Utility of multi-omics data to inform genomic prediction of heifer fertility traits

Abstract

Biologically informed single nucleotide polymorphisms (SNPs) impact genomic prediction accuracy of the target traits. Our previous genomics, proteomics, and transcriptomics work identified candidate genes related to puberty and fertility in Brahman heifers. We aimed to test this biological information for capturing heritability and predicting heifer fertility traits in another breed i.e., Tropical Composite. The SNP from the identified genes including 10 kilobases (kb) region on either side were selected as biologically informed SNP set. The SNP from the rest of the Bos taurus genes including 10-kb region on either side were selected as biologically uninformed SNP set. Bovine high-density (HD) complete SNP set (628,323 SNP) was used as a control. Two populations—Tropical Composites (N = 1331) and Brahman (N = 2310)—had records for three traits: pregnancy after first mating season (PREG1, binary), first conception score (FCS, score 1 to 3), and rebreeding score (REB, score 1 to 3.5). Using the best linear unbiased prediction method, effectiveness of each SNP set to predict the traits was tested in two scenarios: a 5-fold cross-validation within Tropical Composites using biological information from Brahman studies, and application of prediction equations from one breed to the other. The accuracy of prediction was calculated as the correlation between genomic estimated breeding values and adjusted phenotypes. Results show that biologically informed SNP set estimated heritabilities not significantly better than the control HD complete SNP set in Tropical Composites; however, it captured all the observed genetic variance in PREG1 and FCS when modeled together with the biologically uninformed SNP set. In 5-fold cross-validation within Tropical Composites, the biologically informed SNP set performed marginally better (statistically insignificant) in terms of prediction accuracies (PREG1: 0.20, FCS: 0.13, and REB: 0.12) as compared to HD complete SNP set (PREG1: 0.17, FCS: 0.10, and REB: 0.11), and biologically uninformed SNP set (PREG1: 0.16, FCS: 0.10, and REB: 0.11). Across-breed use of prediction equations still remained a challenge: accuracies by all SNP sets dropped to around zero for all traits. The performance of biologically informed SNP was not significantly better than other sets in Tropical Composites. However, results indicate that biological information obtained from Brahman was successful to predict the fertility traits in Tropical Composite population.

Multi-omics data can be used to select single nucleotide polymorphism (SNP) that are biologically relevant. These SNP can be used to aid genomic selection strategies.

Introduction

Early puberty in beef cattle underpins higher first conception rates. First conception and related reproductive traits, which are easy to record, can be useful in beef production systems to perform genetic selection. Adoption of genomic selection strategies has contributed with a 3- to 4-fold increase in the genetic gain for traits with low heritability, such as fertility traits in dairy cattle (Weller et al., 2017). Reproductive traits in beef cattle are often low to moderately heritable (Hawken et al., 2012; Zhang et al., 2014; Ali et al., 2019). The complex and polygenic natures of reproductive traits were reported in cattle and pigs (Fortes et al., 2013; Zak et al., 2017). The complexity of fertility traits emerges from the molecular networking of multiple loci plus the impact of environmental factors (Chen et al., 2008).

Theory shows that the accuracy of genomic prediction is improved by higher trait heritability and larger reference population size (Goddard, 2009). The accuracy of genomic prediction also improves when the reference population consists of a combination of different breeds or populations (Brondum et al., 2011; de Haas et al., 2012; Porto-Neto et al., 2015; Song et al., 2019). It can be deduced that genomic prediction is challenging for fertility traits with low heritability but can be improved by increasing the size of the reference population and combining different populations.

Another aspect to increase the accuracy of genomic prediction is the use of customized sets of single nucleotide polymorphism (SNP; Xiang et al., 2021). The commonly used genomic best linear unbiased prediction (GBLUP) method assumes that all loci have small effects on traits under investigation. While assuming this, GBLUP ignores the fact that biological information retained by specific regions of the genome governs the expression of the phenotypes (Abdollahi-Arpanahi et al., 2017). Multiple studies have demonstrated that the use of informative variants, in linkage disequilibrium (LD) with causal mutations, and ignoring non-informative variants may improve the accuracy of genomic predictions (Sarup et al., 2016; van den Berg et al., 2016; Raymond et al., 2018; Xiang et al., 2019). In addition to genomic prediction, the trait-specific informative SNP also impact heritability estimates (Sarup et al., 2016; Tahir et al., 2022). Various SNP selection strategies have been proposed to improve the accuracy of genomic predictions, such as selecting SNP with large effects for specific traits (Weigel et al., 2009), SNP linked to trait-specific gene sets or pathways (Koufariotis et al., 2014; Edwards et al., 2016; Abdollahi-Arpanahi et al., 2017; Rezende et al., 2019), SNP with presumed functional roles (Koufariotis et al., 2014), and selection of SNP based on machine learning approaches (Li et al., 2018; Yin et al., 2020). There is no current consensus on what the best method for customizing SNP sets for fertility traits is.

Our group previously published genome-wide association, transcriptomics, and proteomics studies that identified candidate genes related to puberty and fertility in Brahman heifers (Nguyen, 2018; Tahir et al., 2019, 2021). In the current study, we used the biological information generated by these previous studies to perform genomic predictions for three female reproductive traits: pregnancy status after the first joining (PREG1, binary), first conception score (FCS, score 1 to 3), and rebreeding score (REB, score 1 to 3.5) in Tropical Composite heifer population. The hypothesis was that the use of SNP located in genomic regions which are biologically relevant to heifer puberty and fertility might result in capturing the genetic variation and producing better genomic prediction accuracies for puberty-based fertility traits in across-breed scenario. To test this hypothesis, we compared three SNP sets: a biologically informed SNP set, a biologically uninformed SNP set, and the high-density (HD) 777K SNP set. In doing so, we investigated the usefulness of the biological information obtained from the studies on Brahman breed to capture genetic variation, estimate heritabilities, and predict phenotypes in another breed Tropical Composite. We compared the SNP sets in two scenarios: 1) across breed use of biological information: 5-fold cross-validation within Tropical Composite heifer population using SNP sets that were selected based on Brahman studies; and 2) across breed use of prediction equations: application of prediction equations (SNP effects) from Brahman to Tropical Composites and vice versa to predict GEBV.

Materials and Methods

Prior biological knowledge from previous multi-omics experiments

Previously, our group used genomics, proteomics, and transcriptomics approaches to identify candidate genes related to puberty and fertility in Brahman heifers. The genomics approach applied multitrait meta-analysis on three early reproductive traits PREG1, FCS, and REB to identify candidate genes related to fertility in Brahman female cattle (Tahir et al., 2021). The proteomics approach used mass-spectrometry technique for hypothalamic, pituitary, and ovarian tissues of pre- vs. postpubertal Brahman heifers to identify proteins/genes related to puberty (Nguyen, 2018; Tahir et al., 2019). The transcriptomics approach used hypothalamic, pituitary, and ovarian tissues of pre- vs. postpubertal Brahman heifers to identify genes related to puberty through RNA-sequencing technique (Nguyen, 2018). A comprehensive list of candidate genes (number of genes = 5,306) was identified in these studies. We named this list as multi-omics gene list. In contrast to multi-omics gene list, we created a list of nonmulti-omics gene (number of genes = 21,927) which included all Bos taurus genes except the multi-omics gene list. Based on these gene lists, different SNP sets were preselected to be used in genetic parameter estimation and genomic prediction of the traits selected for this study.

Preselection of SNP sets

Three SNP sets were selected to test their ability to explain genetic variance of the traits used in this study and to predict these traits. The first set comprised of imputed genotypes from bovine HD complete SNP retained after quality control filtering: MAF > 0.05; and call rate > 0.9. This SNP set was termed as “HD complete SNP set”. From the HD complete SNP set, the SNP within the multi-omics genes including 10 kb up and down stream genomic regions were selected as “biologically informed SNP set”. Similarly, from the HD complete SNP set, the SNP within nonmulti-omics genes including 10 kb up and downstream genomic regions were selected as “biologically uninformed SNP set”. The details about these SNP sets are provided in Table 1. The three SNP sets were used to estimate genetic parameters and to perform genomic predictions for three traits PREG1, FCS, and REB.

Table 1.

Description of SNP sets used for genomic prediction of the heifer reproductive traits measured early in life

SNP Set	SNP Markers	Average MAF1	Description
HD complete SNP Set	669,662	0.295	All BovineHD SNP (MAF > 0.05; call rate > 0.9)
Biologically informed SNP set	110,040	0.295	Within 10 kb around multi-omics genes
Biologically uninformed SNP set	228,308	0.294	Within 10 kb around non-multi-omics genes

¹Minor allele frequency.

Genotype and phenotype datasets

Genotype and phenotype records used in this study were obtained from “Female Fertility PhenoBank” project (L.GEN.1710), a research project funded by Meat and Livestock Australia (MLA). This project maintains a database of reproductive phenotypes recorded in female cattle of different tropically adapted beef breeds, typical of northern Australia. We obtained data of two breeds i.e., Brahman and Tropical Composite from Female Fertility PhenoBank. The cattle populations and scoring criteria of the traits used in this study are described in Table 2.

Table 2.

Scoring criteria and number of animals per trait in each population

Trait	Score	Scoring criteria	Tropical composite number of animals			Brahman number of animals
			CRC4	NT5	Total	CRC	KAM6	NT	Total
PREG11	1	Not pregnant after first joining	138	167	305	261	211	128	600
	2	Pregnant after first joining	737	289	1026	671	997	51	1719
		Total	875	456	1331	932	1208	179	2319
FCS2	1	Never conceived up to 36-m age	136	–	136	230	199	–	429
	2	Conceived in 29 to 36-m age	19	–	19	41	395	–	436
	3	Conceived before 29-mo age	847	–	847	658	492	–	1150
		Total	1002	–	1002	929	1086	–	2015
REB3	1	Not pregnant after the first two joining	–	22	22	26	116	11	153
	2	Pregnant in second joining, not in first	133	144	277	235	198	117	550
	3	Pregnant in first joining, not in second	196	109	305	387	103	16	506
	4	Pregnant twice, after first two joining	541	170	711	284	7	53	326
		Total	870	445	1315	932	424	179	1535

¹Pregnancy status after first joining.

²First conception score.

³Rebreeding score.

⁴Beef CRC population.

⁵Kamilaroi population.

⁶NT DITT population.

Tropical composite population

Since the biologically informed and uninformed SNP sets were based on biological information obtained from Brahman cattle, we planned to test the effectiveness of these SNP sets in capturing genetic variance, estimating heritabilities, and predicting phenotypes in an independent population of a different breed i.e., Tropical Composite. This led to a selection of genotype and phenotype records of 1,350 Tropical Composite cows from the Cooperative Research Centre for Beef Genetic Technologies (Beef CRC; Hawken et al., 2012) and the Northern Territory Department of Industry, Tourism, and Trade (NT DITT; Schatz and Hearnden, 2008) herds. The phenotype records for early reproductive traits PREG1 and REB were available from both herds in this population, while the records of FCS were available with the Beef CRC herd only (Table 2).

Brahman population

Records of 2,319 Brahman cows belonging to the Beef CRC, NT DITT herds, and a commercial Brahman breeder herd (Kamilaroi) were available from the Female Fertility PhenoBank. The same population was used in our previous study on genomics approach (described above) to identify genes related to traits PREG1, FCS, and REB. The purpose of selecting this population in this study was to perform across breed validation of SNP effects from one breed to the other for the estimation of GEBV.

Genotypes and imputation

A combined reference panel of 1,180 Brahman and Tropical Composite cattle, genotyped with the Bovine-HD (777K) SNP chip, was used to impute medium density SNP genotypes (Bovine 50K SNP chip) of Tropical Composite population to HD level (777K). Similarly, a reference panel of 546 Brahman animals, genotyped with the Bovine-HD (777K) SNP chip, was used to impute medium density genotypes (Bovine 50K SNP chip) of Brahman population to HD level. A combination of Eagle v2.4.1 (Loh et al., 2016) and Minimac3 (Das et al., 2016) software was used for imputation. After quality control filtering (allelic R² > 0.4, MAF > 0.05), a total of 669,662 SNP for PREG1, 669,793 SNP for REB, and 671,280 SNP for FCS were available in the Tropical Composite heifers. Similarly, in Brahman population, 587,900 SNP for PREG1, 584,510 SNP for FCS, and 584,344 SNP for REB were available after imputation.

Training and validation groups in the tropical composite population

5-fold cross-validation approach was adapted to determine the accuracy of genomic predictions for each trait in the Tropical Composite population. 5-fold cross-validation provides the advantage of large training groups and avoids high variation in accuracies of genomic prediction (Erbe et al., 2013). For this purpose, five training and five validation groups were selected from the Tropical Composite population. The records in the Tropical Composite population were randomized (Silva et al., 2016; Toghiani et al., 2017), and five equal subsets were made. The training and validation groups were made in a way that every training group consisted of four out of five subsets and the fifth subset was declared as a validation group. This way, five training and validation groups (TR1, TR2, TR3, TR4, TR5 and Val1, Val2, Val3, Val4, Val5) were created for each trait. The composition of validation groups showing the representation of scores for all traits is provided in the Supplementary Figure S1. The training and validation groups were used to determine genetic parameters and the accuracy of genomic predictions, respectively.

Fixed effects

The individual datasets that were combined to create Tropical Composite population came from independently raised herds, therefore, each dataset had its own fixed effects. The Beef CRC dataset had fixed effects: farm location (animals raised together on the same farm), cow crop (birth year), and birth month class (August to November = class A; December to April = class B). The NT DITT dataset had the fixed effects: cow crop and birth month class. Similarly, in the Brahman population, dataset from each herd had its own fixed effects. The Beef CRC dataset had farm location, cow crop, and birth month class as fixed effects. The NT DITT dataset had cow crop and birth month class as fixed effects. Kamilaroi had cow crop as fixed effect. In addition to the above-mentioned fixed effects, “datasets” themselves were also used as fixed effects. All fixed effects in each individual dataset were combined to make a combined fixed effect, which was used in the final model.

Another factor to consider was different breeds within the Tropical Composite population. Tropical Composite is a generic term for stable crossbreeds from different breeds in Australia. Animals from the individual herds within the Tropical Composite population had undefined multiple Tropical Composite breeds (Figure 1). To address breed differences, Tropical Composite genotypes were subjected to principal component analysis using software GCTA and the top two principal components were used as covariates in the final model. The variance explained by the first and second principal components was 5% and 2%, respectively.

Figure 1.

Principal component analysis of breeds from different populations used in this study. CRC_BB, Cooperative Research Centre Brahman cows; CRC_TC, Cooperative Research Centre Tropical Composite cows; Kamilaroi_BB, Kamilaroi herd Brahman cows; NT DITT_BB, Northern Territory Department of Industry, Tourism, and Trade Brahman cows; NT DITT_TC, Northern Territory Department of Industry, Tourism, and Trade Tropical Composite cows.

Estimation of genetic parameters and GBLUP analysis

To investigate the first scenario, 5-fold cross-validation in the Tropical Composite population, the training groups were used to estimate additive genomic breeding values and SNP effects using GBLUP methodology implemented in GCTA software (Yang et al., 2011). The model used was as follows (Zhang et al., 2014):

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{y}=\mathrm{X}\beta +\mathrm{Z}\mathrm{g}+\mathrm{e}}\end{array}}$$ (1)

where, X is the incidence matrix of fixed effect, β is the estimate of fixed effects, Z is the incidence matrix of random polygenic effects, g is the estimate of random additive polygenic effects distributed as (g) ∼ N(0, Gσg²), where G is the genomic relationship matrix (GRM) based on set of SNP markers included in the respective SNP sets. Residual is represented as e and distributed as (e) ~ N(0, Iσe²). The same model was used to estimate heritability of the traits using complete records of the Tropical Composite population.

The GBLUP SNP effects estimated from the training groups were used to calculate the GEBV of the validation groups, as follows (Toghiani et al. 2017):

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{G}\mathrm{E}\mathrm{B}{\mathrm{V}}_{\mathrm{i}}={\sum }_{(j=1)}^n{\mathrm{Z}}_{\mathrm{i}\mathrm{j}}{\mathrm{a}}_{\mathrm{j}}}\end{array}}$$ (2)

where, Z is the incidence matrix designating the additive allele score of jth SNP for ith individual, and a is the estimated SNP effect corresponding to the jth SNP.

In addition to model (1), another model including multiple GRM (biologically informed and biologically uninformed) was also implemented. The purpose of this model was to estimate the additive genetic variance explained by biologically informed and biologically uninformed SNP sets, when fitted together in the model. Following was the multi-GRM model used:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{y}=\mathrm{X}\beta +\mathrm{Z}{\mathrm{g}}_{\mathrm{b}\mathrm{i}}+\mathrm{Z}{\mathrm{g}}_{\mathrm{b}\mathrm{u}}+\mathrm{e}}\end{array}}$$

where, X is the incidence matrix for fixed effect, β is the estimate of fixed effects, Z is the incidence matrix for random polygenic effects, g_bi is the estimate of the parameter with distribution (g_bi) ∼ N(0, G_biσg_bi²) that captured the additive genetic variance explained by the biologically informed SNP set, and g_bu is the estimate of the parameter with distribution (g_bu) ∼ N(0, G_buσg_bu²) that captured the additive genetic variance by the biologically uninformed SNP set. G_bi and G_bu are the genomic relationship matrices that were constructed using the biologically informed and uninformed SNP sets, respectively. Residual is represented as e and distributed as (e) ~ N(0, Iσe²).

Across-breed validation of SNP effects

To investigate the second scenario, across breed validation of SNP effect, complete records of both populations were used to estimate SNP effects for all SNP sets and traits using model (1). The SNP effects from one breed were used to predict GEBV of the other breed.

Accuracy of genomic prediction

The accuracy of genomic prediction was calculated as the correlation coefficient between GEBV and phenotypes adjusted for fixed effects. In case of 5-fold cross-validation in Tropical Composite population, the average of accuracies from five validation groups was considered as the accuracy of genomic prediction for the traits.

Comparison with Bovine XT-50 genotype array

A genotyping array “XT-50” was recently reported to have variants pleiotropic for 34 traits of economic importance including fertility in cattle (Xiang et al., 2021). We compared the genomic regions tagged by biologically informed SNP set in our study with the XT-50 genotype array. Since the XT-50 array included sequence variants, we expected limited commonalities between biologically informed SNP set and the XT-50 genotype array. To make a fair comparison, the extended regions of 10 kb around variants of the XT-50 array were used to extract HD SNP from the Bovine-HD SNP set. This approach resulted in a list of HD SNP in vicinity of the XT-50 variants that were comparable to the biologically informed SNP set of our study.

Results

In the Tropical Composite population, using the HD complete SNP set, PREG1 came up with the highest heritability estimate of 0.16 followed by REB with an estimate of 0.11, and FCS with an estimate of 0.08. When comparing the SNP sets within traits, the biologically informed SNP set produced relatively higher estimates of heritability for PREG1 (0.18) and FCS (0.12) than the HD complete SNP set and biologically uninformed SNP set; however, the differences in estimates were statistically insignificant. For REB, the heritability estimate by the HD complete SNP set was the highest as 0.11 (Table 3).

Table 3.

Estimates of heritability and accuracy of genomic prediction for heifer reproductive traits measured early in life in Bos indicus derived cattle in two across breed scenarios

SNP sets	h 2 1	SE2	Accuracy (r)3	Accuracy (r)	Accuracy (r)
	(TC)		5-Fold validation TC4	SNP effects of BB5 on TC	SNP effects of TC on BB
PREG16
Biologically informed SNP set	0.18	0.05	0.20 (0.012)	-0.03	0.002
HD complete SNP set	0.16	0.05	0.17 (0.01)	-0.001	0.02
Biologically uninformed Set	0.13	0.05	0.16 (0.01)	0.01	0.02
FCS7
Biologically informed SNP set	0.12	0.05	0.13 (0.018)	-0.04	−0.03
HD complete SNP set	0.08	0.05	0.10 (0.017)	−0.02	−0.002
Biologically uninformed Set	0.08	0.05	0.10 (0.017)	−0.01	−0.006
REB8
Biologically informed SNP set	0.09	0.04	0.12 (0.018)	−0.04	−0.03
HD complete SNP set	0.11	0.05	0.11 (0.023)	−0.02	−0.02
Biologically uninformed Set	0.10	0.05	0.11 (0.024)	−0.02	−0.03

¹Heritability of traits in Tropical Composite population.

²Standard error of heritability.

³Accuracy of genomic prediction: correlation between GEBV and adjusted phenotypes. Parentheses: standard error of accuracy of genomic prediction.

⁴Tropical Composite population.

⁵Brahman population.

⁶Pregnancy after first joining.

⁷First conception score.

⁸Rebreeding score.

In multiple GRM-based GBLUP analysis, when GRM based on the biologically informed and uninformed SNP sets were used together in the model, the biologically informed SNP set captured all observed heritability for PREG1 and FCS while biologically uninformed SNP set could not capture any proportion of the heritability for these traits. In the case of REB, the heritability was split between biologically informed and uninformed SNP sets with biologically uninformed SNP set explaining more proportion of heritability than the biologically informed SNP set. The detailed estimates of genetic variance and heritability by all SNP sets in multiple GRM GBLUP models for all traits are provided in Table 4.

Table 4.

Heritability and genetic variance explained by biologically informed and uninformed SNP sets in multiple GRM GBLUP model

Trait	MGRM1 used in model	Genetic variance (SE2)	Heritability (SE)
PREG13	Biologically informed SNP set	0.032 (0.01)	0.23 (0.1)
	Biologically uninformed SNP set	0 (0.01)	0 (0.1)
FCS4	Biologically informed SNP set	0.077 (0.05)	0.16 (0.1)
	Biologically uninformed SNP set	0 (0.05)	0 (0.1)
REB5	Biologically informed SNP set	0.011 (0.04)	0.03 (0.09)
	Biologically uninformed SNP set	0.030 (0.04)	0.08 (0.09)

¹Multiple genomic relationship matrix.

²Standard error.

³Pregnancy status after first joining.

⁴First conception score.

⁵Rebreeding score.

Overall, the accuracy of genomic prediction was low for all traits which is typical for low heritable fertility traits. In the first scenario, 5-fold cross-validation within the Tropical Composite population, biologically informed SNP set allowed estimation of relatively more accurate GEBV for all traits when compared to the HD complete SNP set or the biologically uninformed SNP set; however, these differences were nominal and not statistically significant. The accuracy of genomic prediction by biologically informed SNP set was 0.20 for PREG1, 0.13 for FCS, and 0.12 for REB. The HD complete SNP set produced genomic prediction accuracy of 0.17 for PREG1, 0.10 for FCS, and 0.11 for REB. Biologically uninformed SNP set produced almost similar accuracies of genomic prediction for all traits as did the HD complete SNP set. The details of genomic prediction accuracies by all SNP sets for all traits are provided in Table 3 and Figure 2.

Figure 2.

Accuracies of genomic prediction in Tropical Composite heifer population through 5-fold cross validation. BIN: Biologically informed SNP set included SNP within 10 kb region of the genes identified in puberty and fertility related studies in Brahman cattle through multi-omics techniques. HDC: HD complete SNP set included all SNP from BovineHD SNP chip after quality control filtering. BUN: Biologically uninformed SNP set included SNP within 10 kb of all Bos taurus genes except the genes used in selecting biologically informed SNP set. PREG1, pregnancy after first joining, FCS, first conception score, REB, rebreeding score.

In the second scenario, the application of SNP effects from one breed to the other, the accuracies of genomic prediction for all traits were drastically reduced for all SNP sets when the GEBV of a population were calculated using SNP effects from the other population. All SNP sets resulted in the accuracy of genomic prediction for all traits around zero (Table 3).

In order to compare the genomic regions tagged by biologically informed SNP set and XT-50 genotyping array, we identified 184K HD SNP in extended genomic regions (10 kb) around the variants of XT-50 genotype array. The comparison of these SNP with biologically informed SNP set of the current study (110K) revealed 100K SNP in common.

Discussion

The purpose of this study was to investigate the effectiveness of fertility-related biological information, discovered in the Brahman breed, to estimate genetic parameters of early fertility traits and predict these in another breed i.e., Tropical Composite. We compared a biologically informed SNP set to the HD complete SNP set and to a biologically uninformed SNP set as means to incorporate (or not) biological information from previous studies (Nguyen et al., 2017; Nguyen, 2018; Tahir et al., 2019, 2021) into genomic predictions for heifer fertility traits measured early in life. The idea was to test the effectiveness of the biological information with the challenge of two scenarios: 1) determining the effectiveness of biologically informed SNP set (obtained from Brahman breed) for genetic parameter estimation and genomic prediction of puberty-based reproductive traits in another breed i.e., Tropical Composite through 5-fold cross-validation, and 2) determining the effectiveness of biologically informed SNP set for genomic prediction through application of SNP effects from one breed on other.

Overall, the accuracies of genomic prediction for all traits by all SNP sets were low in this study. Accuracy of genomic prediction is dependent on the heritability of the trait (Hayes and Goddard, 2010; Zhang et al., 2014). The low accuracy of genomic prediction is expected when the heritability of the traits is low which is true in the case of fertility traits (Cammack et al., 2009; Toghiani et al., 2017). The relationship of low heritability and low accuracies of genomic prediction was reflected in the results of this study where estimates of heritability were found to be 0.16, 0.11, and 0.08, and the accuracy of genomic prediction was found to be 0.17, 0.11, and 0.10 for PREG1, REB, and FCS, respectively, using HD complete SNP set.

For all traits, the accuracy of genomic prediction by biologically informed SNP set was slightly higher (statistically insignificant) than the HD complete SNP set and biologically uninformed SNP set in the Tropical Composite population using 5-fold validation approach in this study. A report on genomic prediction of production traits in pigs explained that dilution of SNP panels with less informative SNP (SNP not associated with target trait) can decrease the predictive ability of the SNP panel (Sarup et al., 2016). Similarly, Ling et al. (2021) described the effect of QTL-linked and nonlinked SNP on the genomic prediction of the traits and reported that inclusion of non-linked SNP in the prediction equation decreased the accuracy of genomic prediction. These studies indicate the importance of prioritizing informative genomic regions and SNP for improved genomic prediction. In the current study, the selection of biologically informed SNP set was based on genomic regions related to puberty and fertility in Brahman breed (studies cited above). The results indicate that the biologically informed SNP set was also effective in predicting fertility traits in another breed Tropical Composite. It is possible that the genomic regions tagged by the biologically informed SNP set were enriched for SNP that were informative for fertility traits in both breeds, hence were useful in genomic prediction in across-breed scenario.

The existence of trait-specific biologically informative SNP in specific genomic regions can also be supported by further results of this study which indicate that the biologically informed SNP set having 110K SNP produced slightly higher (statistically insignificant) estimates of heritability for PREG1 and FCS and nominally better prediction accuracies for all traits. The nominal improvements were observed even though the HD complete SNP set and the uninformed SNP set had a higher number of SNP (>228K compared to 110K). Since the biologically informed SNP were also part of the HD complete SNP set, statistically similar estimates of heritability and prediction accuracy were expected. Even though the biologically informed and the uninformed SNP sets were different, their performance in terms of capturing heritability and predicting traits might be similar because of LD between SNP across these sets. We note that LD can occur within nearby genomic regions and inter-chromosomally (Kulminski, 2011). We accounted for any LD possibility by fitting simultaneously the GRM from biologically informed and uninformed SNP sets in a GBLUP model. Through this model, we were able to partition the contribution of biologically informed and uninformed SNP set towards the heritability of the studied traits. The results indicated that the biologically informed SNP set captured all of the observed heritability for PREG1 and FCS when fitted together with the GRM of biologically uninformed SNP. However, this scenario was different for REB, where 73% of heritability was explained by the biologically uninformed SNP set while 27% was explained by the biologically informed SNP set. We believe that REB might be a more complex trait than PREG1 and FCS because it is influenced by postpartum anestrous interval in addition to puberty (Kinder and Werth, 1991; Tahir et al., 2021). We observed small differences in heritability estimates by biologically informed SNP set when fitted alone and together with biologically uninformed SNP set in the models. These differences might be due to interaction between biologically informed and uninformed SNP sets when fitted as covariates with each other.

Another point to observe is that regardless of molecular biology and SNP function, all the SNP sets could estimate the relationships in our study population, generating useful kinship matrices. This might be the reason why SNP sets were not significantly different in performance. It has been documented that as few as 50 to 100 SNP genotyping might be sufficient for parentage tests in cattle (Van Eenennaam et al., 2007). All our SNP sets had a much larger number of markers than the number mentioned in the cited study.

Although the accuracies of genomic prediction by biologically informed SNP set were not significantly higher than the HD complete SNP set and biologically uninformed SNP set in this study, these were comparable to the other similar studies in cattle. A study on US Jersey and Holstein bulls tested SNP with presumed functional role and top SNP significantly associated (P-value =<0.05) with the trait “sire conception rate” for genomic prediction accuracy. Using linear kernel and multikernal Gaussian models, they reported 2% to 5% increase in accuracy of genomic prediction with top SNP, while SNP with presumed functional roles produced comparable results to the SNP set having all markers (Abdollahi-Arpanahi et al., 2017). In the above-cited study, the SNP with presumed biological function were the markers present within all genes that were part of GO terms enriched in their previous studies while the current study selected the biologically informed SNP from the specific genes that were identified in previous puberty and fertility-related studies. Daetwyler et al. (2019) used prioritized biologically informed variants from whole-genome sequence of cattle to use for genomic prediction of milk and fertility traits in Australian Red, Holstein, and Jersey breeds and found an average increase in accuracy of genomic prediction by 2.5% across purebreds and all traits. A study on Danish Jersey cattle used biologically informed variants related to milk and fertility traits from previous studies to perform genomic prediction and reported increase in genomic prediction accuracy for milk traits but not for fertility traits (Liu et al., 2020). Warburton et al. (2020) used three techniques (GWAS, meta-analysis, and conditional and joint analysis) to preselect top variants associated with a fertility trait “age at first corpus luteum” in Bos indicus-influenced cattle. The preselected variants were combined with standard 6K, 50K, and 800K genotype panels to predict another fertility trait “reproductive maturity score”. When compared with control 6K, 50K, and 800K SNP panels, they found increase in the accuracy of genomic prediction by ~0% to 9%, ~0% to 2%, and ~0% to 4% through Bayes-R approach when top variants identified through different techniques were combined with standard 6K, 50K, and 800K SNP panels, respectively. A recent study reported that use of the XT-50 genotyping array resulted in more accurate genomic predictions for multiple milk-related traits in cattle populations across different countries (Xiang et al., 2021). Using a multibreed Australian cow population as a reference to predict New Zealand purebred and cross-bred cows for three milk traits, the average increase in accuracy of genomic prediction by XT-50 panel compared to standard 50K panel was 9%, 11%, and 7.9% for pure Holstein, pure Jersey, and cross-bred cows, respectively. The increase in prediction accuracies achieved with biologically informed/preselected SNP sets in our study and other studies cited above, was not significantly higher when compared to standard SNP sets. However, a recent report (Reverter et al., 2022) highlights the impact of even small increases in accuracies of genomic predictions. According to Reverter et al. (2022), a nominal increase of approximately 0.05 in prediction accuracy can translate into an increase in discrimination power between the top and bottom quartiles of predicted EBV that can be useful to the industry.

The comparison of biologically informed SNP set from the current study with the XT-50 study by Xiang et al. (2021) revealed overlapping genomic regions between the two marker sets. The XT-50 panel used by Xiang et al. (2021) included ~50-K sequence level pleiotropic variants that were prioritized due to their biological relevance with 34 traits including fertility traits (daughter pregnancy rate and productive life) using multiple statistical techniques on data of ~44,000 Australian dairy cattle. Interestingly, out of 110-K SNP in the biologically informed SNP set of our study, 100-K SNP were in common with HD SNP obtained from the extended genomic regions (10 kb) around XT-50 variants. This large proportion of SNP in common with XT-50 based SNP set indicates that the biologically informed SNP in the current study tagged the same genomic regions as did the XT-50 genotyping array. This finding suggests that the traits of economic importance investigated in both studies are controlled by similar genomic regions. Further investigation on use of biologically informed SNP set from the current study can highlight the potential of this SNP set to predict other traits of economic importance in beef cattle.

In the scenario, when the SNP effects from one breed were used to calculate GEBV of the other breed, the accuracy of genomic predictions for all traits was drastically reduced to around zero by all SNP sets. Hayes et al. (2009) reported that the accuracy of genomic predictions was decreased to around zero when the prediction estimates of Holstein cattle were used to predict GEBV of Jersey cattle and vice versa. Additional studies also contextualize the challenge of across-breed predictions (Erbe et al., 2012; Kachman et al., 2013; Calus et al., 2014). A simple explanation can be that an SNP in LD with a QTL in one breed might not be in LD with the same QTL in the other breed. The inconsistent LD between SNP and QTL in across breed scenario results in drastic decrease in accuracy of genomic prediction. Further investigations with focus on understanding LD structures of different populations and prioritization of the informative SNP across breed are needed to overcome the problem of low accuracy of genomic prediction arising from the use of prediction equations from one breed to the other.

Conclusion

The results of this study indicate that the genomic regions that were identified as relevant to puberty and fertility traits in Brahman heifers through multi-omics studies were also informative in terms of capturing heritability and predicting heifer fertility traits in Tropical Composite heifers similar to a standard HD SNP chip. Overall, the results reported in this study suggest that genomic prediction for complex female reproductive traits can potentially be improved by customizing the SNP panel for genomic predictions using multi-omics data. Future strategies for customization of SNP for fertility traits might include several GWAS and other omics studies performed in different breeds, to create a panel that is more likely to work for any breed. Given the past decade of research, we are now in a position to aggregate biological information from multiple studies that have targeted female reproduction traits, and this can inform future genomic selection strategies.

Glossary

Abbreviations

FCS

first conception score

GBLUP

Genomic best linear unbiased prediction

GEBV

Genomic estimated breeding value

GRM

genomic relationship matrix

GWAS

genome-wide association study

HD

high-density

h²

heritability of the trait

LD

linkage disequilibrium

MAF

minor allele frequency

PREG1

pregnancy at first mating opportunity

QTL

quantitative trait loci

REB

rebreeding score

SNP

single nucleotide polymorphism

Conflict of Interest Statement

The authors declare no real or perceived conflicts of interest.