Genomic prediction via reaction norm models for pregnancy loss using temperature humidity index in brahman cattle Authors: Daniel Cardona-Cifuentes, Lucia G de Albuquerque, Ignacio Aguilar, Milagros Arias, Luis Camaripano, Luis Gabriel Gonzalez-Herrera, Lúcio F M Mota, Fernando Baldi, Sabrina T Amorim Published: 2026 ABSTRACT AbstractBrahman cattle are widely used in tropical production systems due to their adaptability; however, reproductive efficiency remains constrained by pregnancy loss (PL), particularly under heat stress. This study evaluated genotype-by-environment (G × E) interactions for PL using reaction norm models incorporating the temperature–humidity index (THI) as an environmental descriptor. Records from Brahman heifers in two Bolivian herds were combined with weather data to calculate accumulated THI for the 30 d before and after rectal palpation. Pregnancy loss, defined from diagnoses performed between 60 and 150 d of gestation, was analyzed as a binary trait using threshold RNM fitted with either a pedigree-based relationship matrix (A) or a combined pedigree–genomic matrix (H). Genetic variation was detected for both the intercept and slope of the reaction norm. Heritability of the intercept increased from 0.23 (A) to 0.37 (H), whereas the slope showed moderate heritability (0.32 for A; 0.33 for H). Mean heritability of PL across accumulated THI levels was 0.10 (A) and 0.13 (H). Negative genetic correlations between intercept and slope (−0.41 for A; −0.34 for H) and between extreme THI levels (−0.65 for A; −0.49 for H) indicate substantial reranking of animals under contrasting thermal conditions. Although variance component estimates were broadly similar between models, incorporating genomic information improved prediction accuracy and reduced dispersion of breeding values. These results demonstrate that accumulated THI captures meaningful environmental sensitivity for PL and that integrating genomic information with reaction norm models enhances selection for reproductive resilience in Brahman cattle under tropical conditions.Keywords: heat stress, reproductive efficiency, genotype-by-environment interaction, threshold model, single-step evaluation, tropical production systems Introduction Brahman cattle (Bos taurus indicus) are widely used in tropical production systems due to their adaptability to grass-based diets, tolerance to high temperature and humidity, and resistance to tropical diseases (Porto-Neto et al. 2014). These characteristics allow them to maintain productivity in environments where other breeds may struggle. However, despite this adaptability, reproductive efficiency remains a major limitation, with pregnancy loss (PL) representing an important source of economic loss in beef operations (Lunstra and Cundiff 2003). Heat stress has been shown to impair follicular development, reduce conception rates, and increase embryonic and fetal mortality (García-Ispierto et al. 2006; Gaughan et al. 2008; Gómez-Guzmán et al. 2024), highlighting the biological sensitivity of reproductive processes to environmental conditions. Breeding programs can be designed to enhance reproductive traits, but progress is often slow due to their typically low heritability (Bessa et al. 2021; Cavani et al. 2015). Performance differences across environments may arise from genotype-by-environment interaction (G × E), whereby genetic merit depends on environmental conditions (Falconer and Mackay 1996; Hayes et al. 2016). Reaction norm models (RNM) provide a framework to quantify G × E by modeling genetic effects along a continuous environmental gradient (Chiaia et al. 2015; Kolmodin et al. 2003). In RNM, the intercept represents overall genetic merit and the slope reflects environmental sensitivity, allowing identification of animals that are either specifically adapted or robust across environments (Carvalheiro et al. 2019; Kolmodin et al. 2002; Shi et al. 2021). Given the environmental diversity of tropical regions and increasing climatic variability, RNM provide a useful framework for identifying adaptable animals (Silva Neto et al. 2024). In indicine cattle, RNM have been applied using climatic descriptors such as the temperature–humidity index (THI), a widely adopted indicator of heat load (Ravagnolo and Misztal 2000; Bohlouli et al. 2019). Studies in Zebu populations have reported G × E for growth and reproductive traits across THI gradients, suggesting that heat stress can alter genetic expression and limit response to selection in extreme environments (Santana et al. 2016, 2018). The integration of genomic information through single-step genomic BLUP (Misztal et al. 2009) further enhances the estimation of realized genetic relationships and prediction accuracy, improving the ability to select animals with greater environmental resilience (Oliveira et al. 2018; Mota et al. 2020a). Pregnancy loss is particularly relevant in tropical systems, as fetal survival is highly sensitive to thermal stress and management variation. In Bos indicus cattle, PL frequently occurs after 45 d of gestation and is more prevalent in heifers (Prado et al. 2024).Although THI has consistently been associated with impaired reproductive performance, little is known about the genetic architecture of PL across climatic gradients in Brahman cattle raised under tropical conditions. In this context, the present study integrates threshold reaction norm models with single-step genomic evaluation to characterize changes in genetic variance across THI levels and to evaluate how genomic information influences environmental sensitivity, reranking of animals, and prediction accuracy under heat stress conditions. Therefore, this study aimed to evaluate G × E interaction for PL using reaction norm models incorporating accumulated THI as an environmental gradient and to assess the impact of genomic information on genetic parameter estimation and genomic prediction in Brahman cattle raised under tropical conditions. Material and methods Ethics statement Institutional Animal Care and Use Committee approval was not required for this study because all data were obtained from the Associação Nacional de Criadores e Pesquisadores (ANCP, Ribeirão Preto, Brazil) computer database. Phenotypic and genotypic information PL records were collected from Brahman heifers with mating seasons between 1998 and 2021 from two herds in Bolivia. The Estancia Espíritu herd is located in the El Beni region, and the San Judas herd is in the Santa Cruz region, approximately 510 km apart (Fig. 1). Although geographically distant, the herds exhibit genetic connectedness using common or related sires. Geolocation of studied herds in Bolivia. The pedigree used in this study included descendants of 886 sires, providing genetic links across herds and years. A total of 13,372 phenotypic records were available for analysis. Among the heifers with phenotypic records, 287 known sires were represented, including 271 sires in Estancia Espíritu and 27 sires in San Judas. Furthermore, the pedigree included 88 known paternal grandsires and 227 known maternal grandsires. Genetic connectedness between herds was established through 11 common sires, 14 common paternal grandsires, and 9 common maternal grandsires shared between both herds. Of the heifers with phenotypic records, 1053 had unknown parentage (both sire and dam). Despite the smaller number of records from San Judas, the presence of common sires and ancestors provided genetic links between herds, supporting joint genetic evaluation. The distribution of age at pregnancy (27–35 mo) and the overall frequency of pregnancy success and loss across the combined dataset are presented in Fig. 2. Of the total records, 12,975 originated from the Estancia Espíritu herd, where pregnancy loss averaged 12%, whereas 397 records originated from the San Judas herd, with an average pregnancy loss rate of 8%. Proportion of pregnancy success and pregnancy losses and distribution of the heifer age at pregnancy in months. In both herds, calves were weaned between 6 and 8 mo of age and subsequently raised in pasture-based systems with mineral supplementation. Heifers were first exposed to breeding at approximately 24 mo of age (22–26 mo), with a minimum weight requirement of 280 kg. Both farms maintain preventive health programs against bovine viral diarrhea, infectious bovine rhinotracheitis, and Leptospira. Females are vaccinated at the beginning of the breeding season and receive a booster 20–30 d later, while pregnant cows receive a third dose at rectal palpation. Sires are vaccinated annually, 15–30 d before the start of the breeding season. Additionally, females between 3 and 8 mo of age are vaccinated against Brucella abortus (strain 19), and males undergo annual serological testing for Brucellosis detection. Females were exposed to reproduction during a 90- to 100-d breeding season between October and January. Artificial insemination was carried out during the first 45–60 d, after which non-pregnant heifers were subjected to controlled natural mating. Pregnancy was confirmed by rectal palpation approximately 60 d after insemination or at the end of the breeding season when heifers did not return to estrus. Additional pregnancy diagnoses were performed between 60 and 150 d of gestation. Heifers confirmed pregnant but that subsequently aborted or failed to calve were assigned a value of 2, indicating pregnancy loss (PL), while cows that successfully calved received a value of 1. The probability of PL and age at pregnancy (27–35 mo) is shown in Fig. 1. Animals were genotyped using a medium-density SNP array containing approximately 54,000 markers (GGP Indicus 50K, Neogen®). Genotype quality control was performed with preGSf90 (Lourenco et al. 2022). Autosomal markers were excluded if they had a minor allele frequency (MAF) lower than 0.05, a significant deviation from Hardy–Weinberg equilibrium (P ≤ 10-5), a Mendelian conflict rate higher than 0.2, or a call rate below 0.90. Samples were excluded if they had a call rate below 0.90 or a Mendelian conflict rate higher than 0.01. After quality control, 39,521 SNP markers remained for 3,436 genotyped animals from multiple reproductive categories including: heifers, primiparous, non-lactating, and multiparous females. Weather information The herds included in this study are located in the north-central (El Beni) and eastern (Santa Cruz) regions of Bolivia. According to data from the Instituto Nacional de Estadística de Bolivia (INE), between 2005 and 2022 the El Beni region had an annual mean temperature of 26 °C (range: 21–32 °C), average relative humidity of 73%, and annual precipitation of 1744 mm. In Santa Cruz, the annual mean temperature was 24 °C (range: 18–30 °C), with 68% relative humidity and 1025 mm of annual rainfall. Both regions exhibit a similar rainfall pattern characterized by a dry season extending from late May to September–October. Despite both regions being classified as tropical environments, El Beni generally exhibited higher humidity and precipitation, whereas Santa Cruz showed comparatively lower humidity and drier conditions. These environmental differences contributed to variation in accumulated THI across herds and breeding seasons, providing the environmental gradient required for reaction norm analyses. Weather data were retrieved from the NASA POWER database (https://power.larc.nasa.gov) using the geographic coordinates of each farm. Because pregnancy diagnosis by rectal palpation was performed between November and April, weather records were collected from October through May for each breeding season from 1998 to 2021. Daily mean air temperature (T) and dew point temperature (Tdp) were used to calculate the temperature–humidity index (THI) according to Hahn (1999): where T is the average daily air temperature (°C), and Tdp is the average daily dew point temperature (°C). Descriptive statistics for weather variables are presented in Table 1. Figure S1 illustrates the monthly THI patterns across farms and years, highlighting both seasonal fluctuations and temporal variation in thermal load throughout the study period. These results demonstrate the environmental heterogeneity underlying the THI gradient used in the reaction norm analyses. Table 1Descriptive statistics for the weather variables from NASA POWER, daily temperature humidity index (THI), and accumulated temperature humidity index.MeanSDMinMaxn Temperature (°C) 26.892.5514.8535.718,471 Dew point temperature (°C) 21.463.140.1425.468,471 THI 75.812.8658.1383.428,471 Accumulated THI* 4,596.8849.104,413.234,724.9713,372*Accumulated THI calculated as the sum of daily THI values during the 30 d before and 30 d after rectal palpation.SD, standard deviation; Min, minimum; Max, maximum; n, number of records; THI, temperature humidity index. Descriptive statistics for the weather variables from NASA POWER, daily temperature humidity index (THI), and accumulated temperature humidity index. Accumulated THI calculated as the sum of daily THI values during the 30 d before and 30 d after rectal palpation. SD, standard deviation; Min, minimum; Max, maximum; n, number of records; THI, temperature humidity index. Statistical overview Generalized linear models were fitted in R (2024) using the glm procedure with a logit link function to evaluate fixed-effects structure and environmental covariates for PL. Model selection for the fixed-effects structure was based on likelihood-ratio tests, Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). The best-fitting model included age at rectal palpation and contemporary group (CG) as fixed effects. Contemporary groups were defined as the combination of herd, year, and month of rectal palpation, representing females managed under similar environmental, nutritional, and reproductive conditions within each breeding season. Selection of the environmental covariate and its time window was performed within the Bayesian reaction norm framework using the Deviance Information Criterion (DIC). The model incorporating accumulated THI, calculated for the period spanning 30 d before and 30 d after rectal palpation, provided the best fit. A first-order Legendre polynomial (p = 1) and heterogeneous residual variance were retained in the final model to capture variation across the environmental gradient. For quality control, contemporary groups (CG) and environmental levels with fewer than five animals or lacking representation of both phenotypic outcomes were excluded (Fig. 3). Distribution of records across accumulated THI levels calculated for the period spanning 30 d before and 30 d after rectal palpation. Blue bars represent the number of pregnancy diagnoses (records) within each accumulated THI class (left y-axis). dots indicate the corresponding percentage of PL observed at each THI level (right y-axis). contemporary groups and environmental classes with fewer than five animals or lacking both phenotypic outcomes were excluded for quality control. Threshold RNM were fitted for PL using either pedigree information through the additive relationship matrix A (BLUP) or the combined pedigree-genomic relationship matrix H (ssGBLUP) (Aguilar et al. 2010). The threshold RNM was specified as follows: where yij is the underlying liability of binary observation of PL for the animal j recorded in the level l of the environmental covariable (accumulated THI); β is the vector of fixed effects (including age at pregnancy as a covariate); c is the vector of random CG effects; xjl and wjl are incidence vectors; Φk(tl)is the kth Legendre polynomial evaluated at environmental level (tl); γk are fixed regression coefficients modeling the population mean trajectory across THI; αjk={b0j,b1j} are random regression coefficients representing the additive genetic intercept and slope for animal j; finally eijl is the residual term. A first-order Legendre polynomial (p = 1) was selected based on the DIC because it provided the best balance between model fit and parsimony (Supplementary File 1). To account for heterogeneity across the environmental gradient, two residual variance classes were defined based on accumulated THI levels (<4,597 and ≥ 4,597 units). For the threshold model, the residual variance on the liability scale was fixed at σe2=1 within each residual class, as recommended by (Sorensen and Gianola 2002). Covariance structure All random effects were assumed to follow a multivariate normal distribution: where P is either the pedigree relationship matrix (A) or the combined pedigree–genomic matrix (H); ⊗ is the Kronecker product; Kα is the covariance matrix of reaction norm coefficients: Kα= [σb02σb0,b1σb0,b1σb12], with b0 representing the intercept b1is the slope; I is an identity matrix, σc2 is the variance for the CG effect, and R represents heterogeneous residual variance. classes. Flat prior distributions were assigned to fixed effects, whereas scaled inverse chi-square priors were assumed for random effects. The inverse of the combined relationship matrix (H-1) was assembled following Aguilar et al. (2010): where and A22-1 represents the inverse of the pedigree-based relationship matrix for genotyped animals and G-1 is the inverse of the genomic relationship matrix as in VanRaden (2008). The G matrix was computed as: where M is the SNP matrix coded as 0, 1, and 2 for AA, AB, and BB, adjusted for allele frequency as 2pj; and pj is MAF of the jth SNP marker. Genetic parameters along THI gradient The additive genetic variance (σatl2) at environmental level (tl) was computed as: where Φ(tl)=[Φ0(tl),Φ1(tl)]. Heritability at level tl was calculated as: where σa(tl)2 and σe(tl)2 are the genetic and residual variance of each environmental level tl; σc2 the variance of the contemporary group effect. The genetic correlations across THI levels (rtl,tln) were computed as: where σ(tl,tl′) is the covariance between THI levels. Lastly, breeding values at each environmental level were predicted as: Models were implemented using Bayesian inference via Gibbs sampling in gibbsf90+ (Lourenco et al. 2022). Chains were run for 1,000,000 iterations, with 200,000 burn-in and a thinning interval of 20, resulting in 40,000 posterior samples. Convergence was assessed by visual inspection of trace plots and by Geweke’s diagnostic (Geweke 1992) using the BOA R package. Validation A forward cross-validation scheme was applied to evaluate the prediction accuracy of breeding values across environmental levels and for the RNM coefficients of additive genetic effects ({b0j,b1j}), which define the intercept and slope of the reaction norm curves. The validation set included 311 genotyped heifers born in 2018 (263 from Estancia Espíritu and 48 from San Judas). For RNM fitted using either the pedigree relationship matrix A or the combined pedigree–genomic relationship matrix H, prediction accuracy was assessed using the linear regression LR method (Legarra and Reverter 2018). In this approach, partial breeding values (u^p) re obtained by analyzing the dataset excluding validation records, whereas whole breeding values (u^w) are estimated using the complete dataset including validation animals. The following validation metrics were calculated: Dispersion: Evaluated as the regression of whole Estimated Breeding Value (EBV) on partial EBV, bw,p=cov (u^w,u^p)var(u^p) with the expectation E(bw,p)=1 in the in the absence of over- or under-dispersion. Dispersion: Evaluated as the regression of whole Estimated Breeding Value (EBV) on partial EBV, bw,p=cov (u^w,u^p)var(u^p) with the expectation E(bw,p)=1 in the in the absence of over- or under-dispersion. Bias: Calculated as the difference between partial and Whole EBV, Δ^p,w=u^p¯-u^w¯ with an expected value of zero if predictions are unbiased. Bias values were standardized by the additive genetic standard deviation in each scenario. Bias: Calculated as the difference between partial and Whole EBV, Δ^p,w=u^p¯-u^w¯ with an expected value of zero if predictions are unbiased. Bias values were standardized by the additive genetic standard deviation in each scenario. Accuracy Ratio: The ratio of accuracies from the partial and whole datasets (accpaccw), estimated as the correlation between the partial and whole EBV, ρw,p=cov (u^w,u^p)var(u^p)var(u^w) Accuracy Ratio: The ratio of accuracies from the partial and whole datasets (accpaccw), estimated as the correlation between the partial and whole EBV, ρw,p=cov (u^w,u^p)var(u^p)var(u^w) Accuracy of partial Genomic Estimated Breeding Value (GEBV): Estimated from the covariance between partial and whole EBVs, acc^p=cov(u^w,u^p)(1+F¯-2f¯)σu,∞2 Accuracy of partial Genomic Estimated Breeding Value (GEBV): Estimated from the covariance between partial and whole EBVs, acc^p=cov(u^w,u^p)(1+F¯-2f¯)σu,∞2 Where F¯ is the average inbreeding coefficient, f¯ is the average relationship among validation individuals, and σu,∞2 is the genetic variance of the validation group. Under the assumption of no prior selection among validation individuals and F≈f, then the denominator simplifies to:(1-F¯)σu2 (Legarra and Reverter 2018; Macedo et al. 2022 a). Results Genetic parameter and (co)variance components The variance attributed to contemporary group was similar under both the pedigree-based (A) and single-step genomic (H) models, whereas additive genetic (co)variances were larger when genomic information was included. Both random regression coefficients showed moderate heritability and a moderate negative genetic correlation. Incorporating the H matrix increased the heritability of b0​, while reducing the magnitude of the genetic correlation between intercept and slope. Posterior standard deviations were generally larger under the H matrix (Table 2). The posterior mean of the fixed regression intercept was −1.47 (SD = 1.16), indicating a low baseline liability for pregnancy loss across the environmental gradient. The fixed regression slope was positive (0.073; SD = 0.12), suggesting a slight increase in population-level liability for pregnancy loss as accumulated THI increased. Table 2Posterior means for (co)variance components and genetic parameters of reaction norm for pregnancy loss, using accumulated THI as environmental covariable and using pedigree-relationship matrix (A matrix) and pedigree-genomic relationship matrix (H matrix).Relationship matrixVariance component A matrix H matrix b0 Genetic variance (σb02) 0.23 (0.10)0.37 (0.15) b1 Genetic variance (σb12) 0.32 (0.15)0.33 (0.15) Genetic covariance (σb0,b1) −0.12 (0.10)−0.13 (0.13) b0 Heritability 0.18 (0.05)0.24 (0.07) b1 Heritability 0.22 (0.07)0.22 (0.07) Genetic correlation (rb0,b1) −0.41 (0.24)−0.34 (0.30) VCG 0.09 (0.02)0.09 (0.03) DIC −27,594.44−29,803.01 b0 and b1: reaction norms coefficients for genetics additive random effect (b0: intercept, b1: slope); VCG: Variance of contemporary group random effect (σc2); DIC: Deviance information criterion. Posterior means for (co)variance components and genetic parameters of reaction norm for pregnancy loss, using accumulated THI as environmental covariable and using pedigree-relationship matrix (A matrix) and pedigree-genomic relationship matrix (H matrix). b0 and b1: reaction norms coefficients for genetics additive random effect (b0: intercept, b1: slope); VCG: Variance of contemporary group random effect (σc2); DIC: Deviance information criterion. Mean heritability for on the liability scale across accumulated THI levels was 0.13 using the A matrix and 0.17 using the H matrix. Heritability was lowest (0.08 and 0.13 for the A and H matrices, respectively) near the average THI and increased toward both extremes of the environmental gradient (Fig. 4). Across THI levels, estimates obtained with the H matrix were consistently higher than those from the A matrix. Posterior means for heritability and standard deviation for pregnancy loss across levels of accumulated THI and using pedigree-relationship matrix (A-matrix) and pedigree-genomic relationship matrix (H-matrix). Genetic correlations between THI levels were high and positive for adjacent environments but declined toward zero or became moderately negative as the environmental distance increased (Fig. 5). Mean correlations were slightly higher under the H matrix (0.77 ± 0.25) than under the A matrix (0.70 ± 0.35). Genetic correlations between distant THI levels were more negative under the A matrix (−0.65) compared with the H matrix (−0.49), indicating stronger reranking of animals across environmental conditions when genomic information was not included. Posterior means for genetic correlations for PL across accumulated THI levels estimated using reaction norm models. Panel (a) presents estimates obtained using the pedigree-based relationship matrix (A-matrix), whereas panel (b) shows estimates using the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. The three-dimensional surfaces represent genetic correlations between PL expressed at different accumulated THI levels (x- and y-axes), with color gradients indicating the magnitude of correlation (red = lower or negative correlations; green = higher positive correlations). Genotype by environment interaction Figure 6 presents the (G)EBV trajectories across accumulated THI levels for the 50 sires with the most favorable predicted values at the mid-range THI (4605 units), selected from the overall population evaluated across both herds. Both the pedigree-based (A) and single-step genomic (H) models revealed substantial reranking of sires across the environmental gradient, providing clear evidence of genotype-by-environment (G × E) interaction. Breeding value trajectories for PL across accumulated THI levels for the 50 sires with the most favorable predicted breeding values at the intermediate THI level (4605 units). Panel (a) shows estimates obtained using the pedigree-based relationship matrix (A-matrix), whereas panel (b) shows estimates from the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. Each line represents the breeding value trajectory of one sire across the environmental gradient. Colors indicate the direction of the reaction norm slope coefficient (a1): red lines represent positive slopes (greater sensitivity to increasing THI), blue lines represent negative slopes (declining breeding values with increasing THI), and green lines represent near-zero slopes (greater environmental robustness across THI levels). The crossing of trajectories across the THI gradient illustrates genotype-by-environment interaction and reranking of sires under contrasting environmental conditions. Differences among trajectories were primarily driven by variation in slope coefficients. Some sires showed increasing PL risk as accumulated THI increased, whereas others maintained relatively stable breeding values across environments. This pattern reflects the moderate negative genetic correlation between intercept and slope, indicating that animals with lower average PL risk may differ in their environmental sensitivity. Sires with flatter trajectories demonstrated greater robustness across THI levels, whereas those with steeper slopes exhibited greater environmental sensitivity. These results highlight the opportunity to select animals not only for lower average PL risk but also for reduced sensitivity to thermal stress. Genetic prediction validation Bias and dispersion for both reaction norm coefficients were close to their expected values (zero for bias and one for dispersion), indicating overall agreement between partial and whole predictions. Incorporating genomic information did not substantially affect bias but reduced dispersion, particularly for the slope coefficient. Accuracy of partial breeding values was generally low for both coefficients, especially for the slope (b1). However, single-step genomic prediction increased the ratio of accuracies (ρ^w,p), demonstrating improved relative accuracy of partial evaluations compared with whole-data predictions (Fig. 7). Accuracy (acc), bias, dispersion (b^w,p), and ratio of accuracies (ρ^w,p) for random coefficients—intercept (b0) and slope (b1), using the pedigree-relationship matrix (A-matrix) and pedigree-genomic relationship matrix (H-matrix). Across accumulated THI levels, bias remained small under both approaches, although slightly higher under single-step prediction at intermediate THI values (Fig. 8). In contrast, dispersion across the environmental gradient was notably reduced under single-step prediction, indicating improved stability of genomic predictions prior to phenotype inclusion (Fig. 9). Bias of predicted breeding values for PL across accumulated THI levels using the pedigree-based relationship matrix (A-matrix) and the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. Bars represent bias estimates at each environmental level. Dispersion of partial breeding values for PL across accumulated THI levels using the pedigree-based relationship matrix (A-matrix) and the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. Bars represent dispersion estimates (expected value ≈ 1.0) at each environmental level. Prediction accuracy was lowest at the lowest THI levels, increased toward intermediate THI values, and slightly declined at the highest THI levels (Fig. 10). Nevertheless, single-step genomic prediction consistently improved accuracy across the environmental gradient compared with the pedigree-based model. Similarly, the ratio of accuracies (ρ^w,p) generally increased across the THI gradient under single-step prediction (Fig. 11). At elevated THI levels, partial and whole breeding value accuracies became comparable. Prediction accuracy of breeding values for PL across accumulated THI levels using the pedigree-based relationship matrix (A-matrix) and the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. Bars represent accuracy estimates at each environmental level. Ratio of accuracies between partial and whole breeding values ρ^wp for PL across accumulated THI levels using the pedigree-based relationship matrix (A-matrix) and the combined pedigree–genomic relationship matrix (H-matrix) under ssGBLUP. Discussion Our study evaluated PL detected between 60 and 150 d of gestation, corresponding to the fetal loss period (Reese et al. 2020). This stage is particularly susceptible to environmental stressors, including thermal stress, which can impair placental development and fetal viability (Pegorer et al. 2007; Diskin and Kenny 2016; Consentini et al. 2023). The selection of a 30-d window before and after rectal palpation likely captured cumulative thermal stress affecting pregnancies that were initially established but later failed. Although repeated pregnancy diagnoses could improve temporal precision in identifying losses (Pohler et al. 2020; Reese et al. 2020), such management practices are uncommon in extensive tropical systems. Covariance components and genetic parameters The model including accumulated THI during the 30 d before and after rectal palpation provided the best fit, supporting the biological relevance of cumulative heat load around pregnancy diagnosis. Thermal stress is known to impair reproductive physiology and pregnancy maintenance, even in heat-adapted indicine breeds (Avendaño-Reyes et al. 2010; Santana et al. 2017; Shi et al. 2021), reinforcing the suitability of accumulated THI as an environmental descriptor for PL. Nevertheless, a potential limitation of this study is that weather variables were obtained from the NASA POWER database rather than from on-farm weather stations. Although NASA POWER data are widely used and generally show strong agreement with ground-based measurements, they may not fully capture local microclimatic conditions experienced by individual animals, such as shade availability, wind exposure, pasture conditions, or short-term heat events (Carrara 2023). Consequently, some degree of measurement error in the THI covariate is expected, which could attenuate estimates of environmental sensitivity and G × E interaction. However, the use of farm-specific geographic coordinates and accumulated THI over the 30 d before and after rectal palpation likely reduced the impact of short-term daily discrepancies. Therefore, the observed G × E patterns should be interpreted as reflecting genetic sensitivity to broad-scale cumulative thermal load rather than exact individual-level heat exposure. Future studies incorporating on-farm sensors or animal-level environmental monitoring may further improve characterization of reproductive resilience under heat stress. Moderate heritability estimates for the reaction norm coefficients indicate genetic variation in environmental sensitivity. In particular, genetic variability in the slope confirms that accumulated THI influences PL through genotype-by-environment interaction, consistent with previous findings in indicine cattle for growth and reproductive traits (Carvalheiro et al. 2019). Furthermore, genetic variability in slopes has been associated with environmental factors affecting proper gestation development in both heifers and cows (Alves et al. 2022; Santana et al. 2023). The moderate negative genetic correlation between intercept and slope provides further evidence of G × E, reflecting reranking of animals across environmental levels. Biologically, this suggests a trade-off between overall genetic merit for PL and environmental sensitivity: animals with lower average PL risk may differ in their response to increasing thermal load. However, because the correlation was far from −1, there remains substantial opportunity to identify animals that combine low average PL with reduced sensitivity to THI, supporting simultaneous selection for performance and robustness. Heritability of PL was low near the average THI and across the overall gradient, consistent with previous estimates reported for the same population (Cardona-Cifuentes 2026), and with reports for reproductive traits in beef and dairy cattle (Carthy et al. 2015; Ask-Gullstrand et al. 2021; Sigdel et al. 2022). Nevertheless, heritability increased toward the extremes of the environmental gradient. Although this pattern has been observed in several species and traits under environmental stress (Brügemann et al. 2013; Shi et al. 2014; Santana et al. 2017), caution is warranted because random regression models may inflate variance estimates at the edges of the covariate distribution due to fewer records. Even so, the increase in heritability under extreme THI suggests amplified genetic variability under thermal stress, possibly reflecting heterogeneous genetic responses among animals. Importantly, genetic correlations across THI levels declined as environmental distance increased, with moderate negative correlations observed between distant environments. Correlations below 0.8 are indicative of meaningful G × E (Robertson 1959), and the values observed here confirm substantial reranking across accumulated THI levels. Similar patterns have been reported for Zebu cattle in Brazil for a range of traits, including scrotal circumference, heifer pregnancy, age at first calving, and growth traits (Chiaia et al. 2015; Santana et al. 2015, 2018; de Araujo Neto et al. 2018; Oliveira et al. 2018; Mota et al. 2020a), reinforcing that environmental heterogeneity in tropical systems can alter rankings. Compared with previous studies that evaluated G × E for heifer pregnancy, age at first calving, scrotal circumference, and growth traits in Zebu cattle, the present results extend this evidence to PL in Brahman cattle, indicating that reproductive resilience after pregnancy establishment is also affected by environmental sensitivity to thermal load. Together, these results indicate that incorporating environmental covariates such as THI into genetic evaluations can improve identification of animals that are either specifically adapted to challenging conditions or capable of maintaining stable performance across diverse environments. Genotype by environment interaction Reaction norms similar to those observed in our study have been reported for reproductive traits in Nellore cattle by Mota et al. (2020a) and Santana et al. (2017), reinforcing the presence of G × E under thermal stress. Together with the moderate negative genetic correlations among regression coefficients and across THI levels, our results provide clear evidence that accumulated heat load alters genetic expression for PL. Although Zebu cattle are generally adapted to tropical environments, substantial genetic variability exists among individuals in their response to thermal stress. The marked reranking of (G)EBV across accumulated THI levels indicates that genetic merit for PL is environment dependent. This finding highlights the importance of accounting for environmental heterogeneity in breeding programs operating under tropical conditions. Selection based solely on average performance may favor animals that perform well under moderate conditions but are more sensitive to thermal stress. Although phenotypic records originated from only two herds, environmental variation was observed across herds and breeding seasons. The El Beni region was characterized by higher humidity and precipitation, whereas Santa Cruz presented comparatively drier conditions, contributing to differences in accumulated THI over time. In addition, the long study period (1998–2021) captured substantial temporal climatic variation within herds. Therefore, the environmental gradient used in the reaction norm analyses reflected both spatial and temporal variation in thermal load rather than a simple herd effect. This environmental heterogeneity likely contributed to the reranking of animals across THI levels and supports the biological relevance of the detected G × E interaction for PL. The observed genetic variation in slope confirms the opportunity to incorporate robustness into selection objectives. Animals with flatter reaction norms maintain more stable performance across environments and therefore represent desirable candidates for breeding in variable climatic conditions. In practical terms, selection strategies should prioritize animals that combine low average PL risk with limited environmental sensitivity, thereby improving reproductive efficiency under both favorable and stressful conditions. Using the H matrix and genetic prediction Incorporating the H matrix increased heritability estimates for the intercept b0 and across levels of accumulated THI. Similar trends have been reported for yearling weight (Oliveira et al. 2018) and heifer pregnancy in Nellore cattle (Mota et al. 2019). The inclusion of genomic information improves estimation of realized genetic relationships and enhances connectedness among animals across herds, particularly in populations with incomplete pedigree information (Tonussi et al. 2021). In addition, Mota et al. (2020a) and Amorim et al. (2020) emphasized that genomic information enhances genetic connectedness across herds by more accurately capturing animal genetic effects. This improved relationship structure likely contributed to the higher genetic variance captured under ssGBLUP. However, higher posterior standard deviations at intermediate THI levels suggest that genomic information also increased variability in parameter estimates, potentially reflecting the limited number of genotyped sires with recorded daughters in this dataset. The inclusion of genomic information slightly reduced genetic correlations between extreme THI levels and between reaction norm coefficients but did not materially change the detection of G × E interaction. Reranking across THI gradients remained evident under both A and H matrices, indicating that genomic information refined but did not alter the overall G × E pattern. Validation results demonstrated that ssGBLUP reduced dispersion and improved the relative accuracy of partial breeding values, particularly for environment-specific predictions. Although bias increased slightly under ssGBLUP at intermediate THI levels, it remained small overall. Reduced dispersion under the H matrix suggests improved stability of genomic predictions prior to phenotype inclusion, consistent with previous reports showing that genomic information enhances predictive ability in reaction norm models (Oliveira et al. 2018; Mota et al. 2020a). As expected for a low-heritability trait, the overall accuracy of (G)EBV for reaction norm coefficients and across THI levels was modest. Lower accuracy at extreme THI levels likely reflects the smaller number of phenotypic records in those environments, a known limitation of random regression approaches when data are sparse (Pégolo et al. 2009). This pattern differed from the dispersion results because calibration of predictions and predictive accuracy measure distinct aspects of model performance; therefore, predictions may remain well calibrated even when accuracy decreases at environmental extremes due to limited data density. Although prediction accuracies were modest, the presence of additive genetic variation and moderate heritability estimates for the reaction norm slope indicates that genetic improvement for environmental sensitivity is feasible. Collectively, these findings support the use of genomic information in RNM for PL. Even though variance component estimates were not substantially more precise under the H matrix, the reduction in dispersion and improvement in prediction accuracy indicate clear advantages for genomic-enabled selection. Integrating ssGBLUP with environmental descriptors such as THI provides a practical framework for identifying animals with both low average PL risk and improved resilience to thermal stress. Conclusion Pregnancy loss in Brahman cattle is influenced by G × E interaction across accumulated THI levels, with clear reranking of animals under contrasting thermal conditions. Accumulated THI proved to be a biologically relevant environmental descriptor for capturing variation in reproductive performance under tropical conditions. The moderate heritability estimates obtained for both the intercept and slope of the reaction norm indicate the presence of additive genetic variation not only in average PL risk but also in environmental sensitivity to heat stress. These results support the feasibility of selecting animals that combine lower PL risk with greater environmental robustness across THI levels. Incorporating genomic information improved the predictive performance of reaction norm models, supporting the use of single-step genomic evaluation as a practical strategy to enhance selection for reproductive resilience and heat-stress adaptation in tropical beef production systems. Supplementary Material Acknowledgments We acknowledge FAPESP (São Paulo Research Foundation, grant #2021/04807-2) for financial support. The authors acknowledge ANCP (Associação Nacional de Criadores e Pesquisadores, Ribeirão Preto, São Paulo, Brazil) for the support regarding data availability, computing resources and all the Brahman farmers and technicians for the data collection. F.B held productivity research fellowships from The Brazilian National Council for Scientific and Technological Development (CNPQ). Glossary List of abbreviations Akaike Information Criterion Bayesian Information Criterion best linear unbiased prediction contemporary group Deviance Information Criterion estimated breeding value genomic estimated breeding value genomic selection genotype by environment interaction Instituto Nacional de Estadística de Bolivia Legarra–Reverter method minor allele frequency NASA Prediction Of Worldwide Energy Resources pregnancy loss reaction norm model(s) single nucleotide polymorphism single-step genomic BLUP temperature–humidity index Contributor Information Daniel Cardona-Cifuentes, Departamento de Zootecnia, Faculdade de Ciências Agrarias e Veterinárias, Universidade Estadual Paulista (UNESP), Jaboticabal, SP, 14884-900, Brazil; Facultad de Ciencias Agrarias, Fundación Universitaria Agraria de Colombia (UNIAGRARIA), Bogotá, 111166, Colombia. Lucia G de Albuquerque, Departamento de Zootecnia, Faculdade de Ciências Agrarias e Veterinárias, Universidade Estadual Paulista (UNESP), Jaboticabal, SP, 14884-900, Brazil. Ignacio Aguilar, Instituto Nacional de Investigación Agropecuaria (INIA), Montevideo, 11500, Uruguay. Milagros Arias, Facultad de Ciencias Veterinarias, Universidad Central de Venezuela, Maracay, 2103, Venezuela. Luis Camaripano, Grupo Ganadero - Estancias Espíritu, Santa Cruz de la Sierra, Bolivia. Luis Gabriel Gonzalez-Herrera, Grupo de Investigación en Biodiversidad y Genética Molecular (BIOGEM), Universidad Nacional de Colombia, Medellín, 050034, Colombia. Lúcio F M Mota, Departamento de Zootecnia, Faculdade de Ciências Agrarias e Veterinárias, Universidade Estadual Paulista (UNESP), Jaboticabal, SP, 14884-900, Brazil. Fernando Baldi, Faculdade de Zootecnia e Engenharia de Alimentos, Universidade de São Paulo, Rua Duque de Caxias Norte, Pirassununga, SP, 13635-900, Brazil. Sabrina T Amorim, Department of Animal and Food Sciences, Oklahoma State University, Stillwater, OK, 74078, United States. Supplementary data Supplementary data is available at Translational Animal Science online. Funding FAPESP (São Paulo Research Foundation, grant #2021/04807-2). Author contributions DCC, LFMM, FB and STA conceived the study. DCC carried out the data analyses and prepared the first draft of the paper. LGA, IA, MA, LC, and LGGH provided technical support for the analyses, revised and edited the manuscript, and contributed to the interpretation and discussion of the results. All authors have read and approved the final manuscript. Conflicts of interest The authors declare that they have no competing interests. Ethical approval The datasets used were obtained from pre-existing databases. Therefore, Animal Care and Use Committee approval was not required for this study. Data availability The data supporting the results of this article are property of the ANCP (Associação Nacional de Criadores e Pesquisadores, Ribeirão Preto, São Paulo, Brazil) cattle producers, and this information is commercially sensitive, therefore the data cannot be made available. Literature cited Associated Data Supplementary Materials Data Availability Statement The data supporting the results of this article are property of the ANCP (Associação Nacional de Criadores e Pesquisadores, Ribeirão Preto, São Paulo, Brazil) cattle producers, and this information is commercially sensitive, therefore the data cannot be made available.