IIIVmrMLM.QEI: An effective tool for indirect detection of QTN-by-environment interactions in genome-wide association studies Authors: Ya-Wen Zhang, Xue-Lian Han, Mei Li, Ying Chen, Yuan-Ming Zhang Published: 2024 ABSTRACT AbstractAlthough 3VmrMLM-MEJA and several indirect indicators have been employed to identify QTN-by-environment interactions (QEIs) in genome-wide association studies (GWAS), there is no convenient, flexible, and accurate method to comprehensively identify QEIs. To address this issue, 3VmrMLM-random was first extended to 3VmrMLM-fixed. Next, the two single-environment QTN detection methods were integrated with trait differences and regression parameters to indirectly detect QEIs. Finally, these indirect indicators were extended to include environmental factors (EFs, such as temperature) and four environmental variation indicators. As a result, both 3VmrMLM-random and 3VmrMLM-fixed, alongside all the indirect indicators, were incorporated into a new tool, IIIVmrMLM.QEI, designed for effective QEI identification. Simulation studies demonstrated that 3VmrMLM-fixed showed significantly higher powers than existing fixed-SNP-effect methods (MLM and EMMAX) because it takes into account all the possible effects and controls for all the possible polygenic backgrounds. 3VmrMLM-random and 3VmrMLM-fixed exhibited superior combination power to 3VmrMLM-MEJA. In the re-analysis of Arabidopsis flowering time across three temperatures, 3VmrMLM-fixed (12) detected more known gene-by-environment interactions (GEIs) than both MLM (1) and EMMAX (1). Additionally, IIIVmrMLM.QEI (18) detected more known GEIs than 3VmrMLM-MEJA (6), when all indirect indicators were analyzed. All 18 GEIs were confirmed by haplotype analysis and associated with temperature variation in previous studies. Two and five GEIs were identified only by 3VmrMLM-fixed and 3VmrMLM-random, respectively, and 12 GEIs were identified only by indirect indicators, indicating the need to expand models and indirect indicators. This study provides a novel tool (https://github.com/YuanmingZhang65/IIIVmrMLM.QEI) for more comprehensive QEI detection.Keywords: QTN-by-environment interaction, IIIVmrMLM.QEI, Genome-wide association studies, Flowering time, Variation indicator, Environmental factor Introduction Gene-by-environment interaction (GEI) is an important genetic component of complex traits, reflecting the fact that the effects of a gene may vary between environments or may only be detectable in certain environments. With the escalating impacts of global climate change, sustainable crop production faces significant challenges. Developing climate-resilient crops necessitates identifying quantitative trait nucleotides (QTNs) and QTN-by-environment interactions (QEIs). Over the past decade, several indirect approaches have been used to identify QEIs in genome-wide association studies (GWAS). First, QEIs are identified using multi-trait GWAS methods, where multiple environments are considered as multiple traits [1]. Their power is equal to the trait difference as phenotype [2], [3]. Second, a number of indirect indicators have been used to detect QEIs via some widely used GWAS methods, such as genetic risk score [4], [5], environmental score [6], and regression parameters [7]. However, the sample sizes for these approaches are equivalent to the number of individuals (n) in a single environmental experiment. This reduces the power of QEI detection and makes QTN detection infeasible [8]. To address this issue, Sul et al. [9] and Moore et al. [10] developed mixed model-based methods. However, their mixed models only considered allelic substitution effects and their polygenic backgrounds, and did not include additive and dominance effects, additive-by-environment (ae) and dominance-by-environment (de) interaction effects, and their corresponding polygenic backgrounds. To address the detection of interactions in human genetics, the Brown-Forsythe test [11], [12] and the double generalized linear model [13] have been used to detect variance controlling loci, although their power is relatively low [14]. It should be noted that these mixed model based methods are based on the assumption of random mating. To overcome these problems, we have recently established a compressed variance component mixed model framework, namely 3VmrMLM [8]. In our 3VmrMLM method, the assumption of random mating is not required, all the effects to be detected and estimated in QTN, QEI, and QTN-by-QTN interaction detection are compressed into an effect-related vector, and all the possible polygenic backgrounds are compressed into an effect-vector-related polygenic background, and the sample size in multi-environment joint GWAS analysis is equal to mn, where m is the number of environments and n is the number of individuals. Although 3VmrMLM has been established, there are limited methods available to comprehensively identify QEIs. To address the above issue, this study developed a novel tool, IIIVmrMLM.QEI, for more comprehensive QEI detection. In IIIVmrMLM.QEI, 3VmrMLM-random was first extended to 3VmrMLM-fixed, and these two methods were integrated with trait differences, regression intercept (RI), and regression coefficient (RC) as phenotypes for indirectly identifying QEIs. More importantly, the above three indirect indicators were extended to environmental factor (EF, such as temperature) and four environmental variation indicators: range, variance (Var), standard deviation (SD), and coefficient of variation (CV). These environmental variation indicators effectively capture phenotypic differences and variability across environments, providing a direct measure of phenotypic response in different environments. In addition, RI evaluates general performance, RC measures genotype-specific phenotypic changes across different environments, and EF assesses adaptability. QEIs, detected using these methods and indicators, are more comprehensive than those identified by existing approaches. Monte Carlo simulations validated all methods in the tool, which were also benchmarked against current methods [8], [15], [16], highlighting the need to extend the fixed model and indirect indicators. Researchers will benefit from the IIIVmrMLM.QEI for more comprehensive QEI and GEI identification, and the extension of year / location environments to meteorological factors and other environmental conditions. To distinguish these new methods, the integration of 3VmrMLM-random with RI as phenotypes is abbreviated as 3VmrMLM-random-RI, and the others are similar. Materials and methods Association mapping population The well-known Arabidopsis thaliana flowering time dataset was downloaded from http://www.arabidopsis.usc.edu/ [17]. A total of 199 accessions were genotyped by 216130 SNP markers, and 192 accessions were phenotyped by flowering time under 16 h daylight at 10 ℃, 16 ℃, and 22 ℃ conditions, representing three distinct environments. The datasets were listed in Data Files S1-S5. 200 kb upstream and downstream regions of significant / suggested QEIs were used to mine known FT genes [18], while the 200 kb was changed as 50 kb for all the candidate GEIs. Monte Carlo simulation study The Monte Carlo simulation studies simulated two scenarios similar to those in our previous studies [19], [20]. The first dataset was simulated by sampling 200 individuals from 1136 Simmental cattle (https://datadryad.org/stash/dataset/doi:10.5061/dryad.4qc06; Zhu et al. [21]), each with 20000 SNP markers selected from 597650 SNP markers. To simulate the phenotypic values in three environments, three QTNs, three QEIs, and four QTNs with main and environmental interaction effects (r2: 2 %∼8 %) were simulated on 10 chromosomes. The numbers, positions, and sizes of QTNs and QEIs are listed in Table S1. All the phenotypes were simulated based on the model y=Xα+ZEe+∑i=110Ziγi+∑i=110Zeiγei+ε, where e is environmental effect, ZE is its design matrix; γi is the effect of the ith QTN, and Zi is its design matrix; γei is the effect of the ith QEI, and Zei is its design matrix; ε∼MVN(0,10×Ibn), b=3, and n=200. 1000 replicates were performed to evaluate the performances of all the methods, including statistical power of all the simulated loci, false positive rate (FPR) and false negative rate (FNR), and the accuracies for the estimates for positions and effects of QTNs and QEIs, as described in our previous studies. Significant loci were determined based on Bonferroni correction, while suggested loci for 3VmrMLM related methods were determined based on LOD score ≥ 3.0 [8]. All the simulation datasets are listed in Data Files S6-S9. The second dataset was simulated using PLINK 1.9 [22], consisting of 1000 (n) individuals with one million markers each. To simulate the phenotypic values in five (b) environments, three QTNs, three QEIs, and four QTNs with main and environmental interaction effects (r2: 1 %∼7 %) were simulated on one chromosome. The number of replicates was 20. The numbers, positions and sizes of the QTNs and QEIs are listed in Table S9, while all the simulation datasets are listed in Data Files S10-S13. 3VmrMLM-fixed: Fixed SNP effect model of QTN detection in single environment data analysis Genetic model As described in Li et al. [8], genetic model for complex trait y in QTN detection using 3VmrMLM method is where α is a fixed effect vector, X is its design matrix for α, γt is a 3 × 1 vector of QTN genotypic effects, and Zt is the n × 3 design matrix for γt, in which marker genotypes are coded as (1, 0, 0) for AA, (0, 1, 0) for Aa, and (0, 0, 1) for aa; u∼MVN(0,Kϕ2) is a vector of QTN genotype polygenic effects, where K=1m∑t=1mZtZtT is kinship matrix, ϕ2 is polygenic background variance, and m is the number of markers; ε∼MVN(0,Iσe2) is an independent and identically distributed vector of residual errors, σe2 is residual variance, and I is unit matrix. If the vector γt in model (1) is treated as fixed, the above model is transformed into Thus, we have where λg=σg2/σe2 is the variance ratios. The λg value can be pre-estimated under pure polygenic model, and λg is fixed as λˆg when testing each SNP effect in genome-wide scanning. Using eigen decomposition for K, we can obtain K=UDU′, where D=diag{δ1,…,δn} is a diagonal matrix for the eigenvalues and U is an n×n matrix for the eigenvectors. Let yc=U′y, Xc*=U′X*, uc=U′u, and εc=U′ε, model (2) was changed into Restricted log-likelihood function Restricted log-likelihood function of yc is where q=rank(Xc*). Thus, restricted maximum likelihood (REML) estimates for parameters of interest can be obtained as below. Estimation of residual varianceσe2and fixed effectαˆ* Using REML function, we can obtain that: where H=Dλˆg+I. Tests for locus related effects Wald test statistic for locus-related effects γt under H0:γt=0 is where varγˆt can be obtained fromvar(αˆ*)=σˆe2(Xc*′H−1Xc*)−1. The probability for H0 is calculated from 1−Pr(χν2 22 °C WT-spa123.636485151 ∼ 648515115.30 ∼ 16.461.71 ∼ 2.04a ∼ b/1VRN12.87E−04 * **4.14E−02 *VernalizationFLC mRNA levelsUnlike fca−1, FLC levels in vrn1 −1fca−1 increased upon return to normal temperatures, highlighting VRN1's role in stably maintaining FLC repression during later development in warm temperatures.74227242 ∼ 3106578.58 ∼ 24.081.60 ∼ 5.49a ∼ b/3,4,7LIF24.54E−03 * *3.35E−02 *VernalizationFlowering timeVernalization: lhp1/FRI-Col - FRI-Col > Non-Vernalization: lhp1/FRI-Col - FRI-Col, LIF2 primarily functions downstream of LHP1.841261211 ∼ 12719209.99 ∼ 25.183.29 ∼ 6.56b/1,4,6 ∼ 8GA12.26E−04 * **2.65E−02 *TemperatureThe accumulation of GFP-RGA29 °C: WT-ga1 −3 > 20 °C WT-ga1 −39410804201 ∼ 111664025.45 ∼ 6.860.41 ∼ 0.63a/1,8TSF7.51E−07 * **1.95E−03 * *VernalizationmRNA levelsIn fca−1 and FRI-Sf2 (Col), TSF mRNA levels were reduced. Vernalization treatment restored TSF expression in fca−1 and FRI-Sf2 (Col).104111664026.860.63a/1JMJ143.83E−08 * **8.66E−05 * **TemperatureExpression of target geneSeven selected genes were upregulated in Col−0 at 27 °C, but downregulated in the icdh−1/jmj14 −1 mutant under the same conditions compared to Col−0.11412628000 ∼ 126357828.90 ∼ 12.530.44 ∼ 0.98a, c/1MOS14.00E−14 * **5.25E−08 * **Cold treatmentExpression of target geneIn both rpp4 −1 and rpp4 −1mos1, the expression of PR1 increased with cold treatment, but in rpp4 −1mos1 it did not rise as much as rpp4 −1.124180678983.45 ∼ 11.100.73 ∼ 2.50a ∼ c/1 ∼ 3,5,8BBX194.72E−04 * **3.02E−02 *High temperatureFlowering timeUnder high temperature, the level of plant endogenous hormone GAs is downregulated, while the delayed flowering in of 35GS:BBX19 is inhibited by GA.1353055565 ∼ 33011103.95 ∼ 22.360.48 ∼ 3.36a ∼ e/1 ∼ 8FLC1.42E−08 * ** ∼2.29E−02 *3.07E−04 * **TemperatureExpressionThe decrease in FLC expression between 9 and 13 DAG: 29 °C < 22 °C.1453284990 ∼ 35858183.95 ∼ 19.780.51 ∼ 3.36a ∼ c/1 ∼ 3,6 ∼ 8MIR156D1.65E−06 * ** ∼5.93E−05 * **3.59E−04 * **Heat stressExpressionmiR156 isoforms are highly induced after heat stress, and miR156 may avoid flowering during a hot spell.1553873896 ∼ 39118234.39 ∼ 4.700.70 ∼ 0.96a/1,5MIR156E2.62E−04 * ** ∼5.67E−04 * **1.46E−02 *Heat stressExpressionmiR156 isoforms are highly induced after heat stress, and miR156 may avoid flowering during a hot spell.1658508968 ∼ 86098574.96 ∼ 31.150.71 ∼ 4.91a ∼ b/2 ∼ 4,6FPF11.13E−08 * ** ∼1.46E−07 * **3.22E−06 * ** ∼5.35E−05 * **High temperatureExpression levelUnder high temperature, the level of plant endogenous hormone GA is inhibited, while the expression of FPF1 is increased by GA.17523373062 ∼ 234900844.94 ∼ 18.641.52 ∼ 2.26a ∼ b/2,5,6COL51.70E−05 * ** ∼2.36E−05 * **1.26E−02 *Low TemperatureThe activity of proteinsThe activity of COL proteins (COL1/2/3/5) is repressed by TOE proteins at low temperatures, which leads to the repression of FT expression.18526137944 ∼ 261379447.61 ∼ 12.051.64 ∼ 2.15a/2,5MAF41.36E−13 * ** ∼4.66E−11 * **2.19E−10 * ** ∼2.54E−09 * **Cold treatmentmRNA expression levelsIn the early stages of cold exposure, the mRNA expression levels of MAF4 are significantly increased.3VmrMLM-random, 3VmrMLM-fixed, 3VmrMLM-MEJA, MLM in GAPIT, and EMMAX were indicated by a to e, respectively; indirect indicators trait difference, regression intercept, regression coefficient, range, variance, standard deviation (SD), coefficient of variation (CV), and RC of flowering time on temperature (EF) were indicated by 1 to 8, respectively. The known GEIs in Supplementary Tables 12 and 14 were summarized in this table, where the QEIs were identified by one of methods or indicators, and their GEIs were confirmed in haplotype analysis using one- (indirect indicators) and two-way (environments and flowering time) ANOVA. P-value1 (indicator) and P-value2 (FT): The probabilities of haplotype analysis for indirect indicator in one-way ANOVA and for flowering time in two-way ANOVA, respectively. All the references were cited in Supplementary Table S12 and S14. Eighteen known gene-by-environment interactions (GEIs) around significant and suggested QTN-by-environment interactions (QEIs) for flowering time (FT) using 3VmrMLM-random, 3VmrMLM-fixed, 3VmrMLM-MEJA, EMMAX, and MLM. 3VmrMLM-random, 3VmrMLM-fixed, 3VmrMLM-MEJA, MLM in GAPIT, and EMMAX were indicated by a to e, respectively; indirect indicators trait difference, regression intercept, regression coefficient, range, variance, standard deviation (SD), coefficient of variation (CV), and RC of flowering time on temperature (EF) were indicated by 1 to 8, respectively. The known GEIs in Supplementary Tables 12 and 14 were summarized in this table, where the QEIs were identified by one of methods or indicators, and their GEIs were confirmed in haplotype analysis using one- (indirect indicators) and two-way (environments and flowering time) ANOVA. P-value1 (indicator) and P-value2 (FT): The probabilities of haplotype analysis for indirect indicator in one-way ANOVA and for flowering time in two-way ANOVA, respectively. All the references were cited in Supplementary Table S12 and S14. In summary, fourteen significant GEIs identified from indirect indicators were directly associated with temperature variation in previous studies, including CSTF77, NF-YC9, PIE1, SPA3, VRN1, GA1, TSF, JMJ14, MOS1, FLC, MIR156D, MIR156E, COL5, and MAF4, while GA3OX2, LIF2, BBX19, and FPF1 were found to be indirectly associated with temperature variation in previous studies (Fig. 4A; Table 3, S13 and S15). In addition, 39 candidate GEIs such as GRP2, FRI, SPL5, and TOE3 were mined (Tables S14 and S16). Manhattan plots of identifying QTN-by-environment interactions (QEIs) for flowering time in 199 Arabidopsis thaliana lines using the new tool IIIVmrMLM.QEI. (A) All the QEIs and their associated gene-by-environment interactions (GEIs) for flowering time and (B) QEIs identified only by multi-environment joint analysis (3VmrMLM-MEJA) and their associated GEIs. The left y axis represents the medians of −log10(P) values across all the approaches, which are obtained from single-marker genome-wide scanning in the first step, and the right y axis represents LOD scores, which are obtained from a likelihood ratio test for QEIs in the second step. The threshold of LOD score for suggested QEIs was 3.0 (red dashed line) in likelihood ratio test. If LOD score ≥ 15, the LOD scores are transformed as LOD=15+(LOD−15)/15. These LOD scores, along with their known GEIs, are shown in points with straight lines. All the GEIs, identified by one (pink), at least two indicators/methods (blue), and also detected by MLM/EMMAX (black), are marked with different colors. In addition, the GEIs with asterisk are around suggested QEIs. To demonstrate the second objective of the simulation studies, in which the extended 3VmrMLM-fixed is better than existing fixed-SNP-effect methods (MLM and EMMAX), all the eight indirect indicators were used as phenotypes to indirectly identify QEIs using EMMAX and MLM in the above FT datasets. In summary, 6 and 2 significant QEIs were identified by EMMAX and MLM, respectively (Table 2; file B: Tables B6-B7), and one and one known GEIs around significant QEIs from EMMAX and MLM were found to be truly associated with FT (Table 2, Table 3, S13 and S15; Fig. 4A). In addition, one candidate GEI was found around significant QEI from EMMAX (Tables S14 and S16). Clearly, the newly extended 3VmrMLM-fixed (97 QEIs and 12 known and 18 candidate GEIs) identifies many more QEIs and known/candidate GEIs than EMMAX (six QEIs and one known and one candidate GEIs) and MLM (two QEIs and one known GEIs) (Table 2). To further demonstrate the third objective of the simulation studies, in which 3VmrMLM-random and 3VmrMLM-fixed are better than 3VmrMLM-MEJA, all the above trait phenotypes under three temperatures were jointly used to associate with 216130 SNP markers using 3VmrMLM-MEJA. As a result, 22 significant and 7 suggested QEIs were identified (Table 2; file B: Table B1), and 19 (3) known FT genes were found to be located around significant (suggested) QEIs (Table 2). To confirm whether these known genes are interacted with temperature change, the above known FT genes were used to perform haplotype analysis by two-way ANOVA for flowering times. As a result, two and one genes were significant around significant and suggested QEIs, respectively (Table 2; Fig. S1; Tables S12-S13). All the three significant GEIs were found to be directly associated with temperature change in previous studies, including SPA3, MOS1, and MIR156D (Table 3; Fig. 4B). Therefore, these known genes were considered as known FT GEIs. In addition, the other 4 known FT genes were considered as candidate FT GEIs in the same way, including ARP4, SPL5, AGL24, and GRP2, pending further experimental validation to explore these novel GEI-FT associations (Table 2 and Table S14). Based on the above results, 3VmrMLM-random identified 141 QEIs along with 15 known and 26 candidate GEIs, while 3VmrMLM-fixed identified 97 QEIs as well as 12 known and 18 candidate GEIs. Therefore, both 3VmrMLM-random and 3VmrMLM-fixed identify substantially more QEIs and known/candidate GEIs than 3VmrMLM-MEJA (29 QEIs and 6 known and 4 candidate GEIs) (Table 2). More importantly, although 3VmrMLM-random can identify numerous QEIs and GEIs, it may still be incomplete, as some known FT GEIs, such as GA3OX2, PIE1, and GA1, were not identified by 3VmrMLM-random (Table 3; Tables S13 and S15). Therefore, the new tool, IIIVmrMLM.QEI, developed in this study, provides a more comprehensive approach to detecting QEIs in multi-environment GWAS. Discussion The genetic variations among individuals are derived from both recombinant in bi-parental segregation populations for linkage analysis and historical recombinant in association mapping populations for GWAS. Hence, the genetic basis for identifying genetic loci is recombinant for linkage analysis and linkage disequilibrium for GWAS, and recombinant-based linkage analysis has a higher repetition rate in detecting genetic loci across different models, indirect indicators, and environments than linkage disequilibrium-based GWAS [25], [26], [27], [28]. Therefore, it is essential to extend models and indirect indicators in GWAS in this study. In GWAS, many software packages were available to identify QTNs associated with complex traits [29]. Although 3VmrMLM-MEJA without random mating assumption was established to identify QTNs and QEIs with high power, the number of known and candidate GEIs was relatively limited in real data analysis, e.g., 3VmrMLM-MEJA (six known and four candidate GEIs) identified fewer known and candidate GEIs than IIIVmrMLM.QEI (18 known and 38 candidate GEIs) (Table 2, Table 3; Fig. 4; Fig. S1; Tables S13-S16). Meanwhile, 3VmrMLM-random-D (2), 3VmrMLM-random-VI (1), and 3VmrMLM-fixed-VI (1) in IIIVmrMLM.QEI can detect some method-specific known GEIs, and 5, 6, and 3 known GEIs were simultaneously identified by two, three, and more than four methods, respectively, demonstrating the repeatability and specificity between different methods (Fig. S2). The new tool provides a more comprehensive framework to identify QEIs and mines known and candidate GEIs, indicating the necessity of developing the new tool in this study. More importantly, multiple lines of evidence were provided to confirm the true GEI-phenotype causality in this study. In detail, thirty-four known FT genes were confirmed by haplotype analysis for indirect indicators, eighteen GEIs were further validated by two-way ANOVA (Tables S11 and S12), and all the eighteen significant GEIs were found to be directly (14) or indirectly (4) associated with temperature change in previous studies (Table 3). Significant advances in methods and applications have been achieved in this study. First, 3VmrMLM-random was extended to 3VmrMLM-fixed for the first time. Unlike in Li et al. [8], where these methods were used to identify QTNs, we employed them to indirectly detect QEIs. As a result, when compared with 3VmrMLM-random, 3VmrMLM-fixed can identify three method-specific GEIs in real data analysis. Furthermore, 3VmrMLM-fixed outperformed MLM and EMMAX in identifying QEIs in both real and simulated datasets. This improvement is attributed to the fact that additive and dominant effects and their polygenic backgrounds are included in our mixed model in this study. In other words, all the possible effects are detected and estimated conditional on controlling for all the possible polygenic backgrounds in this study, while only one confounding effect and its polygenic background are considered in MLM and EMMAX. This conclusion is consistent with the quantitative trait locus mapping in Zhou et al. [30]. Second, 3VmrMLM-random and 3VmrMLM-fixed were applied to analyze trait difference, RC, RI, and extended EF and VIs to indirectly identify QEIs for the first time. We also compared 3VmrMLM-random and 3VmrMLM-fixed with mrMLM in detecting QEIs via indirect indicators. As a result, 3VmrMLM-random (38.9 %, 81.0 %, 80.2 %, 54.7 %, 44.7 %, 54.6 %, and 57.7 %) and 3VmrMLM-fixed (36.6 %, 73.6 %, 73.0 %, 51.4 %, 37.8 %, 51.8 %, and 54.9 %) exhibited higher average powers for difference, RI, RC, range, variance, SD, and CV than mrMLM (23.5 %, 56.1 %, 48.0 %, 36.9 %, 31.9 %, 36.1 %, and 44.9 %) (Table S17), which were observed in Li et al. [8], [29], further confirming the above reason. In conclusion, combining diverse indirect indicators that capture phenotypic changes across environments with optimal QTN detection methods (3VmrMLM-random and 3VmrMLM-fixed) enhances the accuracy of indirect QEI detection. In our 3VmrMLM method, the nominal FPR at the genome-wide level was set at 0.05, and the P-value threshold of significant loci was set at 0.05/m based on the widely used Bonferroni correction. The Bonferroni correction is known to be too stringent in crop GWAS [25]. To mitigate the risk of missing important loci, an additional threshold for suggested loci was applied. In our previous multi-locus GWAS methodologies, the LOD score of 3.0 was deemed an effective threshold [26]. Therefore, this standard was also adopted in this study. Once we obtain more significant and suggested QEIs, multi-omics, bioinformatics, haplotype analysis, and ATAC-seq dataset analysis can be employed to mine candidate functional genes. In this case, we focus on the QEIs with some important candidate functional genes, while all the significant and suggested loci contribute to enhancing the accuracy of genomic selection [31], [32], [33]. When there are more environments, more general indicators of environmental variation are more appropriate. In the real data analysis, 2 known and 15 candidate GEIs were exclusively identified by the four environmental VIs (Table 3, S13 and S16). Therefore, in this study, the trait difference was extended to four environmental VIs in this study. Additionally, 3VmrMLM-random and 3VmrMLM-fixed were applied for the first time to analyze the four environmental VIs. Although VI-based methods were inferior to regression parameter-based methods in simulation studies, the former can identify some method-specific GEIs compared to the latter in real data analysis, e.g., 3VmrMLM-random-VI identified a specific GEI and 3VmrMLM-fixed-VI identified a specific GEI to be truly associated with flowering time (Table 3), indicating the fact that these new methods can be used to not only demonstrate the repeatability of QEI detection, but also to identify new QEIs. Hence, extending these indicators was deemed essential in this study. In this study, we also conducted regression analysis of Arabidopsis flowering time on environmental temperatures (10 ℃, 16 ℃, and 22 ℃), and the regression coefficients were used as phenotypes to identify QEIs. Environmental factors (EF), encompassing specific meteorological factors (e.g. temperature), treatment levels, and soil fertility levels, plays an important role in studying environmental plasticity and predicting crop yields. Compared with RI and RC, EF detected 2 known and 2 candidate method-specific GEIs (Tables S13 and S14), more importantly, the year/location environments are expanded to include meteorological factors, treatment levels, soil fertilizer content, and other environmental conditions, indicating the need to analyze EF by the three methods in a new tool. The newly extended 3VmrMLM-fixed and all the indirect indicators provide a robust theoretical framework for identifying QEIs. First, the statistical foundation of 3VmrMLM-fixed is detailed in the Materials and Methods section, with supporting evidence presented in the Results section. Second, trait differences and regression parameters serve as phenotypes to indirectly identify QEIs, as described in Korte et al. [1], Zan and Carlborg [3], Kerin and Marchini [6], Li et al. [7], and Fu and Wang [34]. Finally, although QTNs and QEIs were simulated in Monte Carlo simulation studies, only QEIs were detected by all the indirect indicators as phenotypes (Tables S1-S9), providing solid evidence for this method. In addition, we calculated correlation coefficients of previously proposed indirect indicators (trait difference, RI, and RC) with newly proposed indirect indicators (range, Var, SD, CV and EF). As a result, very significant correlations (Table S18) support the detection of QEIs by the new indirect indicators in this study. In almost all the existing GWAS methods, marker genotypes are coded as 0, 1, and 2, the effect to be detected is the allelic substitution effect, and the polygenic background to be controlled is also related to the allelic substitution effect. In this case, association mapping population is assumed to be randomly mated [35]. However, this assumption does not exist in crop association mapping populations. 3VmrMLM-related methods do not rely on this assumption. Thus, this tool based on 3VmrMLM-related methods is expected to be more applicable in future GWAS studies for animals, plants, and humans. In particular, the Multi_env module of software IIIVmrMLM [29] can be used to identify QEIs in human genetics, although there are no same individuals in different environments [8]. In addition, if different individuals of each mice family are measured under multiple environments for trait phenotypes, IIIVmrMLM.QEI in this study is feasible to identify QEIs in human genetics. Conclusion To comprehensively detect QEIs, trait phenotype difference, regression parameters (RI and RC), variation indicators (range, variance, standard deviation, and coefficient of variation), and environmental factors (EF) were first integrated with previous 3VmrMLM-random and extended 3VmrMLM-fixed to develop an effective tool IIIVmrMLM.QEI for QEI detection. The new tool (18), validated by Monte Carlo simulation studies, identified much more known GEIs for flowering time than multi-environment joint analysis (6). Funding This work was supported by the National Natural Science Foundation of China (32070557; 32200500; 32270673; 32470657), Huazhong Agricultural University Scientific & Technological Self-Innovation Foundation (2014RC020), Natural Science Foundation of Hubei Province (2022CFB780), and Postdoctoral Innovative Research Position of Hubei Province. We thank Mr. Hanwen Zhang (hywenzhang@henceedu.com; Hence Education Ltd., Vancouver, Canada) for improving the language within the manuscript. CRediT authorship contribution statement Yuan-Ming Zhang: Writing – review & editing, Supervision, Methodology. Chen Ying: Data curation. Mei Li: Data curation. Xue-Lian Han: Data curation. Ya-Wen Zhang: Writing – original draft, Software, Methodology. Conflict of interest We are pleased to submit our manuscript entitled “IIIVmrMLM.QEI: An effective tool for indirect detection of QTN-by-environment interactions in genome-wide association studies” to be considered for publication in Computational and Structural Biotechnology Journal. No conflict of interest exits in the submission of this manuscript, all authors acknowledge the content of this manuscript and consent to its publication. Declaration of Competing Interest All authors acknowledge the content of this manuscript and consent to its publication. Footnotes Supplementary data associated with this article can be found in the online version at doi:10.1016/j.csbj.2024.11.046. Contributor Information Ya-Wen Zhang, Email: yawen@ujs.edu.cn. Xue-Lian Han, Email: luluzi@webmail.hzau.edu.cn. Mei Li, Email: 1286985431@qq.com. Ying Chen, Email: chenying8861@163.com. Yuan-Ming Zhang, Email: soyzhang@mail.hzau.edu.cn. Supplementary material Supplementary material . Data Availability The new tool can be downloaded from https://github.com/YuanmingZhang65/IIIVmrMLM.QEI and has been presented in this manuscript as IIIVmrMLM.QEI_1.0.zip (R code) and IIIVmrMLM.QEI.zip (CLI). All the simulated and real datasets in this study are included in the IIIVmrMLM.QEI software and are also presented as Data Files S1 to S5 (real datasets) and S6 to S9 (simulated datasets). References Associated Data Supplementary Materials Supplementary material Data Availability Statement The new tool can be downloaded from https://github.com/YuanmingZhang65/IIIVmrMLM.QEI and has been presented in this manuscript as IIIVmrMLM.QEI_1.0.zip (R code) and IIIVmrMLM.QEI.zip (CLI). All the simulated and real datasets in this study are included in the IIIVmrMLM.QEI software and are also presented as Data Files S1 to S5 (real datasets) and S6 to S9 (simulated datasets).