# Source paper archive

- Source title: Integrating multi-omics data and machine learning in single-step GBLUP reaction norm models to improve prediction across untested environments in lodgepole pine
- Author(s): Cappa, E. P., Chen, C., Benowicz, A., Thomas, B. R., El-Kassaby, Y. A.
- Source landing URL: https://www.biorxiv.org/content/10.64898/2026.11.16.767158v1
- PDF URL: https://www.biorxiv.org/content/10.64898/2026.11.16.767158v1.full.pdf
- PDF SHA-256: a8b96ebf863a6e4b9e7aa7512628c50dce00fb875cc259ef1feac6cc4218dd01
- Converter: pymupdf4llm 0.0.27
- Limits: Automated PDF text extraction; reading order, equations, tables and formatting may be incomplete or incorrect. Images are omitted, and no OCR was performed. Markup and links are neutralized for safe downloading. Consult the original PDF for the authoritative version.

---

&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


1

# **Integrating multi-omics data and machine learning in single-step GBLUP** **reaction norm models to improve prediction across untested environments in** **lodgepole pine**


Eduardo P. Cappa &#91;1,2*&#93;, Charles Chen &#91;3,4*&#93;, Andy Benowicz &#91;5&#93;, Barb R. Thomas &#91;4&#93;, Yousry A. El-Kassaby &#91;6&#93;


1 Instituto Nacional de Tecnología Agropecuaria (INTA), Instituto de Recursos Biológicos, Centro de


Investigación en Recursos Naturales, De Los Reseros y Dr. Nicolás Repetto s/n, 1686, Hurlingham, Buenos


Aires, Argentina.


2 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Buenos Aires, Argentina.


3 Department of Biochemistry and Molecular Biology, Oklahoma State University, Stillwater, Oklahoma


74078, USA


4 Department of Renewable Resources, University of Alberta, 442 Earth Sciences Bldg., Edmonton, Alberta,


T6G 2E3, Canada.


5 Forest Stewardship and Trade Branch, Alberta Forestry and Parks, Edmonton, Alberta, T6H 5T6, Canada


6 Department of Forest and Conservation Sciences, Faculty of Forestry, University of British Columbia,


Vancouver, British Columbia, V6T 1Z4, Canada.


*** Corresponding authors**

Eduardo Pablo Cappa (ORCID: 0000-0002-6234-2263)
e-mail: &#91;cappa.eduardo@inta.gob.ar&#93;(mailto:cappa.eduardo@inta.gob.ar)


Charles Chen (ORCID: 0000-0002-2203-0433)
e-mail: charleschen@ualberta.ca


**Running title:** Multi-omics and machine learning prediction


**Keywords:** multi-omics integration; genotype-by-environment interaction; reaction norm models;
nonlinear kernels; genomic prediction


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


2


**Abstract**


The effectiveness of genomic prediction in forest trees is often constrained by genotype-by-environment


(G×E) interactions, particularly when predicting performance in environments not represented in the


training population. We evaluated the integration of genomic, enviromic (climatic and soil), and


metabolomic (targeted monoterpene) information within multi-environment single-step genomic best linear


unbiased prediction (ssGBLUP) reaction norm models to improve prediction in unobserved environments.


Models incorporated environmental covariates (W) and genotype-specific responses using linear and


nonlinear kernels, including Gaussian and Arc-Cosine formulations. The study examined 12 traits measured


on 11,124 trees across four lodgepole pine progeny trials, with 1,490 trees genotyped with 25,099 SNPs.


Predictive performance was assessed using leave-one-environment-out cross-validation. Additive genetic


correlations among environments revealed substantial G×E for early-growth traits and drought-related


decline, with correlations frequently below 0.65. Inclusion of environmental covariates reduced variance


attributed to site effects, indicating that these descriptors captured environmental heterogeneity.


Incorporating G×E and genotype-by-environmental covariate (G×W) interactions improved model


performance, increasing explained phenotypic variance by up to 29.58% and reducing residual variance by


up to 3.13% relative to baseline models. Across traits, G×E explained more phenotypic variation than G×W,


contributing on average 2.31% and 0.60%, respectively. Predictive ability was high across models (0.688

0.701), with gains of up to 7.85% when nonlinear kernels were combined with metabolomic covariates,


particularly for traits exhibiting stronger G×E. Monoterpene profiles yielded predictive performance


comparable to climate and soil descriptors, highlighting the potential of integrating genomic, enviromic,


and metabolomic information to improve prediction across environments and support climate-resilient


forest tree breeding and deployment.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


3


**Introduction**


Genomic prediction (GP) has emerged as a powerful tool to increase selection accuracy and accelerate


breeding cycles by incorporating genome-wide marker information into breeding programs (Meuwissen _et_


_al._, 2001). In forest tree species, GP is commonly implemented using genomic best linear unbiased


prediction (GBLUP), which relies on genomic relationship matrices derived from SNP markers (VanRaden,


2008), and has been successfully applied across a range of species (Grattapaglia, 2022; Isik, 2022). More


recently, single-step GBLUP (ssGBLUP) approaches have gained attention because they jointly integrate


pedigree, phenotypic, and genomic information within a unified framework (Legarra _et al._, 2009; Misztal


_et al._, 2009; Aguilar _et al._, 2010; Christensen and Lund, 2010). This feature is particularly attractive for


forest tree breeding programs, where only a subset of individuals is typically genotyped (e.g., (Ratcliffe _et_


_al._, 2017; Cappa _et al._, 2019; Callister _et al._, 2021)).


Despite these advances, the predictive ability of GP models can be compromised when genotype

by-environment (G×E) effects are strong (Resende _et al._, 2012a; Beaulieu _et al._, 2014; Grattapaglia, 2014;


Duarte _et al._, 2024), as marker effects may not be consistently expressed across environments (Cappa _et_


_al._, 2025). Multi-environment models have therefore been developed to account for G×E, either by


modeling genetic correlations among environments or through reaction norm approaches. Traditional multi

trait models treat the same trait measured across environments as distinct traits and quantify G×E through


genetic correlations among sites (Falconer and Mackay, 1996). Although effective in forest trees (e.g., (Li


_et al._, 2007; Belaber _et al._, 2018; Cappa, Chen, _et al._, 2022)), these models become increasingly complex


as the number of environments increases, even when more parsimonious structures such as factor-analytic


models are used (Cullis _et al._, 2014). Moreover, they cannot predict performance in untested environments


(Smith _et al._, 2015).


Reaction norm models provide an alternative framework by incorporating environmental covariates


(ECs) to characterize environmental variation. Within this context, the enviromics framework uses large

scale environmental data (Costa-Neto and Fritsche-Neto, 2021) to construct environmental relationship


matrices and model genotype-specific responses across environmental gradients (Jarquín _et al._, 2014).


Building on this concept, (Pérez-Rodríguez _et al._, 2017) incorporated pedigree and genomic information


within a ssGBLUP reaction norm model to predict wheat performance across international environments,


demonstrating the potential of ssGBLUP models integrating G×E interactions for prediction in untested


environments. While this formulation, using pedigree or genomic information, or both, has been widely


applied in agricultural crops (Pérez-Rodríguez _et al._, 2015; Sukumaran _et al._, 2017; Morais Júnior _et al._,


2018), apple (Jung _et al._, 2025), alfalfa (Filho _et al._, 2023) and animal breeding (Ribeiro _et al._, 2015; Tiezzi


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


4


_et al._, 2017; McWhorter _et al._, 2023; Bussiman _et al._, 2025), its application in forest trees remains limited


(Ratcliffe _et al._, 2019; Oliveira Silva da Costa _et al._, 2025).


The increasing availability of high-resolution environmental datasets has further expanded


&#91;enviromics approaches. Climate information from sources such as ClimateNA (https://climatena.ca/) and&#93;(https://climatena.ca/)


NASA POWER &#91;(https://power.larc.nasa.gov/), together with edaphic information from the Global Soil&#93;(https://power.larc.nasa.gov/)


Information System (GLOSIS, &#91;https://data.apps.fao.org/glosis/), can be used to construct environmental&#93;(https://data.apps.fao.org/glosis/)


similarity matrices (Costa-Neto _et al._, 2021; Fritsche-Neto, 2024). Beyond these traditional descriptors,


emerging -omics technologies offer new opportunities to characterize biological responses to environmental


variation and improve the modeling of G×E interactions.


Among the available -omics layers, metabolomics is particularly promising because metabolites


are downstream products of gene expression and are therefore closely associated with phenotypic variation


(Guijas _et al._, 2018). Metabolite profiles integrate both genetic and environmental influences, making them


valuable biomarkers of environmental adaptation and G×E interactions (Wang et al., 2023). Unlike climatic


and soil descriptors, metabolomic variables may also capture integrated physiological responses associated


with environmental adaptation and stress (Yoshioka _et al._, 2026). Targeted metabolomics approaches, which


quantify specific metabolites or metabolic pathways (Patti _et al._, 2012), provide a powerful framework to


investigate biologically meaningful processes underlying plant responses. In conifers, for instance,


constitutive monoterpene profiles have been linked to resistance against insect-pathogen complexes and


may provide useful predictors for defense-related traits (Ullah _et al._, 2021). Consequently, metabolomic


variables may serve not only as intermediate phenotypes but also as biologically informed descriptors of


environmental similarity within reaction norm frameworks.


Multi-omics prediction integrates multiple sources of biological information to capture


complementary processes underlying complex traits (see Figure 1 in (Robinson et al., 2025)). By


incorporating intermediate phenotypes and regulatory mechanisms linking genotype to phenotype, these


approaches can improve both predictive performance and biological interpretability (Xu _et al._, 2022;


Montesinos-López _et al._, 2025; Yoshioka _et al._, 2026). However, identifying which omics variables are


biologically meaningful and effectively improve prediction accuracy remains largely unexplored,


particularly in the context of G×E interactions (Yoshioka _et al._, 2026).


Integrating multiple omics layers into GP presents an important analytical challenge because these


datasets differ substantially in dimensionality, scale, and structure (Picard _et al._, 2021; Montesinos-López


_et al._, 2025). Moreover, the relationship within and among genomic, environmental, and other omics layers


are often nonlinear and difficult to capture using conventional linear models. Machine learning (ML) and


kernel-based approaches provide flexible alternatives for modeling such complex relationships (Yan and


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


5


Wang, 2023; Crossa _et al._, 2025). While linear kernels form the basis of traditional GBLUP models,


nonlinear kernels such as Gaussian and arc-cosine functions can capture more complex patterns, including


gene-gene interactions (Alves _et al._, 2019; Cappa _et al._, 2026) and nonlinear G×E interactions (De Los


Campos _et al._, 2010; Cuevas _et al._, 2016, 2017; Bandeira e Sousa _et al._, 2017; Crossa _et al._, 2019; Costa

Neto _et al._, 2020).


Recent studies in crops have demonstrated that integrating genomic, metabolomic, transcriptomic,


and phenomic information can improve prediction accuracy when combined with advanced ML approaches


(Wu _et al._, 2024; Amin _et al._, 2025; Montesinos-López _et al._, 2025; Mohamedikbal _et al._, 2025). However,


relatively few studies have evaluated multi-omics prediction in multi-environment settings (Hu _et al._,


2021), despite growing recognition that robust multi-omics frameworks should be validated across diverse


environments and incorporate G×E interactions (Montesinos-López _et al._, 2025; Yoshioka _et al._, 2026).


Moreover, applications of these approaches in forest tree breeding remain limited (Wang _et al._, 2023). To


our knowledge, no previous study has jointly integrated genomic, enviromic, and metabolomic information


within multi-environment ssGBLUP reaction norm models using both linear and nonlinear kernels in forest


tree breeding. In this study, we conducted a comprehensive evaluation of multi-environment ssGBLUP


reaction norm models integrating genomic, enviromic (i.e., climatic and edaphic), and targeted


metabolomic information to predict 12 productivity- and climate-adaptation-related traits measured across


four open-pollinated lodgepole pine progeny trials. Specifically, our objectives were to: (i) assess the


relative contribution of environmental relationship matrices derived from different omics covariates, and


their interactions with genotypes, modeled using alternative linear and nonlinear kernels, in explaining


phenotypic variation across environments; and (ii) evaluate the predictive ability of these ssGBLUP


reaction norm models for prediction in untested (new) environments under variable climatic conditions.


**Material and Methods**


**Genetic material and trial description**


Genetic material and field trials used in this study were previously described by (Cappa, Chen, _et al._, 2022).


Briefly, the genetic material originated from four open-pollinated (OP) lodgepole pine progeny test trials


established under Alberta Forestry and Park’s Region C lodgepole pine breeding program (FGRMS, 2016)


and currently managed by Blue Ridge Lumber. The study population was comprised of 224 OP families


derived from five natural stands (provenances); Deer Mountain, Inverness River, Judy Creek, Swan Hills,


and Virginia Hills, all located within the Region C breeding region. These progeny tests were established


at four sites in Alberta: Judy Creek (JUDY; 54°26'42.4'' N, 115°34'10.7'' W, 1097 m elevation), Virginia


Hills (VIRG; 54°28'26.2'' N, 115°50'42.1'' W, 1127 m elevation), Swan Hills (SWAN; and 54°42'43.9'' N,


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


6


115°29'54.0 W, 1033 m elevation), and Timeau (TIME; 54°43'49.6'' N, 115°17'57.6'' W, 1064 m elevation).


All trials followed a “set in replications” experimental design consisting of five replications (blocks), 21


sets per replication, and 4-tree row plots planted at a spacing of 2.5 × 2.5 m. All trial sites were fenced and


surrounded by a single buffer row of lodgepole pine seedlings planted around the perimeter to minimize


edge effects. Additional information on site characteristics and experimental design is provided in (Cappa,


Ratcliffe, _et al._, 2022) (Table 1), while the geographic distribution of the test sites across Alberta, Canada,


is illustrated in (Cappa, Chen, _et al._, 2022) (Supplementary Information Fig. S2).


**Phenotypic data**


The four progeny trials were assessed for two growth productivity traits, diameter at breast height (DBH,


measured at 1.3 m above ground) and total tree height (HT). DBH were measured at ages 23, 30, and 35,


whereas HT was measured at ages 9, 23, 30, and 35. At age 35, DBH and HT were assessed only in trees


selected for genomic sampling (see Genomic Data subsection below). Wood density (WD) was assessed at


age 35 years using 5-mm increment cores extracted at DBH (for details see (Cappa, Chen, _et al._, 2022)).


Resistance to _Endocronartium harknessii_ (J.P. Moore) Y. Hirats (western gall rust, WGR) was


evaluated at age 36 using a qualitative seven-class scoring system ranging from the absence of gall


symptoms (class 1) to dead trees (class 7). Because classes 3, 5, and 7 occurred infrequently across the trial


sites, they were merged with classes 2, 4, and 6, respectively, resulting in a four-class resistance scale. These


four WGR classes were subsequently transformed into normal score (NSWGR) following the method


described by (Gianola and Norton, 1981).


The dendrochronological growth decline index (Decline) was derived from tree-ring measurements


obtained from the same increment cores used for wood quality assessment. This index quantifies long-term


growth reduction associated with repeated drought events by comparing the mean basal area increment


(BAI) during a pre-drought reference period (1997-2001) with the mean BAI during the final five years of


the study period (2012-2016). Larger decline values indicate greater growth reductions. Further details


regarding the calculation and interpretation of this index are provided in (Sebastian-Azcona _et al._, 2025).


The two outer slabs remaining after pneumatic processing of the increment cores for wood density


determination, were subsequently used to characterize lifetime variation in stable carbon isotope


composition (C13). In addition, the carbon fraction (FC), defined as the proportion of carbon relative to


total dry biomass, was measured during elemental analysis. A detailed description of sample preparation


and analytical procedures is provided by (Cappa, Klutsch, _et al._, 2022).


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


7


Table S1 summarizes all evaluated traits, including the number of observations and descriptive


statistics for each trait on their original scale (i.e., prior to standardization). The distribution of the twelve


traits across the four environments is illustrated by Figure 1 using boxplots.


**Genomic data**


As described by (Cappa, Ratcliffe, _et al._, 2022), 40 OP families were selected from the original population


of 224 families, to represent the full range of height variation at age 30, including families with high,


intermediate, and low growth performance. Approximately 10 trees per OP family were sampled at each of


the four sites, yielding a total of 1,600 sampled individuals (see (Cappa, Ratcliffe, _et al._, 2022) for details).


Additionally, 35 candidate forward-selected trees, previously identified based on their height breeding


values (BVs) at age 30, and representing an additional 28 OP families, were included in the sequencing


efforts. Consequently, a total of 1,635 trees from 68 OP families were selected and needles sampled for


genotyping.


Of the 1,635 sampled trees, 1,554 DNA samples passed quality-control procedures after extraction


and were subsequently genotyped using the genotyping-by-sequencing (GBS) platform (Chen _et al._, 2013)


following the protocol described by (Cappa, Ratcliffe, _et al._, 2022). After filtering for a minor allele


frequency (MAF) ≥ 1% and allowing up to 30% missing genotype data per marker, the final dataset was


comprised of 1,490 trees genotyped at 25,099 (25K) single nucleotide polymorphisms (SNPs). Missing


genotypes were imputed using the mean allele frequency method described by (Ratcliffe _et al._, 2015).


Because a reference genome assembly was unavailable for lodgepole pine at the time oof analysis, SNP


calling and genotyping were conducted using the reference-free UNEAK pipeline.


**Enviromic data**


Climatic environmental covariates were obtained using the get_weather() function from the R package


envRtype (Costa-Neto _et al._, 2021). Fourteen climatic variables (Table 1) were retrieved from NASA


POWER satellite-derived datasets using the geographic coordinates of each trial site. Climate records


covering the period from plantation establishment (between May 31 and June 18, 1982) (see Table 1 in


(Cappa, Ratcliffe, _et al._, 2022)) to the final measurement dates (July 12 to 20, 2017). To better represent


conditions influencing tree growth, only data corresponding to the active growing season (April 15 to


October 15) were retained (Chhin _et al._, 2008). The resulting 35-year climate dataset was used to


characterize historical environmental variability across sites and to approximate the range of climate


conditions that trees may encounter in future deployment environments (Gillberg _et al._, 2019).


Soil environmental covariates were obtained for each site using the get_soil() function from the R


package SoilType (Fritsche-Neto, 2024). This package also uses the geographic coordinates of the trial sites


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


8


to retrieve soil information from global databases, particularly the World Soil Information Service (WoSIS).


Based on the WoSIS, seven physical and chemical soil covariates (Table 1) were obtained for the locations


closest to each experimental site.


**Targeted metabolomic data**


Secondary metabolites, primarily monoterpenes associated with tree defense, were identified and quantified


from phloem tissue collected between May and July 2017 across the four lodgepole pine test sites. A


detailed description of the sampling, extraction, and analytical procedures are described in (Cappa, Chen,


_et al._, 2022). In total, 15 hexane-extractable monoterpenes were identified. From these, seven compounds


(Table 1) were selected for further analysis, based on their relatively high abundance in the phloem and


their documented role in lodgepole pine defense against _Dendroctonus ponderosae_ Hopkins (mountain pine


beetle) (Erbilgin, 2019). These compounds were used as targeted monoterpene covariates in subsequent


analyses.


Although monoterpenes are metabolomic traits, they were incorporated into the reaction norm


framework as biologically informed descriptors of environmental similarity. As outlined in the Introduction,


metabolite profiles reflect the combined influence of genetic and environmental conditions. Consequently,


targeted monoterpenes may capture adaptive and stress-related processes that are not fully represented by


conventional climatic and soil variables. Thus, these metabolite profiles provide offering complementary


information for characterizing environmental heterogeneity and genotype-by-environment interactions.


**Quantitative genetics analyses**


_**Pedigree-based multi-environment mixed model**_


To obtain heritability estimates and additive genetic correlations across sites, a classical pedigree-based


multi-environment mixed model (ABLUP) was fitted for each trait as follows:


𝒚= 𝛃𝑿 + 𝒁 ! 𝒓 + 𝒁 " 𝒂 +𝒆 &#91;1&#93;


where, 𝒚 is the vector of individual tree phenotypes; 𝛃 is the vector of fixed effects of genetic group nested


within site formed according to the five provenances; 𝒓 is the vector of random replicate effect nested within


site; 𝒂 is the random vector of additive genetic effects ( _i.e._, breeding values) distributed as 𝒂~𝑵(𝟎, 𝜮 𝒂 ⨂𝑨),


where 𝜮 𝒂 is the multi-environment genetic effects (co)variance matrix and 𝑨-matrix is the average


numerator relationship matrix derived from the pedigree information (Henderson, 1984). Finally, 𝒆 is the


vector of random residual distributed as 𝒆~𝑵(𝟎, 𝑹 𝟎 ⨂𝑰), where 𝑹 𝟎 is the residual (co)variance matrix for


the four environments (sites) with dimension 4 × 4, and 𝑰 denotes the identity matrix, while the symbol ⨂


stands for the Kronecker operator. We assumed an unstructured (co)variance matrix for the genetic effects


(𝜮 𝒂 ). However, given that the sites were assessed separately, the residual covariance across sites is assumed


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


9


to be zero. The matrices 𝑿, 𝒁 !, and 𝒁 " relate the observation to the means of the site effects in 𝛃, the


replicate effects in 𝒓, and the additive genetic effects for each tree in 𝒂.


Individual narrow-sense heritability estimates for each environment _i_ (ℎ &#91;6&#93; %&amp; ) and genetic correlations


between environment _i_ and _j_ (𝑟̂ " !" ), were estimated as follows:



&amp; &#91;;  𝑟̂&#93; &#91;"&#93; !" &#91;=&#93;



ℎ &#91;6&#93; %&amp; =



𝜎: " !



&amp; ! + 𝜎: '&amp; !



𝜎: "&amp; !



&amp;

" !,!
=𝜎:



&amp; !,! × 𝜎: "&amp; ","



𝜎: " !,"



&amp;



&amp;" ! is the estimated variance for the additive genetic effects, and σ@ &amp;' !



where, σ@ " !



&amp; is the estimated residual error



&amp; !,! and 𝜎: "&amp; ","



for environment _i_ . 𝜎: " !,!



&amp; are the estimated additive genetic variances in the environment _i_ and _j_,



respectively, and 𝜎: " !," is the estimated additive genetic covariance between environments _i_ and _j_ .


Variance components, narrow-sense heritability estimates and genetic correlations between sites


from model &#91;1&#93; (and their standard errors), were estimated using average information restricted maximum


likelihood (AI-REML) with the ASReml-R package (version 4.2.0) (Butler _et al._, 2023).


_**Genomic-pedigree-based multi-environment ssGBLUP predictive mixed models**_


To assess the impact of multi-omic covariates and their interactions with the environment on the proportion


of variance explaining and genomic prediction, five hierarchical ssGBLUP prediction models were fitted


following the (Jarquín _et al._, 2014) approach.


Assuming that phenotypic data were collected from _q_ trees evaluated across _r_ sites. Trees may have


different means depending on their membership in different genetic groups, and replicates were nested


within sites. Let _i_ ( _i_ = 1, 2, …, _r_ ) denote the subscript for site or environment, _j_ ( _j_ = 1, 2, …, _q_ ) the subscript


for tree, _k_ ( _k_ = 1, 2, …, _g_ ) the corresponding genetic group, and _l_ ( _l_ = 1, 2, …, _s_ ) the replicate within site.


Let 𝑦 %() and 𝑒 %() denote the observed phenotype and the residual effect of individual _j_ evaluated at site _i_,


and replicate _l_, respectively. The first ssGBLUP model included a genetic group (β - ) as fixed effects,


together with random environment effects (𝐸 % ), random replicate within site effects (𝑅 )(%) ), and additive


genetic value (𝑎 ( ). Then, the first ssGBLUP mixed model is defined as:


Model 1 (M1, E + G)


𝑦 %() = β              - + 𝐸 % + 𝑅 )(%) + 𝑎 ( + 𝑒 %() &#91;2&#93;


where the assumed distributions are: 𝑬 ~ 𝑁(𝟎, 𝑰 𝜎 -&amp; ), 𝑹 ~ 𝑁(𝟎, 𝑰 𝜎 !&amp; ), 𝒂 ~ 𝑁(𝟎, 𝑯 𝜎 "&amp; ), and


&amp;

𝒆 ~ 𝑁(𝟎, 𝑹 𝟎 ⨂𝑰). The variance components 𝜎 ! &#91;&amp;&#93;, 𝜎 -, and 𝜎 "&amp;, represent the variances associated with each


effect, assumed to be common across environments. Matrices 𝑹 𝟎 and 𝑰 were defined previously, and 𝑯


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


10


corresponds to the combined pedigree-genomic relationship matrix. The inverse of the relationship matrix


that combines combined pedigree- and marker information (𝑯 &#91;.𝟏&#93; ) was derived by (Legarra _et al._, 2009;


Aguilar _et al._, 2010; Christensen _et al._, 2012), and calculated for each _**G**_ -matrix as:


𝟎
𝑯 &#91;.𝟏&#93; = 𝑨 &#91;.𝟏&#93; + J𝟎 &#91;𝟎&#93; 𝝀(𝑮 &#91;.𝟏&#93; −𝑨 .𝟏𝟐𝟐 ) &#91;N&#93;


where 𝝀 scales the differences between genomic and pedigree-based information, 𝑮 &#91;.𝟏&#93; is the inverse of the


genomic-based relationship matrices, and 𝑨 .𝟏𝟐𝟐 is the inverse of the pedigree-based relationship matrix for


the genotyped individuals only (the subscript 2 denotes genotyped individuals). The weighting factor λ was


set to 0.95 for all ssGBLUP models. The _**G**_ -matrix was estimated following the first method proposed by


(VanRaden, 2008):


𝑴𝑴´
𝑮=
2 ∑𝑝                 -                 - (1 − 𝑝                 - )


where, 𝑴 is the centered matrix of SNP covariates, and 𝑝   - is the current (or observed) allele frequency of


the genotyped trees for marker _k_ . The _**G**_ -matrix was scaled to have the same diagonal and off-diagonal


averages as the corresponding _**A**_ _**22**_ -matrix, according to (Christensen _et al._, 2012) (eq. 4). Finally, in the


model &#91;2&#93;, the fixed-effect vector 𝛃 represents the systematic effects associated with the _k_ genetic groups.


Model 2 (M2, E + W + G)


𝑦 %() = β             - + 𝐸 % + 𝑅 )(%) + 𝑤 %( + 𝑎 ( + 𝑒 %() &#91;3&#93;


Model M2 extends Model M1 (Equation &#91;2&#93;) by incorporating a new random effect associated with


environmental covariates, including climatic, soil, or targeted monoterpene covariables (𝒘). This effect is


assumed to follow 𝒘 ~ 𝑁(𝟎, 𝛀 𝜎 1 &#91;&amp;&#93; ), where 𝛀 represents the environmental kernel, i.e., the relationship


matrix among environments (sites) derived from climatic, soil, or monoterpene information using linear


and nonlinear kernel functions (see below), and 𝜎 1 &#91;&amp;&#93; is the variance associated with the environmental


covariates.


Model 3 (M3, E + G + W + G×W)


𝑦 %() = β             - + 𝐸 % + 𝑅 )(%) + 𝑤 %( + 𝑎 ( + 𝑎𝑤 %( + 𝑒 %() &#91;4&#93;


Model M3 incorporates the random interaction vector between additive genetic effects ( _**a**_ ) and


environmental covariates (𝒘). This effect is modeled as 𝒂𝒘 ~ 𝑁(𝟎, 𝑯∘𝒁 𝑬 𝛀𝒁 𝑬 𝜎 "1&amp; ), where 𝜎 "1&amp; is the


variance associated with the genotype × environmental covariate interaction, 𝒁 𝑬 is the incidence matrix for


site (or environment) effects, and ∘ denotes the Hadamard product of matrices.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


11


Model 4 (M4, E + G + W + G×E)


𝑦 %() = β             - + 𝐸 % + 𝑅 )(%) + 𝑤 %( + 𝑎 ( + 𝑎𝐸 %( + 𝑒 %() &#91;5&#93;


Model M4 incorporates the random interaction effect between additive genetic effects (𝒂) and the



&amp; &amp;

"- ), where 𝜎 "


evaluation sites (𝑬): 𝒂𝑬 ~ 𝑁(𝟎, 𝑯∘𝒁 𝑬 𝒁 𝑬 𝜎 "


"-&amp; is the variance associated with the genotype ×



enviroment (G×E) interaction.


Model 5 (M5, E + G + W + G×E + G×W)


𝑦 %() = β           - + 𝐸 % + 𝑅 )(%) + 𝑤 %( + 𝑎 ( + 𝑎𝑤 %( + 𝑎𝐸 %( + 𝑒 %() &#91;6&#93;


Finally, Model M5 simultaneously includes both interaction effects: genotype × environment (G×E)


(Equation &#91;5&#93;) and genotype × environmental covariates (G×W) (Equation &#91;4&#93;).


Table 2 summarizes the five multi-environment ssGBLUP models evaluated across the 12 traits,


including the covariates, kernel types, model abbreviations, and corresponding model formulations. In total,


37 model-covariate-kernel combinations were assessed.


_**Linear and nonlinear environmental and targeted metabolomic kernels**_


Climatic and soil environmental data, together with targeted metabolomic covariates (monoterpenes), were


used to estimate the environmental similarity matrix among environments (𝛀). In this study, three methods


were evaluated to estimate environmental similarity kernels. As a reference, a linear kernel was computed


from the environmental covariate matrix 𝑾, which included climatic, soil, and monoterpene variables. In


this case, 𝑾 corresponds to a centered and standardized matrix of covariates with dimensions equal to the


number of sites ( _r_ ) by the number of covariates ( _m_ ) ( _m_ = 14 for climatic variables and _m_ = 7 for soil and


targeted monoterpenes) (Table 1). The linear relationship matrix (Linear) was then calculated as:


𝑾𝑾´
Linear =
𝒕𝒓𝒂𝒄𝒆(𝑾𝑾´)⁄𝒏𝒓𝒐𝒘(𝑾)


Additionally, two nonlinear kernels from machine learning were evaluated to capture more complex


relationships among environments: the Gaussian kernel (GK) (Gianola _et al._, 2006; Gianola and Van Kaam,


2008) and the arc-cosine kernel (AK) (Cuevas _et al._, 2019).


The GK estimates similarity between environments based on the Euclidean distance between their


covariate vectors, allowing the model to account for nonlinear patterns that cannot be captured by linear


kernels. A bandwidth parameter ( _h_ ) controls the rate at which similarity decays with increasing distance


between environments. In this study, four different _h_ values (0.1, 0.25, 0.5, and 1) were initially evaluated,


and the final analyses were conducted using _h_ = 1.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


12


The Linear and GK relationship kernels were calculated using the W_matrix() function from the R


package envRtype (Costa-Neto _et al._, 2021).


The AK was also evaluated following the approach proposed by (Cuevas _et al._, 2019), originally


introduced by (Cho and Saul, 2009). This kernel measures similarity between environments based on the


angle between their environmental covariate vectors, allowing the model to capture complex nonlinear and


hierarchical relationships. The arc-cosine kernel can be interpreted as an approximation to computations


performed by deep neural networks, enabling flexible representations of interactions among environmental


variables. In this study, the AK was computed from the environmental covariate matrix 𝑾 using a single


layer (layer = 1), as implemented by (Cuevas _et al._, 2019).


Following to (Prado _et al._, 2024), the environmental similarity structures (𝛀) generated by the linear


and nonlinear kernels were compared with the pattern of the pedigree-based genetic correlations among


sites for each trait. For each covariate source (climatic, soil, and targeted monoterpene data), pairwise


environmental similarities were calculated, and their distributions were visualized using density plots.


_**Assessment of the multi-environment ssGBLUP models**_


First, models M1-M5 (Table 2) were fitted using the complete dataset available for each trait (Table S1).


This analysis enabled the estimation of variance components (proportion of variance) and the calculation


of goodness-of-fit statistics for model comparison, including the log-likelihood, and Akaike Information


Criterion (AIC).


Second, to evaluate the predictive ability of the five proposed ssGBLUP models, a Leave-One

Environment-Out cross-validation scheme was implemented. The objective was to predict the performance


of genotypes already observed in a novel (unobserved) climate-variable environment. To mimic that


situation, each environment (i.e., site) was treated in turn as a future environment. At each iteration, the


training population included phenotypic observations from three environments, which were used to predict


the performance of genotypes in the excluded environment (validation). This procedure was repeated across


all four environments.


Predictive ability was quantified as the Pearson correlation ( _r_ ) between the predicted genetic values


and the corresponding observed values in the omitted environment.


**Results**


**Pedigree-based heritability estimates and additive genetic correlations across sites**


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


13


Pedigree-based heritability estimates varied across traits and sites (Table 3). Growth traits showed a clear


age-dependent pattern, with moderate heritability estimates at early ages (e.g., HT09: 0.19-0.38; DBH23:


0.18-0.29) that increased substantially at later ages (e.g., HT35: 0.72-0.87; DBH35: 0.33-0.61). Among


traits, HT35, C13, and WD exhibited the highest average heritability estimates (0.80, 0.59, and 0.46,


respectively), whereas DBH at younger ages and Decline showed lower values, with Decline presenting


particularly low and variable estimates across sites (0.00-0.41). Across environments, SWAN and JUDY


tended to show slightly higher heritability estimates (0.40 and 0.42, respectively) on average compared to


VIRG and TIME. It should be noted that HT35 and DBH35 were assessed only in the genomic sampling


subset; therefore, the relatively high heritability estimates observed for these traits should be interpreted in


light of this sampling design.


Genetic correlations among sites were positive but varied considerably depending on the trait


(Figure 2). Early-age growth traits showed the lowest and most variable correlations across environments


(HT09: 0.25-0.71, mean = 0.50; HT23: 0.55-0.70, mean = 0.65; HT30: 0.50-0.75, mean = 0.56), indicating


a stronger genotype-by-environment (G × E) interaction at younger ages. Similarly, Decline exhibited


relatively low, and highly variable, correlations (0.19-1.00; mean = 0.52), further supporting substantial G


× E effects for this trait. In contrast, later-age traits such as HT35 (0.70-1.00; mean = 0.90) and DBH35


(0.52-1.00; mean = 0.81) showed higher and more consistent correlations across sites, indicating greater


stability of genetic effects over time. However, both traits were assessed only in the genomic sampling


subset at age 35. Intermediate patterns were observed for DBH23 (0.49-0.83; mean = 0.70) and wood


density (WD: 0.26-1.00; mean = 0.76), suggesting more moderate levels of G × E interaction. Notably,


lower correlations were more frequently observed between TIME and JUDY or VIRG (the more distant


sites) (e.g., HT09: 0.25 and 0.34; Decline: 0.19 and 0.44), whereas correlations between TIME and SWAN


were generally higher (e.g., HT09: 0.71; Decline: 0.86).


**Covariate patterns across sites**


The 𝑾 matrices show differences in the clustering patterns of environments across the three sets of


covariates (Figure 3). For climatic variables (A), TIME was separated from the other sites, showing


contrasting patterns for most temperature- and precipitation-related covariates, while SWAN, VIRG, and


JUDY clustered more closely, indicating more similar climatic conditions among them. In contrast, soil


covariates (B) showed a different clustering pattern, where VIRG and JUDY grouped together, suggesting


similar soil composition, whereas SWAN and TIME appeared more differentiated from both each other and


the VIRG–JUDY group. For targeted monoterpenes (C), TIME again showed a markedly different profile,


with consistently similar values across all compounds, separating from the remaining sites. Overall, these


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


14


results highlight that environmental similarity among sites depends on the type of covariates considered,


with TIME consistently representing the most distinct environment.


**Environmental similarity patterns across covariates and kernel methods**


The environmental similarity patterns across sites varied depending on both the type of covariates and the


kernel used (Figure 4). For climatic covariates (A), the linear kernel showed moderate similarities among


SWAN, VIRG, and JUDY, with TIME remaining clearly distinct, a pattern that was further accentuated


under the nonlinear kernels (GK and AK), which increased the contrast among environments. For soil


covariates (B), the linear kernel suggested closer similarity between VIRG and JUDY, while SWAN and


TIME were more differentiated; this structure was partially preserved under GK, although similarities


became more polarized, whereas AK highlighted sharper contrasts across all sites. In the case of targeted


monoterpenes (C), TIME consistently appeared as the most dissimilar environment across all kernels,


although the strength of the pairwise dissimilarities differed among kernels. In summary, the nonlinear


kernels produced clustering patterns that differed from those obtained with the linear kernel and, in some


cases, highlighted environmental contrasts more clearly. Nevertheless, the overall relationships among sites


were broadly consistent across kernel approaches.


The environmental kernels derived from climatic, soil, and targeted monoterpene covariates


generally displayed broader similarity distributions than the G×E interaction estimated from the pedigree

based additive genetic correlations (Figure S1). This indicates that environmental covariates capture a wider


spectrum of environmental heterogeneity among sites, whereas genetic correlations tend to be concentrated


within a narrower range of similarities. Consequently, the G×E structure inferred from the genetic


correlations may only partially reflect the environmental contrasts present across these multi-environment


lodgepole pine trials, potentially masking part of the underlying environmental heterogeneity.


**Model fit of multi-environment ssGBLUP models across traits**


Complete log-likelihood (logL) and Akaike Information Criterion (AIC) values for all evaluated traits,


models, kernel functions, and covariate types are presented in Table S2. Overall, the effect of including


interaction terms on model fit was trait-dependent (Table 4). Models incorporating interactions yielded


lower AIC values for several traits, particularly early growth and Decline-related traits, whereas the baseline


model (M1: E + G) remained competitive or achieved the lowest AIC for others. In particular, the inclusion


of G×E interaction (M4) produced the largest reduction in AIC for early growth traits (HT09, HT23, HT30,


DBH23) and for Decline. Notably, these traits also exhibited lower to moderate genetic correlations among


sites (Figure 2), consistent with a stronger influence of a G×E interaction. In contrast, traits with higher and


more stable genetic correlations across sites (e.g., HT35 and DBH35) showed little to no improvement,


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


15


with negligible AIC differences among models. The addition of genotype-by-environmental covariate


interaction (G×W; M3) led to moderate improvements in some traits, but the full model (M5) did not


consistently outperform M4, indicating limited additional improvement in model fit from including both


interaction terms simultaneously. For traits with stronger G×E effects (HT09, HT23, HT30), M5 showed


AIC values very similar to M4, although slightly higher, indicating no additional improvement over M4.


**Estimates of variance components across the multi-environment ssGBLUP models**


_**Impact of covariates and kernels on the variance attributable to environment**_


Under the M1 (E + G) model, the main environmental (site) effect accounted for 0.00% to 16.59% of the


phenotypic variance (Table S2). The largest site effects were observed for HT09 (16.59%), C13 (9.48%),


and HT35 (6.97%), whereas no detectable site effect was found for NSWGR.


The inclusion of covariates (W; M2: E + G + W) led to a moderate but consistent reduction in the


magnitude of the main site effect compared with the baseline M1 (E + G) model. Across traits, the average


site effect decreased from 5.00% under M1 to 3.97% under M2 (-1.03 percentage points (20.48%)),


indicating that part of the environmental variance previously attributed to site was captured by explicit


covariates. On average, site effects were lower when soil (3.40%) or climate (4.05%) covariates were


included than when monoterpene covariates were fitted (4.48%).


The impact of W on the variance attributable to environment was strongly trait-dependent. For


instance, in HT09 the site effect dropped from 15.32% (M1) to 0.00% when soil covariates were included,


suggesting a near-complete absorption of site differences by soil descriptors. A similar pattern was observed


for DBH23, where the site effect was also reduced to 0.00% under soil covariates using nonlinear kernels.


In contrast, for traits such as HT30 and DBH30, site variance remained virtually unchanged across all


covariates and kernels, indicating that the evaluated covariates had limited influence in explaining site


effects. For WD, climate covariates led to a complete reduction of the site effect (0.00%), whereas soil


covariates retained values similar to M1, highlighting that the relevance of environmental descriptors


depends on both the trait and the covariate considered.


When further examining site effects, differences among kernels (linear, GK, and AK) were


generally modest compared with the impact of including W, although some patterns emerged (Table S2).


The linear kernel tended to produce results similar to M1 for several traits, suggesting limited ability to


capture additional environmental structure. In contrast, nonlinear kernels occasionally captured more


complex patterns, leading to a slight reduction (e.g., DBH23 under soil) depending on the trait-covariate


combination.


_**Reduction of residual variance**_


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


16


Overall, the inclusion of environmental covariates (W) and G×W interactions had limited effects


on reducing residual variance (Figure 5 and Table 5). Models including W alone (M2) showed only minor


and inconsistent changes relative to the baseline model, regardless of covariate type or kernel. Incorporating


G×W interactions (M3) led to greater reductions in residual variance, with the largest improvements


observed under the GK kernel (1.70% - 2.78%), especially when soil covariates were fitted. However,


reductions under the linear and AK kernels were generally less than 1.5%. Taken together, these results


suggest that environmental covariates and reaction-norm terms explained only a small fraction of the


residual variation not accounted for by the baseline model.


When modeling the G×E interaction (M4), reductions in residual variance became more


pronounced across all covariates. The linear kernels showed consistent moderate reduction (-2.25% to 

2.42%). In contrast, nonlinear kernels provided slightly stronger reductions, particularly for the GK kernel


(-2.76% to -3.11%), while the AK kernel showed similar but slightly smaller effects (-1.30% to -2.28%).


These results indicate that a portion of the remaining unexplained variation is associated with G×E


interactions not fully captured by W or G×W alone, although differences among kernels remain moderate.


When both interaction terms were included (M5: E + G + W + G×E + G×W), reductions in residual


variance remained consistent across covariates and kernels but did not substantially exceed those observed


for M4. The linear kernel showed stable reductions ranging from -2.25% to -2.46%, while the GK kernel


again provided the largest effects (-2.75% to -3.13%), particularly for soil covariates. The AK kernel


showed slightly smaller but comparable reductions (-1.30% to -2.28%). Across covariates, values were


relatively homogeneous, ranging from -1.98% to -2.82% for climate, -1.30% to -3.13% for soil, and -2.28%


to -2.75% for monoterpenes.


In summary, averaging across traits, differences among kernels were evident but depended on


model complexity. Nonlinear approaches, particularly GK, generally outperformed the linear kernel,


capturing a larger proportion of residual variance, whereas AK showed more variable performance and did


not consistently improve over the linear approach in the most complex models. Moreover, the relative


performance of covariates was stable, with soil variables generally providing the greatest reduction,


followed by climate and monoterpenes.


_**Difference in explaining sources of variation**_


Consistent with this expectation, Figure 6 (and Table S3) shows that across the 12 traits, the additive


genetic component (G) consistently explained the largest proportion of variance, averaging 8.14%,


exceeding the combined environmental contribution of site and replication within site (4.10% and 0.58%,


respectively). In the baseline model (M1; E + G), genetic variance accounted for 8.37% out of a total


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


17


explained variance of 13.92%, indicating that the genomic component was the largest modeled source of


variance. The inclusion of covariates in M2 resulted in modest changes in total explained variance (12.80

14.31%), with W explaining between 0.13% and 1.95% depending on the covariate type and kernel.


Nonlinear kernels, particularly Gaussian (GK), consistently captured a greater proportion of variance. When


genotype-by-covariate (G×W) interactions were included (M3), the G×W component explained an


additional proportion of variance (mean = 0.60%, up to 2.94%), accompanied by a slight reduction in the


additive genetic component (e.g., down to 7.15%). As previously mentioned, this suggests a redistribution


of variance from G to interaction effects, indicating that part of the additive genetic signal is being captured


by the G×W interaction component. Total explained variance increased up to 16.31%. The inclusion of G×E


interaction (M4) resulted in an appreciable contribution of G×E (mean = 2.31%), increasing total explained


variance to 15.04-16.60%. A slight reduction in G (8.02-8.13%) further indicates that part of the genetic


variance is environment-dependent. Finally, the full model (M5), incorporating both G×E and G×W


interactions, yielded the highest total explained variance (up to 16.61%). G×E remained the dominant


interaction term, while G×W contributed more modestly (generally ≤0.50%). Nonlinear kernels,


particularly GK, showed superior performance, explaining a slightly larger proportion of the variance


(15.69%) than linear kernels (15.01%). Overall, total explained variance generally increased with model


complexity from M1 to M5.


This general increase in explained variance with model complexity was also evident at the trait


level, although the relative importance of individual variance components varied among traits (Figure 7 and


Table S2). Although G remained predominant, its importance differed substantially, ranging from high


values in C13 (13.96-14.32%) and NSWGR (18.49-18.93%) to much lower contributions in Decline (mean


= 3.01%) and early growth traits such as HT09 (2.44-4.99%). Height traits showed moderate G


contributions in HT23 (5.79-7.25%) and HT30 (4.61-6.56%), increasing in HT35 (12.95-13.09%), while


WD exhibited consistently high genetic contributions (9.05-11.77%). Environmental effects were relevant


in early growth (HT09, up to 16.59%), remained moderate in HT23-HT35 (0.00-6.97%), and were


negligible in NSWGR. The contribution of covariates (W) also varied widely: soil covariates explained


substantial variance in HT09 (up to 20.60%) and DBH23 (up to 4.39%), while remaining negligible for


NSWGR and HT30; climate covariates showed moderate effects in HT23 (up to 5.91%) and WD (up to


6.30%), while being minimal in HT30 and absent in NSWGR; monoterpenes showed generally limited and


trait-dependent effects. Interaction components also showed strong heterogeneity. G×E was particularly


relevant for Decline (up to 6.34%) and consistently important across height traits (HT09, HT23, HT30:


2.83-4.17%), as well as for WD (2.79-3.96%), but negligible in HT35 and NSWGR. Similarly, G×W


interactions were trait-dependent, with notable contributions in Decline (up to 8.59%), WD (up to 6.02%),


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


18


and HT09-HT30 (up to 4.21 and 4.95%, respectively), while remaining close to zero in HT35, NSWGR,


and C13.


**Predictive ability in untested environments**


The Leave-One-Environment-Out cross-validation scheme revealed that the predictive performance of


multi-environment ssGBLUP models in untested environments was generally high across traits and


environments, with average predictive ability ranging from 0.688 to 0.701, depending on the model


structure, covariate type, and kernel formulation (Figure 8 and Table S4). These results indicate that


substantial predictive information could be transferred across environments, even under the baseline model


(M1), suggesting moderate environmental connectivity among sites. Overall, predictive ability tended to


increase with model complexity, although the magnitude of improvement varied considerably among


environments, covariate types, and kernel formulations.


The incorporation of enviromic and targeted metabolomic covariates together with G×E and G×W


interaction structures generally improved predictive ability relative to the baseline model, although the


magnitude of improvement depended strongly on the environmental sensitivity of the trait and the target


environment. Across all evaluated traits (Table 6 and Table S5), gains were moderate, typically ranging


between 1 and 3% depending on the model and kernel. The largest improvements were consistently


observed under models M4 and M5 (up to 3.65%), indicating that explicitly modeling genotype-by

environment heterogeneity improved the prediction of genotype performance in unobserved environments.


Most of this gain appeared to be driven by the G×E component, whereas the additional contribution of


G×W was comparatively small once G×E was included in the model. For example, under AK formulations


with monoterpene covariates, predictive ability at SWAN increased from 0.653 under M1 to values close


to 0.677 under M5. Overall, these results indicate that the benefits of increasing model complexity were


generally positive but highly dependent on trait, environment, covariate type, and kernel formulation.


Differences among covariate types were generally modest. Climatic and soil covariates produced


relatively stable responses across kernels and environments, whereas targeted metabolomic covariates


generated more variable responses, particularly under nonlinear kernels. Although linear kernels generally


provided the highest and most stable average predictive performance across all traits and environments


(Figure 9 and Table S6), nonlinear kernels occasionally produced larger gains in specific scenarios (Figure


8 and Table S4).


Moreover, the advantages of incorporating covariates and interaction structures became


substantially more evident for traits exhibiting stronger G×E patterns (average pedigree-based genetic


correlations across site pairs &lt; 0.65, Figure 2), namely HT09, HT23, HT30, and Decline (Table 7 and Table


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


19


S7). For these traits, predictive ability improvements reached up to 5.81% under linear kernels, 5.94% under


GK, and 7.85% under AK. For instance, under AK formulations with monoterpene covariates, predictive


ability at SWAN increased from 0.594 under M1 to approximately 0.640 under M5. These gains were


particularly evident in environmentally contrasting sites such as SWAN, where baseline predictive ability


was comparatively low. In contrast, environments with already high baseline predictive performance


showed comparatively smaller responses to model complexity.


Overall, these results indicate that the incorporation of enviromic and targeted metabolomic


covariates within multi-environment ssGBLUP models can improve predictive performance in untested


environments, particularly for traits exhibiting stronger G×E interaction patterns. The linear kernel was the


strongest general-purpose model, while GK and AK functioned as context-dependent alternatives,


suggesting that their utility may depend on the degree of nonlinear genotype-by-environment relationships


underlying the trait.


**Discussion**


This empirical study evaluated, for the first time in a forest tree species, the integration of genomic,


enviromic (climatic and soil), and metabolomic (targeted monoterpenes) information within multi

environment ssGBLUP reaction norm models using both linear and nonlinear kernels. Three major findings


emerged. First, the benefits of incorporating environmental information depended strongly on the


magnitude of G×E interaction, with the largest improvements observed for traits exhibiting lower genetic


correlations among environments. Second, environmental covariates improved environmental


characterization and facilitated prediction in untested environments, although most gains were attributed to


the explicit modeling of G×E rather than to environmental covariates alone. Third, targeted monoterpene


profiles achieved predictive performance comparable to that of climatic and soil covariates, suggesting that


targeted metabolomic information can serve as a biologically meaningful descriptor of environmental


similarity.


Previous studies evaluating genomic prediction (GP) across environments in forest trees have


reported contrasting results. These results range from substantial declines in predictive ability when


predictions are transferred to environmentally distinct sites (Resende _et al._, 2012b; Gamal El-Dien _et al._,


2015; Shalizi _et al._, 2022) to relatively stable performance when G×E interaction is limited (Beaulieu _et al._,


2014; Lenz _et al._, 2017), and even to improved prediction when G×E relationships are explicitly modeled


through multi-environment kernels (Hu _et al._, 2023). This study demonstrates that the benefit of


incorporating environmental information depends strongly on the magnitude of G×E interaction. Traits


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


20


exhibiting lower genetic correlations among environments (&lt;0.65; i.e., moderate G×E interaction; Figure


S2), including HT09, HT23, HT30, and Decline, showed the largest improvements in model fit (Table 4),


explained variance (Figure 7 and Table S2), and predictive ability (Figure 8 and Table S4) following the


incorporation of G×E and environmental covariates. In contrast, traits characterized by high genetic


correlations across sites, such as HT35 and DBH35, exhibited only marginal gains. These findings highlight


that the value of reaction norm models increases as environmental sensitivity becomes more pronounced,


providing greater insight into the environmental and biological factors underlying G×E interactions.


Historically, G×E in forest tree breeding has been modeled within multi-trait frameworks, in which


environments are treated as discrete entities and genetic correlations are estimated between pairs of sites


(Falconer and Mackay, 1996). While effective, this approach simplifies environmental variation into a


limited set of site categories. From the reaction norm perspective, genotype performance is instead viewed


as a continuous function across environmental gradients, with each environment representing a point along


a continuum rather than a discrete category (Kirkpatrick and Heckman, 1989). By incorporating


environmental covariates, reaction norm models characterize environments through continuous descriptors,


allowing information to be shared across environmental gradients and facilitating prediction in


environments that have not been directly observed (Resende _et al._, 2021). The relatively high predictive


abilities obtained in the leave-one-environment-out validation suggest that substantial predictive


performance could be transferred across sites, supporting this framework for prediction beyond directly


observed trial environments. This capacity is particularly relevant as tree deployment environments may


increasingly depart from the conditions represented in current trials.


The variance partitioning results further support this interpretation. Although the inclusion of


environmental covariates reduced the proportion of variance attributed to the main site effect, indicating


that climatic, soil, and targeted metabolomic descriptors captured part of the environmental heterogeneity


among trials, the environmental similarity kernels generally exhibited broader distributions than the realized


G×E correlation patterns observed among environments (Figure S1). This indicates that similarity in


measured environmental conditions does not necessarily translate directly into similarity in genetic


responses across sites. Nevertheless, despite providing an explicit characterization of site differences, G×W


interactions accounted for less phenotypic variation than G×E interactions, suggesting that the measured


covariates did not fully represent the environmental dimensions underlying genotype-specific responses


among sites. This interpretation is consistent with the limitation noted by (Jarquín _et al._, 2014) and other


authors (e.g., (Mumford _et al._, 2023)) for reaction norm models, namely that measured environmental


covariates may not fully describe environmental differences among sites because some relevant


environmental factors may not have been measured. Such a limitation could explain why G×E remained


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


21


the dominant source of interaction variance in our analyses and produced the greatest reductions in residual


variance (Figure 5 and Table 5). Moreover, the most complex models (M4 and M5) consistently achieved


the highest levels of explained variance (Table 5). This shift in variance structure was most evident for


environmentally sensitive traits (Figure 2; genetic correlation &lt; 0.65), which showed a redistribution of


explained variation away from additive genetic effects and toward G×E, G×W, site effects, and


environmental covariates (Table S2). Despite the increased importance of environmental and interaction


components for these traits, additive genetic effects remained the largest source of explained variation


across most trait-model combinations (Figure 6 and Table S2).


Beyond these broad differences in variance partitioning, the specific environmental factors


underlying site variation also differed among traits (Figure 7 and Table S2). For example, soil covariates


almost completely absorbed the site effect for HT09 under both linear and nonlinear kernels, suggesting


that differences in early growth among sites were predominantly driven by edaphic conditions rather than


by macro-climatic variation. This pattern is consistent with the importance of soil properties and resource


availability during the establishment phase, when juvenile trees are sensitive to local micro-environmental


conditions (e.g., (Henneb _et al._, 2020)). Similar patterns have been reported in _Pinus radiata_ D. Don (Li _et_


_al._, 2015), where soil characteristics were associated with genetic and site-related variation at ages 1 and 6.


Interestingly, no significant associations with these soil factors were detected at age 8, suggesting that the


environmental drivers of genetic and site variation may change as trees develop. In contrast, climatic


covariates completely absorbed the site effect for WD, suggesting that differences in wood density among


sites were primarily driven by climatic conditions. The influence of climate on wood density has been


extensively documented in forest tree species (Thomas _et al._, 2004; Nabais _et al._, 2018), with temperature


and moisture conditions playing a key role in shaping wood formation and resulting wood properties. For


instance, (Cabon _et al._, 2020) showed that temperature and water potential jointly regulate cambial activity


and xylem development in conifers, thereby influencing wood properties across contrasting environments.


Importantly, these differences in variance partitioning were also reflected in model predictive performance


(Table 6 and Table 7). Across traits, particularly those exhibiting moderate to strong G×E interactions


(Figure 2), the models accounting for a greater proportion of phenotypic variance generally achieved higher


predictive ability (Figure S2), especially when additional variation was captured through G×E and G×W


interaction terms. This finding suggests that improvements in predictive performance were associated with


the representation of additional sources of genetic and environmental heterogeneity relevant to trait


expression.


An interesting finding was that targeted monoterpene covariates achieved predictive performance


comparable to climatic and soil descriptors, despite representing a fundamentally different source of


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


22


information (Tables 6 and 7). Whereas climate and soil covariates characterize external environmental


conditions, monoterpene profiles represent physiological and biochemical phenotypes that may integrate


genetic effects with responses to the conditions experienced by individual trees. Their comparable


predictive performance suggests that targeted monoterpene profiles contain information relevant to cross

environment prediction and may provide a biologically mediated complement to conventional


environmental descriptors. Collectively, these findings suggest that targeted metabolomics can serve not


only as an additional omics layer within genomic prediction frameworks, as demonstrated by (Guo _et al._,


2023), but also as a complementary representation of the physiological state associated with environmental


and genetic differences among trees. These findings are consistent with recent studies highlighting the value


of metabolomic information as an intermediate layer linking genotype and phenotype (Wang _et al._, 2023;


Yoshioka _et al._, 2026). Because metabolite profiles integrate both genetic and environmental influences,


they may capture biological variation that is not fully explained by additive genetic effects alone. In this


context, the comparable predictive performance of monoterpenes and conventional environmental


descriptors suggests that part of the environmental signal relevant for prediction may also be represented


through the physiological responses of trees. Within reaction norm models, such information may help


characterize genotype-specific responses along biologically relevant environmental gradients associated


with G×E. However, the generally small contribution of G×W interactions, when G×E was also fitted,


suggests that monoterpene profiles capture only a limited fraction of genotype-specific environmental


variation beyond that already represented by the broader G×E term.


Although nonlinear kernels, particularly GK, generally improved model fit (Table S2), reduced


residual variance, and explained a larger proportion of phenotypic variation (Figure 5 and Table 5), these


advantages did not translate into higher predictive ability. Linear kernels instead generally produced the


highest and most stable average predictive performance across traits and environments (Figure 9). This


highlights an important distinction between explanatory and predictive performance. While nonlinear


kernels may more effectively capture complex environmental and biological covariance structures


(Montesinos-López _et al._, 2021), the additional complexity introduced by these models does not necessarily


improve their ability to generalize to novel environments. Similar observations have been reported by


(Lourenço _et al._, 2024), who showed that increasing model complexity does not automatically translate into


superior prediction and that simpler linear approaches often remain competitive because of their robustness


and reduced tuning requirements. More broadly, accumulating evidence suggests that the relative


performance of machine-learning algorithms depends strongly on trait architecture and dataset


characteristics, making it unlikely that any single method will consistently outperform all others across


scenarios (Reinoso-Peláez _et al._, 2022; Lourenço _et al._, 2024).


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


23


The trait-dependent responses observed in our study are consistent with recent evidence suggesting


that additional omics layers may provide complementary information to genomic data in some cases,


although the magnitude of these benefits can vary substantially among traits (Montesinos-López _et al._,


2021). When genomic effects accounted for most of the explained variation, as observed for HT35 and C13


(Figure 7 and Table S2), the inclusion of more complex kernels and additional covariates yielded only


marginal improvements. In contrast, environmentally sensitive traits characterized by stronger G×E


interactions (Figure 2) benefited more from nonlinear kernels and metabolomic information (Table S5 and


Table S7). Notably, the combination of targeted monoterpene covariates with the AK kernel produced the


largest predictive gains, reaching up to 7.85% in environmentally contrasting sites such as SWAN. These


findings suggest that the complementary information provided by metabolomic covariates may arise, at


least in part, from their ability to capture environmental variation that is not fully represented by genomic


data alone, rather than from providing substantial additional genetic information, as suggested by (Guo _et_


_al._, 2025) in a metabolomic-genomic BLUP study. Such complementary environmental information may


contribute to improved prediction across contrasting environments for environmentally sensitive traits.


An important aspect of this study is that prediction was evaluated under a leave-one-environment

out framework, which closely mimics the practical challenge of predicting performance of genetically


connected material in environments not represented in the training population. Although future climatic


conditions cannot be observed directly, the use of long-term historical weather records allowed


environmental variability to be represented through continuous descriptors that capture both past


environmental variation and cyclical climatic patterns, thereby introducing a temporally resolved dimension


to GP (Howard and Lipka, 2025). Nevertheless, the relatively small number of environments evaluated in


this study highlights the need for validation across broader environmental gradients before these approaches


can be deployed for climate-adaptation strategies in forest tree breeding. Previous studies have reported


larger predictive gains when reaction norm models are trained across a wider range of environments,


suggesting that increasing both the number and diversity of test sites may further improve the


characterization of environmental relatedness and genotype-specific responses (Resende _et al._, 2021).


Overall, our results demonstrate that the benefits of integrating environmental information and


interaction terms into GP models depends strongly on the magnitude of the G×E interaction.


Environmentally responsive traits benefited substantially from the incorporation of environmental


descriptors and interaction terms, resulting in improved characterization of G×E patterns and enhanced


prediction in untested environments. The comparable performance of metabolomic, climatic, and soil


covariates further suggests that biologically informed omics layers can complement traditional enviromic


approaches. Collectively, these findings demonstrate that integrating genomic, enviromic, and metabolomic


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


24


information within multi-environment ssGBLUP reaction norm models provides a promising framework


for improving prediction across heterogeneous and unobserved environments and informing climate

resilient breeding and deployment policies in forest tree species.


**Data availability statement**


Genotyping-by-sequencing (GBS) raw reads used in this study have been deposited in NCBI SRA


BioProject - PRJNA715165: &#91;https://www.ncbi.nlm.nih.gov/bioproject/715165. Information on the&#93;(https://www.ncbi.nlm.nih.gov/bioproject/715165)


lodgepole pine trials, including pedigree and phenotypic data used and analyzed during the current study,


will be made available in a GitHub repository upon acceptance of the manuscript.


**Acknowledgements**


We would like to acknowledge the RES-FOR staff that collected and prepared the many lodgepole pine


samples for this research: Laura Vehring, Pablo Chung, Jillian Dyck, Sarah Suzuk, Kristie Bui, Chris Arbter,


Rob Johnstone, Jesse Shirton, Arial Eatherton, Calvin Jensen, and Michael Thomson. We also wish to thank


all the technicians that work in the Thomas´, Erbilgin´ and Mansfield´ labs. We also want to acknowledge


the administrative coordination work of Stacy Bergheim. Finally, we thank Dr. Diego Jarquin for his


insightful comments on the implementation and validation of reaction norm models.


**Study funding**


&#91;This work was funded by Genome Canada (https://www.genomecanada.ca/) RES-FOR ID 10207, grants&#93;(https://www.genomecanada.ca/)


&#91;RES0031330 to BRT, and 16R75036 to YAE; Genome Alberta (https://genomealberta.ca/)&#93;(https://genomealberta.ca/) RES-FOR ID:


LRF, RES0034657 to BRT; University of Alberta/Faculty ALES/Dept RR


&#91;(https://www.ualberta.ca/index.html) grant RES0034569 to BRT; Alberta Innovates – BioSolutions&#93;(https://www.ualberta.ca/index.html)


&#91;(https://albertainnovates.ca/) RES0028979 to BRT; Genome BC&#93;(https://albertainnovates.ca/) &#91;(https://www.genomebc.ca/) grants&#93;(https://www.genomebc.ca/)


&#91;16R75421 to YAE; Forest Resource Improvement Association of Alberta (FRIAA, https://friaa.ab.ca/)&#93;(https://friaa.ab.ca/)


&#91;grants RES0037021 and RES0036845 to BRT; National Science Foundation (NSF, tps://www.nsf.gov/)&#93;(http://www.nsf.gov/)


grants MRI-1531128, ACI-1548562, and ACI-1445606 to CC; The Extreme Science and Engineering


Discovery (XSEDE, &#91;https://xras.xsede.org/public/requests/29304-XSEDE-MCB180177)&#93;(https://xras.xsede.org/public/requests/29304-XSEDE-MCB180177) grant


MCB180177 to CC. The funding bodies played no role in the design of the study and collection, analysis,


and interpretation of data and in writing the manuscript.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


25


**Author contributions**


Eduardo P. Cappa: Conceptualization, Formal analysis, Investigation, Methodology, Software,


Visualization, Writing - Original Draft, Writing - Review and Editing. Charles Chen: Data Curation,


Resources, Writing - Review and Editing. Andy Benowicz: Data Curation, Resources, Funding acquisition,


Writing - Review and Editing. Barb R. Thomas: Project Administration, Resources, Supervision, Validation,


Writing - Review and Editing. Yousry A. El-Kassaby: Project Administration, Supervision, Validation,


Writing - Review and Editing.


**Conflicts of interest**


None declared.


**References**


Aguilar I, Misztal I, Johnson DL, Legarra A, Tsuruta S, Lawlor TJ (2010). Hot topic: a unified approach

to utilize phenotypic, full pedigree, and genomic information for genetic evaluation of Holstein final
score. _J Dairy Sci_ **93** : 743–52.


Alves FC, Granato ÍSC, Galli G, Lyra DH, Fritsche-Neto R, de los Campos G (2019). Bayesian analysis

and prediction of hybrid performance. _Plant Methods_ **15** : 14.


Amin A, Zaman W, Park S (2025). Harnessing multi-omics and predictive modeling for climate-resilient

crop breeding: from genomes to fields. _Genes (Basel)_ **16** : 809.


Bandeira e Sousa M, Cuevas J, de Oliveira Couto EG, Pérez-Rodríguez P, Jarquín D, Fritsche-Neto R, _et_

_al._ (2017). Genomic-enabled prediction in maize using kernel models with genotype × environment
interaction. _G3 (Bethesda)_ **7** : 1995–2014.


Beaulieu J, Doerksen TK, MacKay J, Rainville A, Bousquet J (2014). Genomic selection accuracies

within and between environments and small breeding groups in white spruce. _BMC Genomics_ **15** :

1048.


Belaber EC, Gauchat ME, Reis HD, Borralho NM, Cappa EP (2018). Genetic parameters for growth,

stem straightness, and branch quality for Pinus elliottii var. elliottii × Pinus caribaea var. hondurensis
F 1 hybrid in Argentina. _Forest Science_ **64** : 595–608.


Bussiman F, Lourenco D, Hidalgo J, Chen C-Y, Holl J, Misztal I, _et al._ (2025). Genotype-by-environment

interaction with high-dimensional environmental data: an example in pigs. _Genetics Selection_
_Evolution_ **57** : 28.


Butler DG, Cullis BR, Gilmour AR, Gogel BJ, Thompson R (2023). ASReml-R Reference Manual

Version 4.2. VSN International Ltd., Hemel Hempstead, HP2 4TP, UK. _Technical report_ .


Cabon A, Peters RL, Fonti P, Martínez-Vilalta J, De Cáceres M (2020). Temperature and water potential

co-limit stem cambial activity along a steep elevational gradient. _New Phytologist_ **226** : 1325–1340.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


26


Callister AN, Bradshaw BP, Elms S, Gillies RAW, Sasse JM, Brawner JT (2021). Single-step genomic

BLUP enables joint analysis of disconnected breeding programs: an example with Eucalyptus
globulus Labill. _G3 Genes|Genomes|Genetics_ **11** .


Cappa EP, Chen C, Klutsch JG, Sebastian-Azcona J, Ratcliffe B, Wei X, _et al._ (2022). Multiple-trait

analyses improved the accuracy of genomic prediction and the power of genome-wide association of
productivity and climate change-adaptive traits in lodgepole pine. _BMC Genomics_ **23** : 536.


Cappa EP, Chen C, Klutsch JG, Sebastian-Azcona J, Ratcliffe B, Wei X, _et al._ (2025). Revealing stable

SNPs and genomic prediction insights across environments enhance breeding strategies of
productivity, defense, and climate-adaptability traits in white spruce. _Heredity (Edinb)_ **134** : 186–

199.


Cappa EP, Klutsch JG, Benowicz A, Munilla S, Mansfield SD, Erbilgin N, _et al._ (2026). Bayesian

networks and structural equation models reveal genetic causal relationships between productivity,
defense, and climate-adaptability traits in interior lodgepole pine. _G3 Genes|Genomes|Genetics_ **16** :
jkaf308.


Cappa EP, Klutsch JG, Sebastian-Azcona J, Ratcliffe B, Wei X, Da Ros L, _et al._ (2022). Integrating

genomic information and productivity and climate-adaptability traits into a regional white spruce
breeding program. _PLoS One_ **17** : e0264549–e0264549.


Cappa EP, de Lima BM, da Silva-Junior OB, Garcia CC, Mansfield SD, Grattapaglia D (2019).

Improving genomic prediction of growth and wood traits in Eucalyptus using phenotypes from nongenotyped trees by single-step GBLUP. _Plant Science_ **284** : 9–15.


Cappa EP, Ratcliffe B, Chen C, Thomas BR, Liu Y, Klutsch J, _et al._ (2022). Improving lodgepole pine

genomic evaluation using spatial correlation structure and SNP selection with single-step GBLUP.
_Heredity (Edinb)_ **128** : 209–224.


Chen C, Mitchell SE, Elshire RJ, Buckler ES, El-Kassaby YA (2013). Mining conifers’ mega-genome

using rapid and efficient multiplexed high-throughput genotyping-by-sequencing (GBS) SNP
discovery platform. _Tree Genet Genomes_ **9** : 1537–1544.


Chhin S, Hogg EH (Ted), Lieffers VJ, Huang S (2008). Influences of climate on the radial growth of

lodgepole pine in Alberta. _Botany_ **86** : 167–178.


Cho Y, Saul LK (2009). Kernel methods for deep learning. In: _Proceedings of the 22nd International_

_Conference on Neural Information Processing Systems (NIPS)_,, pp 342–350.


Christensen OF, Lund MS (2010). Genomic prediction when some animals are not genotyped. _Genet Sel_

_Evol_ **42** : 2.


Christensen OF, Madsen P, Nielsen B, Ostersen T, Su G (2012). Single-step methods for genomic

evaluation in pigs. _Animal_ **6** : 1565–1571.


Costa-Neto G, Fritsche-Neto R (2021). Enviromics: bridging different sources of data, building one

framework. _Crop breeding and applied biotechnology_ **21** : e393521S12.


Costa-Neto G, Fritsche-Neto R, Crossa J (2020). Nonlinear kernels, dominance, and envirotyping data

increase the accuracy of genome-based prediction in multi-environment trials. _Heredity (Edinb)_ **126** :

92–106.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


27


Costa-Neto G, Galli G, Carvalho HF, Crossa J, Fritsche-Neto R (2021). EnvRtype: a software to interplay

enviromics and quantitative genomics in agriculture. _G3 Genes|Genomes|Genetics_ **11** : jkab040.


Crossa J, Martini JWR, Gianola D, Pérez-Rodríguez P, Jarquin D, Juliana P, _et al._ (2019). Deep kernel

and deep learning for genome-based prediction of single traits in multienvironment breeding trials.
_Front Genet_ **10** : 1168.


Crossa J, Montesinos-Lopez OA, Costa-Neto G, Vitale P, Martini JWR, Runcie D, _et al._ (2025). Machine

learning algorithms translate big data into predictive breeding accuracy. _Trends Plant Sci_ **30** : 167–

184.


Cuevas J, Crossa J, Montesinos-López OA, Burgueño J, Pérez-Rodríguez P, de los Campos G (2017).

Bayesian genomic prediction with genotype × environment interaction kernel models. _G3: Genes,_
_Genomes, Genetics_ **7** : 41–53.


Cuevas J, Crossa J, Soberanis V, Pérez-Elizalde S, Pérez-Rodríguez P, Campos G de los, _et al._ (2016).

Genomic Prediction of Genotype × Environment Interaction Kernel Regression Models. _Plant_
_Genome_ **9** : plantgenome2016.03.0024.


Cuevas J, Montesinos-López O, Juliana P, Guzmán C, Pérez-Rodríguez P, González-Bucio J, _et al._

(2019). Deep kernel for genomic and near infrared predictions in multi-environment breeding trials.
_G3 Genes|Genomes|Genetics_ **9** : 2913–2924.


Cullis BR, Jefferson P, Thompson R, Smith AB (2014). Factor analytic and reduced animal models for the

investigation of additive genotype-by-environment interaction in outcrossing plant species with
application to a Pinus radiata breeding programme. _Theoretical and Applied Genetics_ **127** : 2193–

2210.


Duarte D, Jurcic EJ, Dutour J, Villalba P V, Centurión C, Grattapaglia D, _et al._ (2024). Genomic selection

in forest trees comes to life: unraveling its potential in an advanced four-generation Eucalyptus
grandis population. _Front Plant Sci_ **15** .


Erbilgin N (2019). Phytochemicals as mediators for host range expansion of a native invasive forest insect

herbivore. _New Phytologist_ **221** : 1268–1278.


Falconer DS, Mackay TFC (1996). Introduction to Quantitative Genetics. 4th Longman.


FGRMS (2016). Alberta forest genetic resource management and conservation standards. Alberta

Agriculture and Forestry, Government of Alberta, Edmonton, Alberta. : 158 p.


Filho CCF, Andrade MHML, Nunes JAR, Jarquin DH, Rios EF (2023). Genomic prediction for complex

traits across multiples harvests in alfalfa (Medicago sativa L.) is enhanced by enviromics. _Plant_
_Genome_ **16** : e20306.


Fritsche-Neto R (2024). SoilType: An R package to interplay soil characterization in plant science. _Agron_


_J_ **116** : 848–854.


Gamal El-Dien O, Ratcliffe B, Klapste J, Chen C, Porth I, El-Kassaby YA, _et al._ (2015). Prediction

accuracies for growth and wood attributes of interior spruce in space using genotyping-bysequencing. _BMC Genomics_ **16** : 370.


Gianola D, Fernando RL, Stella A (2006). Genomic-assisted prediction of genetic value with

semiparametric procedures. _Genetics_ **173** : 1761–1776.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


28


Gianola D, Van Kaam JBCHM (2008). Reproducing kernel Hilbert spaces regression methods for

genomic assisted prediction of quantitative traits. _Genetics_ **178** : 2289–2303.


Gianola D, Norton HW (1981). Scaling threshold characters. _Genetics_ **99** : 357–364.


Gillberg J, Marttinen P, Mamitsuka H, Kaski S (2019). Modelling G × E with historical weather

information im- proves genomic prediction in new environments. _Bioinformatics_ **35** : 4045–4052.


Grattapaglia D (2014). Breeding Forest Trees by Genomic Selection: Current Progress and the Way

Forward. In: _Genomics of Plant Genetic Resources_, Springer Netherlands: Dordrecht, pp 651–682.


Grattapaglia D (2022). Twelve years into genomic selection in forest trees: climbing the slope of

enlightenment of marker assisted tree breeding. _Forests_ **13** : 1554.


Guijas C, Montenegro-Burke JR, Warth B, Spilker ME, Siuzdak G (2018). Metabolomics activity

screening for identifying metabolites that modulate phenotype. _Nat Biotechnol_ **36** : 316–320.


Guo X, Sarup P, Bay Nord A, Henryon M, Ostersen T, Christensen OF (2025). Metabolomic-genomic

prediction realizes small increases in accuracy of estimated breeding values for daily gain in pigs.
_Genetics Selection Evolution_ **57** : 24.


Guo X, Sarup P, Jahoor A, Jensen J, Christensen OF (2023). Metabolomic-genomic prediction can

improve prediction accuracy of breeding values for malting quality traits in barley. _Genetics_
_Selection Evolution_ **55** : 61.


Henderson CR (1984). _Applications of linear models in animal breeding_ . University of Guelph: Guelph.


Henneb M, Thiffault N, Valeria O (2020). Regional climate, edaphic conditions and establishment

substrates interact to influence initial growth of black spruce and jack pine planted in the boreal
forest. _Forests_ **11** : 139.


Howard R, Lipka AE (2025). Genomic selection and reproducibility: are complex models distracting us

from true scientific validity in the presence of genotype-by-environment interaction? _G3_
_Genes|Genomes|Genetics_ **15** : jkaf244.


Hu H, Campbell MT, Yeats TH, Zheng X, Runcie DE, Covarrubias-Pazaran G, _et al._ (2021). Multi-omics

prediction of oat agronomic and seed nutritional traits across environments and in distantly related
populations. _Theor Appl Genet_ **134** : 4043–4054.


Hu X, Carver BF, El-Kassaby YA, Zhu L, Chen C (2023). Weighted kernels improve multi-environment

genomic prediction. _Heredity (Edinb)_ **130** : 82–91.


Isik F (2022). Genomic prediction of complex traits in perennial plants: A case for forest trees - genomic

prediction of complex traits: methods and protocols. In: Ahmadi N, Bartholomé J (eds) Springer US:
New York, NY, pp 493–520.


Jarquín D, Crossa J, Lacaze X, Du Cheyron P, Daucourt J, Lorgeou J, _et al._ (2014). A reaction norm

model for genomic selection using high-dimensional genomic and environmental data. _Theoretical_
_and Applied Genetics_ **127** : 595–607.


Jung M, Quesada-Traver C, Roth M, Aranzana MJ, Muranty H, Rymenants M, _et al._ (2025). Integrative

multi-environmental genomic prediction in apple. _Hortic Res_ **12** : uhae319.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


29


Kirkpatrick M, Heckman N (1989). A quantitative genetic model for growth, shape, reaction norms, and

other infinite-dimensional characters. _J Math Biol_ **27** : 429–450.


Legarra A, Aguilar I, Misztal I (2009). A relationship matrix including full pedigree and genomic

information. _J Dairy Sci_ **92** : 4656–4663.


Lenz PRN, Beaulieu J, Mansfield SD, Clément S, Desponts M, Bousquet J (2017). Factors affecting the

accuracy of genomic selection for growth and wood quality traits in an advanced-breeding
population of black spruce (Picea mariana). _BMC Genomics_ **18** : 335.


Li Y, Dutkowski GW, Apiolaza LA, Pilbeam D, Costa e Silva J, Potts B (2007). The genetic architecture

of a Eucalyptus globulus full-sib breeding population in Australia. _Forest Genetics_ **12** : 167–179.


Li Y, Xue J, Clinton PW, Dungey HS (2015). Genetic parameters and clone by environment interactions

for growth and foliar nutrient concentrations in radiata pine on 14 widely diverse New Zealand sites.
_Tree Genet Genomes_ **11** : 10.


De Los Campos G, Gianola D, Rosa GJM, Weigel KA, Crossa J (2010). Semi-parametric genomic
enabled prediction of genetic values using reproducing kernel Hilbert spaces methods. _Genet Res_
_(Camb)_ **92** : 295–308.


Lourenço VM, Ogutu JO, Rodrigues RAP, Posekany A, Piepho H-P (2024). Genomic prediction using

machine learning: a comparison of the performance of regularized regression, ensemble, instancebased and deep learning methods on synthetic and empirical data. _BMC Genomics_ **25** : 152.


McWhorter TM, Sargolzaei M, Sattler CG, Utt MD, Tsuruta S, Misztal I, _et al._ (2023). Single-step

genomic predictions for heat tolerance of production yields in US Holsteins and Jerseys. _J Dairy Sci_

**106** : 7861–7879.


Meuwissen THE, Hayes BJ, Goddard ME (2001). Prediction of total genetic value using genome-wide

dense marker maps. _Genetics_ **157** : 1819–1829.


Misztal I, Legarra A, Aguilar I (2009). Computing procedures for genetic evaluation including

phenotypic, full pedigree, and genomic information. _J Dairy Sci_ **92** : 4648–4655.


Mohamedikbal S, Al-Mamun HA, Bestry MS, Batley J, Edwards D (2025). Integrating multi-omics and

machine learning for disease resistance prediction in legumes. _Theoretical and applied genetics_ **138** :

163.


Montesinos-López OA, Montesinos-López A, Mosqueda-González BA, Delgado-Enciso I, Chavira
Flores M, Crossa J, _et al._ (2025). Genomic prediction powered by multi-omics data. _Front Genet_ **16** :

1636438.


Montesinos-López OA, Montesinos-López A, Pérez-Rodríguez P, Barrón-López JA, Martini JWR,

Fajardo-Flores SB, _et al._ (2021). A review of deep learning applications for genomic selection. _BMC_
_Genomics_ **22** : 19.


Morais Júnior OP, Duarte JB, Breseghello F, Coelho ASG, Magalhães Júnior AM (2018). Single-step

reaction norm models for genomic prediction in multienvironment recurrent selection trials. _Crop_
_Sci_ **58** : 592–607.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


30


Mumford MH, Forknall CR, Rodriguez D, Eyre JX, Kelly AM (2023). Incorporating environmental

covariates to explore genotype × environment × management (G × E × M) interactions: A one-stage
predictive model. _Field Crops Res_ **304** : 109133.


Nabais C, Hansen JK, David-Schwartz R, Klisz M, López R, Rozenberg P (2018). The effect of climate

on wood density: What provenance trials tell us? _For Ecol Manage_ **408** : 148–156.


Oliveira Silva da Costa L, da Rocha GN, Jatzek VA, Ramos E, da Silva P, Lipka AE, _et al._ (2025).

Predicting and mapping Eucalyptus urophylla family performance using large-scale environmental
covariates for targeted deployment. _Available at SSRN 5289714_ .


Patti GJ, Yanes O, Siuzdak G (2012). Metabolomics: the apogee of the omics trilogy. _Nat Rev Mol Cell_

_Biol_ **13** : 263–269.


Pérez-Rodríguez P, Crossa J, Bondalapati K, De Meyer G, Pita F, Campos G de los (2015). A Pedigree
Based Reaction Norm Model for Prediction of Cotton Yield in Multienvironment Trials. _Crop Sci_

**55** : 1143–1151.


Pérez-Rodríguez P, Crossa J, Rutkoski J, Poland J, Singh R, Legarra A, _et al._ (2017). Single-step genomic

and pedigree genotype × environment interaction models for predicting wheat lines in international
environments. _Plant Genome_ **10** : 0.


Picard M, Scott-Boyer M-P, Bodein A, Périn O, Droit A (2021). Integration strategies of multi-omics data

for machine learning analysis. _Comput Struct Biotechnol J_ **19** : 3735–3746.


Prado M, Famoso A, Guidry K, Fritsche-Neto R (2024). Optimizing multi-environment trials in the

Southern US Rice belt via smart-climate-soil prediction-based models and economic importance.
_Front Plant Sci_ **15** : 1458701.


Ratcliffe B, El-Dien OG, Klápště J, Porth I, Chen C, Jaquish B, _et al._ (2015). A comparison of genomic

selection models across time in interior spruce (Picea engelmannii × glauca) using unordered SNP
imputation methods. _Heredity (Edinb)_ **115** : 547–555.


Ratcliffe B, Gamal El-Dien O, Cappa EP, Porth I, Klapste J, Chen C, _et al._ (2017). Single-Step BLUP

with Varying Genotyping Effort in Open-Pollinated Picea glauca. _G3: Genes|Genomes|Genetics_ .


Ratcliffe B, Thistlethwaite F, El-Dien OG, Cappa EP, Porth I, Klápště J, _et al._ (2019). Inter- and intra
generation genomic predictions for douglas-fir growth in unobserved environments. _bioRxiv_ :

540765.


Reinoso-Peláez EL, Gianola D, González-Recio O (2022). Genome-Enabled Prediction Methods Based

on Machine Learning. In: Ahmadi N, Bartholomé J (eds) Springer US: New York, NY, pp 189–218.


Resende MFR, Muñoz P, Acosta JJ, Peter GF, Davis JM, Grattapaglia D, _et al._ (2012a). Accelerating the

domestication of trees using genomic selection: Accuracy of prediction models across ages and
environments. _New Phytologist_ **193** : 617–624.


Resende MFR, Muñoz P, Acosta JJ, Peter GF, Davis JM, Grattapaglia D, _et al._ (2012b). Accelerating the

domestication of trees using genomic selection: Accuracy of prediction models across ages and
environments. _New Phytologist_ **193** : 617–624.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


31


Resende RT, Piepho H-P, Rosa GJM, Silva-Junior OB, e Silva FF, de Resende MD V, _et al._ (2021).

Enviromics in breeding: applications and perspectives on envirotypic-assisted selection. _Theoretical_
_and Applied Genetics_ **134** : 95–112.


Ribeiro S, Eler JP, Pedrosa VB, Rosa GJM, Ferraz JBS, Balieiro JCC (2015). Genotype×environment

interaction for weaning weight in Nellore cattle using reaction norm analysis. _Livest Sci_ **176** : 40–46.


Robinson H, Robles-Zazueta CA, Voss-Fels KP (2025). Accelerating perennial crop improvement via

multi-omics-based predictive breeding. _Plant Genome_ **18** : e70058.


Sebastian-Azcona J, Cappa EP, Da Ros L, Ratcliffe B, Chen C, Wei X, _et al._ (2025). Short- and long-term

growth response to multiple drought episodes: evidence of genetic adaptation in a conifer species.
_Ecol Evol_ **15** : e71398.


Shalizi MN, Payn KG, Isik F (2022). Genetic linkage between the training and selection sets impacts the

predictive ability of SNP markers in a cloned population of Pinus taeda L. _Tree Genet Genomes_ **18** :

4.


Smith AB, Ganesalingam A, Kuchel H, Cullis BR (2015). Factor analytic mixed models for the provision

of grower information from national crop variety testing programs. _Theoretical and Applied_
_Genetics_ **128** : 55–72.


Sukumaran S, Crossa J, Jarquin D, Lopes M, Reynolds MP (2017). Genomic prediction with pedigree and

genotype × environment interaction in spring wheat grown in south and west Asia, north Africa, and
Mexico. _G3 Genes|Genomes|Genetics_ **7** : 481–495.


Thomas DS, Montagu KD, Conroy JP (2004). Changes in wood density of Eucalyptus camaldulensis due

to temperature—the physiological link between water viscosity and wood anatomy. _For Ecol_
_Manage_ **193** : 157–165.


Tiezzi F, de los Campos G, Parker Gaddis KL, Maltecca C (2017). Genotype by environment (climate)

interaction improves genomic prediction for production traits in US Holstein cattle. _J Dairy Sci_ **100** :

2042–2056.


Ullah A, Klutsch JG, Erbilgin N (2021). Production of complementary defense metabolites reflects a co
evolutionary arms race between a host plant and a mutualistic bark beetle-fungal complex. _Plant_
_Cell Environ_ **44** : 3064–3077.


VanRaden PM (2008). Efficient methods to compute genomic predictions. _J Dairy Sci_ **91** : 4414–23.


Wang M, Li R, Zhao Q (2023). Multi-Omics Techniques in Genetic Studies and Breeding of Forest

Plants. _Forests_ **14** : 1196.


Wu C, Luo J, Xiao Y (2024). Multi-omics assists genomic prediction of maize yield with machine

learning approaches. _Mol Breed_ **44** : 14.


Xu Y, Zhang X, Li H, Zheng H, Zhang J, Olsen MS, _et al._ (2022). Smart breeding driven by big data,

artificial intelligence, and integrated genomic-enviromic prediction. _Mol Plant_ **15** : 1664–1695.


Yan J, Wang X (2023). Machine learning bridges omics sciences and plant breeding. _Trends Plant Sci_ **28** :


199–210.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


32


Yoshioka H, Morota G, Iwata H (2026). Reciprocal BLUP: A predictability-guided multi-omics

framework for plant phenotype prediction. _Plants_ **15** : 17.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


33


**Figure legends**


**Figure 1. Boxplot of standarized phenotipic data for the 12 traits studied in the lodgepole pine**


**population in four enviroment.** See text for trait and enviroment abbrevations.


**Figure 2. Estimated genetic correlations between sites for each trait from the multiple-site model**


**using pedigree- (** _**A**_ **-matrix) relationship matrices for the lodgepole pine population** . The dark red to


dark blue color of each individual cell above the diagonal reflects the strength of the genetic correlation.


Abbreviations used for the traits and sites are described in the text.


**Figure 3. Heta maps of the climatic (A), soil (B) and targeted monoterpenes (C) covariates across the**


**four lodgepole pine environments.** Covariate abbreviations are provided in Table 1, and site abbreviations


are described in the text.


**Figure 4. Heat maps of the environment similarity matrix among the four lodgepole pine**


**environments considering climatic (A), soil (B) and targeted monoterpenes (C) covariates, estimated**


**using a linear kernel (Linear) and two nonlinear kernels: the Gaussian kernel (GK; bandwidth**


**parameter** _**h**_ **= 1) and the deep arc-cosine kernel (AK; one layer).** Site abbreviations are described in


the text.


**Figure 5. Averaged percentage reduction in residual (unexplained) variance relative to the baseline**


**multi-environment ssGBLUP Model 1, which includes only environmental (site) and genetic effects**


**(M1; E + G). Results are shown for Model 2 (M2; E + G + W), Model 3 (M3; E + G + W + G×W),**


**Model 4 (M4; E + G + W + G×E), and Model 5 (M5; E + G + W + G×E + G×W), across traits,**


**environmental covariates, and kernels approaches.** Residual variances were obtained by averaging


across traits. For nonlinear kernel approaches, results correspond to the Gaussian kernel (GK) with


bandwidth _h_ = 1 and the deep arc-cosine kernel (AK) with a single (1) layer.


**Figure 6** . **Variance component partitioning averaged across the twelve studied traits.** Bars represent


the percentage of variance explained by each source for each model averaged across trait including


environment (Site), replication nested within site (Rep), additive genetic (G), genotype-by-environment


interaction (G×E), environment covariates (W), and genotype-by-covariate interaction (G×W) components,


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


34


as estimated using the multi-environment ssGBLUP mixed models. Results are shown for the five models


studied. Models differ in the inclusion of climatic (c), soil (s), and targeted monoterpene (m) covariates,


fitted using linear (L) and nonlinear kernels: Gaussian (GK; _h_ = 1) and arc-cosine (AK; one layer). Model


abbreviations are described in the text and Table 2.


**Figure 7. Variance component partitioning across twelve traits.** Bars represent the percentage of


variance explained by each source for each model, including environment (Site), replication nested within


site (Rep), additive genetic (G), genotype-by-environment interaction (G×E), covariates (W), and


genotype-by-covariate interaction (G×W) components, as estimated using the multi-environment


ssGBLUP mixed models. Results are shown for the five models studied. Models differ in the inclusion of


climatic (c), soil (s), and targeted monoterpene (m) covariates, fitted using linear (L) and nonlinear kernels:


Gaussian (GK; _h_ = 1) and arc-cosine (AK; one layer). Trait abbreviations are described in the text and Table


S1.


**Figure 8. Predictive ability of multi-environment ssGBLUP reaction norm mixed models evaluated**


**across twelve traits in each untested environment.** Results are shown for the five models studied,


covariate types (climate, soil and targeted monoterpenes), and kernels approaches (linear, Gaussian kernel


and arc-cosine kernel). Trait and model abbreviations are described in the text and Table S1 and Table 2,


respectively.


**Figure 9. Predictive ability averaged across the twelve studied traits and four sites.** Results are shown


for the five models studied. Symbols indicate the type of covariates included: climatic (+), soil (×), and


targeted monoterpenes (•), while shapes represent kernel approaches: linear, Gaussian kernel (GK;


bandwidth _h_ = 1), and arc-cosine kernel (AK; one layer). Numbers above symbols indicate the mean


predictive ability for each model averaged across traits and sites.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


35


**Table 1. List of covariates considered in this study including climatic, soil and targeted monoterpenes**


**compounds, units, abbreviations and descriptions.**


**Variable** **Unit** **Abbreviation** **Description**

**Climatic**

Total Precipitation mm PRECTOT Total Precipitation

Relative air humidity at 2 m
Relative Humidity % RH2M
above the surface of the earth

Temperature °C T2M Temperature at 2 Meters

Minimum air temperature at 2 m
Minimum temperature °C T2M_MIN
above the surface of the earth

Maximum air temperature at 2
Maximum Temperature °C T2M_MAX
m above the surface of the earth

The dew-point temperature at 2
Dew-point temperature °C T2MDEW
m above the surface of the earth



All Sky Surface

Total Thermal Infrared ALLSKY_SFC_PAR
MJ m &#91;-&#93;&#91;2&#93; d &#91;-&#93;&#91;1&#93; Photosynthetically Active
Radiative Flux _TOT

Radiation (PAR) Total
Potential Evapotranspiration mm ETP
Effect of Temperature on Effect of Temperature on

                           - FRUE
Radiation use Efficiency Radiation use Efficiency
Growing Degree-days °C d &#91;-1&#93; GDD
Daylight hours hour N Daylight hours
Extraterrestrial radiation MJ m &#91;-&#93;&#91;2&#93; d &#91;-&#93;&#91;1&#93; RTA
Slope of saturation vapor
kPa °C SPV
pressure curve



Total Thermal Infrared ALLSKY_SFC_PAR
MJ m &#91;-&#93;&#91;2&#93; d &#91;-&#93;&#91;1&#93;
Radiative Flux



_TOT



Deficit of vapor Pressure kPa d &#91;-1&#93; VPD The deficit of vapor Pressure

**Soil**

Cation exchange capacity- cmol(c)

CECPH7

buffered at pH7 kg &#91;-1&#93;



cmol(c)



CECPH7
kg &#91;-1&#93;



Clay total g/100 g &#91;-1&#93; CLAY
Total nitrogen g kg &#91;-1&#93; NITKJD
Organic carbon g kg &#91;-1&#93; ORGC
pH H 2 O - PHAQ
Sand Total g/100 g SAND
Silt Total g/100 g SILT
**Chemical monoterpenes** **compounds**
α-pinene ng mg &#91;-1&#93; α-pinene
β-pinene ng mg &#91;-1&#93; β-pinene
β-phellandrene ng mg &#91;-1&#93; β-phellandrene
camphene ng mg &#91;-1&#93; camphene
limonene ng mg &#91;-1&#93; limonene
myrcene ng mg &#91;-1&#93; myrcene
terpinolene ng mg &#91;-1&#93; terpinolene


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


36


**Table 2. Summary of the multi-environment ssGBLUP models evaluated for each of the twelve traits,**


**including order, covariates, kernel type, model-kernel-covariate abbreviations, and the**


**corresponding model.** Kernels, and covariates are described in the text. Model abbreviations include the


following components: environment effects (S), additive genetic effects (G), environment covariates effects


(W), genotype-by-environment interaction (G×E), and genotype-by-covariate interaction (G×W).


Covariates correspond to climatic (c), soil (s), and chemical monoterpene (m) data. Nonlinear kernels


include the Gaussian kernel (GK) and deep arc-cosine kernel (AK) approaches. While the deep arc-cosine


kernel (AK) was fitted using a single layer (1), the Gaussian kernel (GK) was fitted using a bandwidth


parameter equal to 1 ( _h_ = 1).


**Order** **Covariates** **Kernel** **Abbreviation** **Model**

1     -     - M1 E + G

2 Climatic Linear M2 E + G + Wc

3 Climatic Nonlinear GK M2 E + G + Wc (GK)

4 Climatic Nonlinear AK M2 E + G + Wc (AK)

5 Soil Linear M2 E + G + Ws

6 Soil Nonlinear GK M2 E + G + Ws (GK)
7 Soil Nonlinear AK M2 E + G + Ws (AK)

8 Monoterpenes Linear M2 E + G + Wm
9 Monoterpenes Nonlinear GK M2 E + G + Wm (GK)

10 Monoterpenes Nonlinear AK M2 E + G + Wm (AK)

11 Climatic Linear M3 E + G + Wc + G×Wc

12 Climatic Nonlinear GK M3 E + G + Wc + G×Wc (GK)
13 Climatic Nonlinear AK M3 E + G + Wc + G×Wc (AK)
14 Soil Linear M3 E + G + Ws + G×Ws

15 Soil Nonlinear GK M3 E + G + Ws + G×Ws (GK)
16 Soil Nonlinear AK M3 E + G + Ws + G×Ws (AK)
17 Monoterpenes Linear M3 E + G + Wm + G×Wm
18 Monoterpenes Nonlinear GK M3 E + G + Wm + G×Wm (GK)
19 Monoterpenes Nonlinear AK M3 E + G + Wm + G×Wm (AK)

20 Climatic Linear M4 E + G + Wc + G×E

21 Climatic Nonlinear GK M4 E + G + Wc + G×E (GK)
22 Climatic Nonlinear AK M4 E + G + Wc + G×E (AK)
23 Soil Linear M4 E + G + Ws + G×E

24 Soil Nonlinear GK M4 E + G + Ws + G×E (GK)
25 Soil Nonlinear AK M4 E + G + Ws + G×E (AK)
26 Monoterpenes Linear M4 E + G + Wm + G×E
27 Monoterpenes Nonlinear GK M4 E + G + Wm + G×E (GK)
28 Monoterpenes Nonlinear AK M4 E + G + Wm + G×E (AK)

29 Climatic Linear M5 E + G + Wc + G×E + G×Wc

30 Climatic Nonlinear GK M5 E + G + Wc + G×E + G×Wc (GK)


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


37


31 Climatic Nonlinear AK M5 E + G + Wc + G×E + G×Wc (AK)
32 Soil Linear M5 E + G + Ws + G×E + G×Ws

33 Soil Nonlinear GK M5 E + G + Ws + G×E + G×Ws (GK)
34 Soil Nonlinear AK M5 E + G + Ws + G×E + G×Ws (AK)
35 Monoterpenes Linear M5 E + G + Wm + G×E + G×Wm
36 Monoterpenes Nonlinear GK M5 E + G + Wm + G×E + G×Wm (GK)
37 Monoterpenes Nonlinear AK M5 E + G + Wm + G×E + G×Wm (AK)


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


38


**Table 3. Estimated heritabilities (with their approximate standard error) for the 12 traits evaluated**


**in the lodgepole population using a pedigree-based relationship matrix (** _**A**_ **-matrix).** Abbreviations used


for the traits and sites are described in the text.


**Traits/sites** **JUDY** **VIRG** **SWAN** **TIME** **Average**
**HT09** 0.30 (0.06) 0.27 (0.07) 0.38 (0.05) 0.19 (0.04) **0.29 (0.06)**
**HT23** 0.42 (0.07) 0.28 (0.07) 0.46 (0.06) 0.35 (0.05) **0.38 (0.06)**
**HT30** 0.35 (0.06) 0.33 (0.07) 0.42 (0.06) 0.33 (0.05) **0.36 (0.06)**
**HT35** 0.72 (0.21) 0.85 (0.25) 0.87 (0.21) 0.76 (0.21) **0.80 (0.22)**
**DBH23** 0.18 (0.05) 0.29 (0.07) 0.23 (0.04) 0.18 (0.04) **0.22 (0.05)**
**DBH30** 0.22 (0.06) 0.30 (0.07) 0.25 (0.04) 0.20 (0.04) **0.24 (0.05)**
**DBH35** 0.61 (0.21) 0.57 (0.22) 0.33 (0.16) 0.33 (0.16) **0.46 (0.19)**
**NSWGR** 0.43 (0.07) 0.51 (0.08) 0.55 (0.06) 0.49 (0.06) **0.50 (0.07)**
**WD** 0.65 (0.22) 0.38 (0.21) 0.33 (0.16) 0.48 (0.19) **0.46 (0.20)**
**Decline** 0.41 (0.20) 0.00 (0.00) 0.24 (0.16) 0.23 (0.17) **0.22 (0.13)**
**C13** 0.65 (0.20) 0.53 (0.19) 0.63 (0.19) 0.54 (0.17) **0.59 (0.19)**
**FC** 0.58 (0.21) 0.24 (0.18) 0.22 (0.14) 0.17 (0.14) **0.30** **(0.17)**
**Average** **0.42 (0.14)** **0.34 (0.11)** **0.40 (0.12)** **0.33 (0.12)** **0.37 (0.12)**


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


39


**Table 4. Model fit statistics for all evaluated multi-environment ssGBLUP models across traits,**


**including log-likelihood (LogL) and Akaike Information Criterion (AIC).** Values are reported for each


trait and model, with both LogL and AIC representing averages across kernel types and covariate sets.


Lower AIC (highlighted in bold) indicating better model fit. Trait and model abbreviations are described in


the main text.


**HT09** **HT23** **HT30** **HT35** **DBH23** **DBH30** **DBH35** **NSWGR** **WD** **Decline** **C13** **FC**


**LogL**


**M1:** E+G -3513.9 -4389.4 -4681.9 -517.7 -4971.8 -4950.6 -635.2 -5055.3 -585.6 -586.2 -525.9 -690.1


**M2:** E+G+W -3513.8 -4389.4 -4681.9 -517.7 -4971.8 -4950.6 -635.2 -5055.3 -585.0 -586.2 -525.9 -690.1


**M3** : E+G+W+G×W -3499.1 -4382.1 -4669.2 -517.7 -4969.8 -4950.2 -635.2 -5054.9 -584.4 -585.2 -525.9 -689.5


**M4:** E+G+W+G×E -3487.7 -4375.2 -4659.7 -517.7 -4967.3 -4949.3 -635.2 -5053.8 -584.1 -583.8 -525.9 -688.7


**M5:** E+G+W+G×E+G×W -3487.7 -4375.2 -4659.7 -517.7 -4967.3 -4949.3 -635.2 -5053.8 -584.0 -583.8 -525.9 -688.7


**AIC**


**M1:** E+G 7041.8 8792.8 9377.8 **1049.4** 9957.6 **9915.2** 1284.3 **10124.6** **1185.2** 1186.5 **1065.9** **1394.2**


**M2:** E+G+W 7043.6 8794.8 9379.8 1051.4 9959.6 9917.2 1286.3 10126.6 1186.1 1188.5 1067.8 1396.2


**M3** : E+G+W+G×W 7016.2 8782.3 9356.5 1053.4 9957.6 9918.5 1288.3 10127.8 1186.9 1188.4 1069.8 1397.0


**M4:** E+G+W+G×E **6993.4** **8768.4** **9337.3** 1053.4 **9952.6** 9916.5 **1288.3** 10125.7 1186.1 **1185.6** 1069.8 1395.4


**M5:** E+G+W+G×E+G×W 6995.4 8770.4 9339.3 1055.4 9954.6 9918.5 1290.3 10127.7 1188.0 1187.6 1071.8 1397.4


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


40


**Table 5.** **Average proportion of phenotypic variance explained by each variance component, and**


**residual variance, across all studied traits for the fitted multi-environment ssGBLUP models.** Models


incorporated enviromic covariates (climatic and soil variables) and targeted metabolomic covariates


(monoterpenes), modeled using a linear kernel (Linear) and nonlinear kernels, including the Gaussian


kernel (GK; bandwidth parameter _h_ = 1) and the deep arc-cosine kernel (AK; single hidden layer). Reported


values represent averages across traits. See text for model abbreviations.


**Total explained phenotypic**
**Model** **Residual variance**

**variance**


**Linear** **GK** **AK** **Linear** **GK** **AK**

S+G 13.92 86.08

**Climate**

S+G+W 13.64 14.02 13.35 86.36 85.98 86.65

S+G+W+G×W 14.32 15.82 13.35 85.68 84.18 86.65

S+G+W+G×E 15.92 16.31 15.62 84.08 83.69 84.38

S+G+W+G×E+G×W 15.95 16.34 15.62 84.05 83.66 84.38

**Soil**

S+G+W 13.59 14.31 12.80 86.41 85.69 87.20

S+G+W+G×W 14.92 16.31 12.96 85.08 83.69 87.04

S+G+W+G×E 15.86 16.60 15.04 84.14 83.40 84.96

S+G+W+G×E+G×W 15.86 16.61 15.04 84.14 83.39 84.96

**Monoterpenes**
S+G+W 13.73 14.01 13.62 86.27 85.99 86.38

S+G+W+G×W 14.33 15.39 13.60 85.67 84.61 86.40

S+G+W+G×E 16.00 16.29 15.89 84.00 83.71 84.11

S+G+W+G×E+G×W 16.04 16.29 15.89 83.96 83.71 84.11


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


41


**Table 6.** **Average Predictive ability across all studied traits for the fitted multi-environment ssGBLUP**


**models.** Models incorporated enviromic covariates (climatic and soil variables) and targeted metabolomic


covariates (monoterpenes), modeled using a linear kernel (Linear) and nonlinear kernels, including the


Gaussian kernel (GK; bandwidth parameter _h_ = 1) and the deep arc-cosine kernel (AK; single hidden layer).


Site and model abbreviations are described in the main text. See text for trait and enviroment abbrevations.


**Model** **Linear** **GK** **AK**


**JUDY** **VIRG** **SWAN** **TIME** **JUDY** **VIRG** **SWAN** **TIME** **JUDY** **VIRG** **SWAN** **JUDY**


S+G 0.700 0.721 0.653 0.699


**Climate**


S+G+W 0.699 0.721 0.653 0.700 0.700 0.701 0.663 0.699 0.699 0.701 0.653 0.700


S+G+W+G×W 0.699 0.722 0.666 0.697 0.706 0.704 0.662 0.706 0.691 0.699 0.671 0.694


S+G+W+G×E 0.702 0.722 0.669 0.708 0.702 0.703 0.670 0.708 0.702 0.703 0.669 0.708


S+G+W+G×E+G×W 0.703 0.723 0.670 0.708 0.703 0.703 0.670 0.708 0.709 0.703 0.669 0.708


**Soil**


S+G+W 0.699 0.721 0.653 0.699 0.700 0.701 0.653 0.699 0.700 0.701 0.653 0.699


S+G+W+G×W 0.702 0.723 0.663 0.698 0.696 0.703 0.665 0.705 0.698 0.699 0.669 0.701


S+G+W+G×E 0.702 0.722 0.669 0.708 0.702 0.703 0.669 0.708 0.702 0.703 0.669 0.708


S+G+W+G×E+G×W 0.702 0.722 0.669 0.708 0.703 0.703 0.668 0.708 0.700 0.703 0.669 0.708


**Monoterpenes**


S+G+W 0.699 0.721 0.653 0.700 0.699 0.701 0.653 0.700 0.699 0.701 0.653 0.700


S+G+W+G×W 0.699 0.722 0.666 0.696 0.714 0.699 0.658 0.705 0.697 0.698 0.672 0.692


S+G+W+G×E 0.702 0.722 0.669 0.708 0.702 0.703 0.669 0.708 0.702 0.703 0.669 0.708


S+G+W+G×E+G×W 0.703 0.723 0.669 0.709 0.702 0.699 0.668 0.708 0.703 0.703 0.677 0.708


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


42


**Table 7.** **Average Predictive ability for the four traits showing the strongest G×E patterns (HT09,**


**HT23, HT30, and Decline) (average pedigree-based genetic correlations across site pairs &lt; 0.65,**


**Figure 2) across multi-environment ssGBLUP reaction-norm models.** Models incorporated enviromic


covariates (climatic and soil variables) and targeted metabolomic covariates (monoterpenes), modeled


using a linear kernel (Linear) and nonlinear kernels, including the Gaussian kernel (GK; bandwidth


parameter _h_ = 1) and the deep arc-cosine kernel (AK; single hidden layer). Site and model abbreviations


are described in the main text. See text for trait and enviroment abbrevations.


**Model** **Linear** **GK** **AK**


**JUDY** **VIRG** **SWAN** **TIME** **JUDY** **VIRG** **SWAN** **TIME** **JUDY** **VIRG** **SWAN** **JUDY**


S+G 0.702 0.728 0.594 0.667


**Climate**


S+G+W 0.702 0.727 0.594 0.667 0.702 0.727 0.613 0.667 0.702 0.727 0.594 0.667


S+G+W+G×W 0.700 0.733 0.623 0.659 0.712 0.730 0.614 0.679 0.695 0.722 0.636 0.652


S+G+W+G×E 0.703 0.728 0.627 0.684 0.703 0.728 0.616 0.684 0.703 0.728 0.627 0.684


S+G+W+G×E+G×W 0.703 0.729 0.628 0.683 0.704 0.728 0.629 0.684 0.726 0.729 0.628 0.683


**Soil**


S+G+W 0.702 0.727 0.594 0.667 0.702 0.727 0.594 0.667 0.702 0.727 0.594 0.667


S+G+W+G×W 0.706 0.729 0.614 0.662 0.698 0.728 0.620 0.677 0.697 0.723 0.628 0.669


S+G+W+G×E 0.703 0.728 0.627 0.684 0.703 0.728 0.627 0.684 0.703 0.728 0.627 0.684


S+G+W+G×E+G×W 0.703 0.728 0.627 0.684 0.703 0.728 0.610 0.684 0.703 0.729 0.627 0.684


**Monoterpenes**


S+G+W 0.702 0.727 0.594 0.667 0.702 0.727 0.594 0.667 0.702 0.727 0.594 0.667


S+G+W+G×W 0.701 0.733 0.622 0.656 0.705 0.723 0.606 0.677 0.695 0.721 0.636 0.648


S+G+W+G×E 0.703 0.728 0.627 0.684 0.703 0.728 0.627 0.684 0.703 0.728 0.627 0.684


S+G+W+G×E+G×W 0.703 0.729 0.627 0.683 0.703 0.728 0.627 0.684 0.704 0.728 0.640 0.684


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


43


**Figure 1. Boxplot of standarized phenotipic data for the 12 traits studied in the lodgepole pine**
**population in four enviroment.** See text for trait and enviroment abbrevations.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


44


**Figure 2. Estimated genetic correlations between sites for each trait from the multiple-site model**
**using pedigree- (** _**A**_ **-matrix) relationship matrices for the lodgepole pine population** . The dark red to
dark blue color of each individual cell above the diagonal reflects the strength of the genetic correlation.
Abbreviations used for the traits and sites are described in the text.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


45


**Figure 3. Heta maps of the climatic (A), soil (B) and targeted monoterpenes (C) covariates across**
**the four lodgepole pine environments.** Covariate abbreviations are provided in Table 1, and site
abbreviations are described in the text.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


46


**Figure 4. Heat maps of the environment similarity matrix among the four lodgepole pine**
**environments considering climatic (A), soil (B) and targeted monoterpenes (C) covariates,**
**estimated using a linear kernel (Linear) and two nonlinear kernels: the Gaussian kernel (GK;**
**bandwidth parameter** _**h**_ **= 1) and the deep arc-cosine kernel (AK; one layer).** Site abbreviations are
described in the text.


&#91;bioRxiv preprint doi: https://doi.org/10.64898/2026.11.16.767158; this version posted November 17, 2026. The copyright holder for this preprint&#93;(https://doi.org/10.64898/2026.11.16.767158)

(which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.


47


**Figure 5. Averaged percentage reduction in residual (unexplained) variance relative to the baseline**
**multi-environment ssGBLUP Model 1, which includes only environmental (site) and genetic effects**
**(M1; E + G). Results are shown for Model 2 (M2; E + G + W), Model 3 (M3; E + G + W + G×W),**
**Model 4 (M4; E + G + W + G×E), and Model 5 (M5; E + G + W + G×E + G×W), across traits,**
**environmental covariates, and kernels approaches.** Residual variances were obtained by averaging
across traits. For nonlinear kernel approaches, results correspond to the Gaussian kernel (GK) with
bandwidth _h_ = 1 and the deep arc-cosine kernel (AK) with a single (1) layer.


48


**Figure 6** . **Variance component partitioning averaged across the twelve studied traits.** Bars represent the percentage of variance explained by
each source for each model averaged across trait including environment (Site), replication nested within site (Rep), additive genetic (G), genotypeby-environment interaction (G×E), environment covariates (W), and genotype-by-covariate interaction (G×W) components, as estimated using the
multi-environment ssGBLUP mixed models. Results are shown for the five models studied. Models differ in the inclusion of climatic (c), soil (s),
and targeted monoterpene (m) covariates, fitted using linear (L) and nonlinear kernels: Gaussian (GK; _h_ = 1) and arc-cosine (AK; one layer).
Model abbreviations are described in the text and Table 2.


49


**Figure 7. Variance component partitioning across twelve traits.** Bars represent the percentage of variance explained by each source for each
model, including environment (Site), replication nested within site (Rep), additive genetic (G), genotype-by-environment interaction (G×E),
covariates (W), and genotype-by-covariate interaction (G×W) components, as estimated using the multi-environment ssGBLUP mixed models.
Results are shown for the five models studied. Models differ in the inclusion of climatic (c), soil (s), and targeted monoterpene (m) covariates,
fitted using linear (L) and nonlinear kernels: Gaussian (GK; _h_ = 1) and arc-cosine (AK; one layer). Trait abbreviations are described in the text and
Table S1.


50



**Figure 8. Predictive ability of multi-environment ssGBLUP reaction norm mixed models evaluated across twelve traits in each untested**
**environment.** Results are shown for the five models studied, covariate types (climate, soil and targeted monoterpenes), and kernels approaches
(linear, Gaussian kernel and arc-cosine kernel). Trait and model abbreviations are described in the text and Table S1 and Table 2, respectively.


51


**Figure 9. Predictive ability averaged across the twelve studied traits and four sites.** Results are shown for the five models studied. Symbols
indicate the type of covariates included: climatic (+), soil (×), and targeted monoterpenes (•), while shapes represent kernel approaches: linear,
Gaussian kernel (GK; bandwidth _h_ = 1), and arc-cosine kernel (AK; one layer). Numbers above symbols indicate the mean predictive ability for
each model averaged across traits and sites.


