the research study conducted for GS of seed weight based on low-density SCAR markers in soybean Material consists of 288 soybean varieties : TP: N=238; VP: N=50 For genotyping 79 SCAR markers were used.
• #34: For GS prerequisite is TP and BP. TP is genotyped and phenotyped Genotyping done by using 79 SCAR markers in 288 soybean varieties Phenotyping data were collected by the CGRIS – Chinese crop germplasm information system From both this train the model using RR-BLUP and BLR – Bayesian linear regression BP was genotyped and the data fed to the GS training model to estimate the GEBVs Once GEBVs estimated, it was compared to evaluate the predictive effects
• #35: This slide indicates the genotyping results of soybean individuals based on SCAR markers.
• #36: This indicates the correlation coefficient of GEBVs and TBVs for Hundred seed weights of soybean The prediction and true value is 0.9042 That is indicated in the small dots
• #37: GEBVs were highly correlated with true breeding values by using SCAR markers The maximum relationship values were 0.854 and 0.904 Results indicated that HSW was controlled by many small-effect genes, which was more suited to GS than MAS Therefore GS would be suitable for estimation of crop breeding traits in soybean
• #40: Vivek and his coworkers published a paper in the year 2017 in the article The Plant Genome entitled Use of Genomic Estimated Breeding Values Results in Rapid Genetic Gains for Drought Tolerance in Maize. The objective of this study was to report the genetic gains made through GS and compare breeders practice of developing source i.e training populations through S1 testcrosses and subsequent per se selections with that of GS. Material: For this study they used a two bi-parental maize populations referred to as CAP i.e CIMMYT Asia populations. One population is CML470 and CML444 the other population is VL1012767 and CML444 here parent 2 i.e CML444 is common in the table 1 they mentioned the size of F2.3 population CAP1 is 276 with 342 polymorphic SNPs and CAP2 is 178 with 377 polymorphic SNPs.
• #41: This slide indicates the GS flowchart of work flow of genomic selection procedures used in development of various improved cycles of selection. After polymorphism screening the parents were crossed and developed the F1s and those F1s were selfed to form F2’s and those F2’s were again selfed to produce F2.3 populations of each cross of CAP1 and CAP2. Then the F2.3 families of each population were crossed to tester CML474. Testcrosses were divided into several trails based on seed availability and were phenotyped under drought and optimal or well watered environments. CAP1 were evaluated in ICRISAT, Sabour and Ludhiana and the CAP2 were evaluated in ICRISAT, Sabour, Belgaum and Davanagere. Alpha lattice design with two to three replications per location Drought stress trial conducted during dry season. Drought stress at flowering to mid-grain fill was imposed on the crop by withdrawing subsequent irrigations. Recommended POP was followed in the specific location as recommended by the respective state department of Agriculture. 4m row 0.75m spacing between rows and 0.2 m spacing between plants were maintained and a plant population of 63,636 plants per hectare was maintained. Grain yield was recorded by adjusting the moisture to 12.5 percent. Best Linaera Unbiased Predictors were calculated for each entry to assess the performances across sites and also BLUPs were calculated for plant stand adjusted GY for both the drought and well watered conditions and BLUPs for ASI was calculated at drought situation/condition. Genotyping was done for the polymorphic markers in the F2.3 families of both CAP1 and CAP2 populations using SNP markers. Formation of Cycle1: A selection index was calculated based on BLUPs from test cross data and weighted as 35% grain yield under drought, 25% ASI under drought and 40% Grain yield under optimal or well watered conditions. For the formation of CYCLE1 top 10% of the F3 families are selected based on these selection index and then recombined. Marker effects were calculated by correlating the testcross phenotypic performance with genotypic data of respective F2.3 families using R software. Formation of Cycle 2: 350 seeds of C1 were planted and they were genotyped. Based on genotyping data GEBVs were calculated. A larger GEBV indicates a favorable plant. 24-30 Plants with top GEBVs were identified and were recombined. Such population is termed as C2(TC-GS) because C2 generated by GS using marker effects generated from testcross data. C2(PerSe-PS): C2 generated by perse phenotype. C2 plants grown under optimal conditions visually appealing good (stainability/no lodging, vigor, general plant aspect including ear position and disease) and further visually appealing cobs were selected based on ear rots, texture, color, general ear aspects, grain fill and cob size were used for C2 perse-PS. Superior families identified under drought are recombined under optimal conditions.
• #42: Based on average grain yield across the drought locations and two populations, per se performance of C2(PerSe-PS) Representing the phenotypic selection ranged from 32 to 39%. While that of C2(TC-GS) representing GS ranged from 53 to 59%. The per se performance of C2(TC-GS) was 10 to 20% better than C2(Perse-PS)
• #45: Zhang et al in the year 2017 published a paper in G3 i.e Genes Genomes and Genetics entitled Rapid Cyclig Genomic Selection in a Multiparental Tropical Maize Population. The objective of there study was to report the genetic gains in four cycles say C1, C2, C3 and C4 and in Original population C0 in multi-environmental field trails of Rapid Cycling Genomic Selection assisted breeding with four checks in two Mexican environments or locations.
• #46: The rapid cycling GS experiment was started in 2009. By using 18 CIMMYT Tropical Maize Inbred lines. The list of inbred lines is as follows.
• #47: The steps or methods followed in Rapid Cycling GS is presented in the figure 1. The selected CIMMYT Tropical maize inbred lines were used as parents and crossed between them in a half diallele fashion in the year 2010B they were belong to flint type kernel heterotic group. In 2011A season they intermated the F1s to form S1s there were 4800 individuals. Among them they selected 1000 best ears were selected and planted in ear to row in 2012A season. They were testcrossed with a tester CML495 from the complementary dent heterotic group. These are the training population (C0) for developing genomic prediction models. The genotyping was done using the SNP markers by GBS platform. The phenotyping was done on the same population and >10 agronomic characters were recorded. The best 50 families with best plant type, flowering and maturity were selected and planted ear to row, 25 plants per family. Cycle 1 was formed by intermating 50 selected families. Based on visual evaluation of flowering time, plant type, plant/ear height, well-filled ears and reaction to naturally occurring diseases along with among and within family selection 157 ears were harvested and shelled individually to form C1. In C1 DNA was extracted and genotyping was done and calculated the GEBVs. Based on highest GEBValues top 25 families were selected and intermated to form C2 population. Based on visual evaluation of flowering time, plant type, plant height, well filled ears and reaction to naturally occurring diseases within family selection was implemented. A total of 91 ears were harvested and individually to form C2. DNA was isolated and genotyping was done using the SNP markers. GEBVs caluculated. Based on highest GEBValues top 22 families were selected. They were intermated to form C4 cycle/population. Total 45 cobs were harvested. And these C4 were testcrossed using testers to evaluate the genetic gains across cycles in 2 locations. individuals the top The produced F1 is intermated between them they were self pollinated. From those populations testcrosses were made with a tester CML495/CML549. They evaluated the test crosses in 4 optimal locations with location specific checks. From the each cycle i.e in C0 they selected 50 families, in C1 they selected 25 families, in C2 they selected 18 families and in C3 they selected 22 families
• #48: This table indicates the number of families and the individual plants sown, individual plants selected and the plants advanced in each breeding cycle and among family, within family and the total selection intensity. In the different cycles I.e in CO 1000 families sown
• #49: Coming to the results: Table 2 indicates the mean grain yield tons per hectare for each genomic cycle ie. C0, C1, C2, C3 and C4, Broad sense heritability and mean of four testers to 2 locations and combined across two locations. The average genetic gain in Grain yield across cycles was estimated for each location and across locations including the only genomic selection cycles (C1 – C4) and the all selection cycles (C0-C4). They observed the genetic gain in C4 cycle is highest compared to the other cycles in each location and the combined locations.
• #51: This table indicates the means of entry and checks for traits anthesis days, silking dates, plant height, ear height and moisture content in each cycle and across the two locations. On an average the anthesis and silking days did not increase with respect to average of GS cycles C0, C1, C2, C3 and C4. They ranged from 56 days for anthesis and 57 days for silking and showd good synchrony between the flowering times. However, GS produced taller plants and ear insertions during cycles C3 and C4 compared to C1 and C2. Grain moisture content did not seem to have affected after the three cycles of Rapid cycle GS.
• #52: The genetic diversity was well controlled upto C3 then it is declined.
• #53: These are the some examples of GS work recently taken place in different crops. In maize for different traits say for grain yield, anthesis date, ASI, plant height under normal and water stress conditions and for diseases like Northern corn leaf blight and I am working on GS for Fusarium stalk rot in Maize.
• #54: These are the projects on GS in and around the world In tomato GS for quality, shape and shelf life by SNP markers In Barley GS for FHB resistance by using SNP markers In Trifolium GS for yield by using SNP markers In Wheat GS by sequencing In Maize GS for drought using SNP markers In Maize GS for total biomass yield and silage quality by SNP markers and Sugarbeet GS for white sugar yield and sugar content by SNP markers These projects were funded by different organizations for the improvement of the desired characters.
• #56: Thank you one and all for patience listening