Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce

Autores
Cappa, Eduardo Pablo; Chen, Charles; Benowicz, Andy; El-Kassaby, Yousry A.; Thomas, Barb R.
Año de publicación
2026
Idioma
inglés
Tipo de recurso
documento de conferencia
Estado
versión publicada
Descripción
Póster
Background: Genomic prediction in forest tree breeding has traditionally relied on parametric linear models such as genomic best linear unbiased prediction (GBLUP), which assume additive genetic effects and linear genotype–phenotype relationships. Although effective for additive traits, these models may be suboptimal for complex traits influenced by dominance and epistasis, particularly in long-lived tree species where productivity, defense, and climate-related traits are often polygenic. Machine-learning (ML) kernel methods, including Gaussian and arc-cosine kernels, offer a flexible alternative by implicitly capturing non-linear and interaction genetic effects, yet their integration into single-step GBLUP (ssGBLUP) is limited in forest trees. Methods: We evaluated single- and multi-trait ssGBLUP models using linear and non-linear kernels for 30 traits in white spruce (Picea glauca (Moench) Voss). Phenotypic data were obtained from an open-pollinated progeny test in Alberta, Canada (up to 3,936 trees). A subset of 612 trees was genotyped using genotyping-by-sequencing, yielding 211,061 SNPs after quality control. Standard ssGBLUP using VanRaden’s genomic relationship matrix served as a benchmark and was compared with two ML-based extensions: (i) a Gaussian kernel using bandwidth averaging and (ii) an arc-cosine kernel (depth 1–20). Predictive ability was assessed by ten-fold cross-validation using Pearson correlations between predicted and reference breeding values. Models partitioning additive, dominance, and epistatic variance were also fitted. Results: Genetic variance partitioning showed marked heterogeneity. Additive variance predominated in 21 of 30 traits. Several defense traits exhibited strong dominance (>50% of total variance), while physiological and climate-related traits (e.g., stomatal conductance, water-use efficiency, intercellular CO₂ concentration) showed high epistatic contributions (>25%). Predictive ability varied by trait and kernel. The arc-cosine kernel achieved the highest accuracy in 29 of 30 traits, with gains >23% for physiological traits and improvements for some secondary metabolites. Moderate gains (5–10%) were observed for defense traits and wood density, whereas growth traits showed minimal improvement. Diameter at breast height was the only trait where the linear kernel performed best. The Gaussian kernel produced smaller but positive gains. In multi-trait analyses, performance was similar across kernels, although integration improved accuracy for low-heritability traits such as drought resistance (15.9–18.7%). Conclusions: Non-linear kernels improved genomic prediction in white spruce, with the arc-cosine kernel outperforming linear and Gaussian models for traits with complex non-addative architectures, while linear models remained competitive for core productivity traits. Integrating non-linear kernels into ssGBLUP can enhance prediction accuracy and accelerate genetic gain in climate-resilient breeding programs.
Instituto de Recursos Biológicos
Fil: Cappa, Eduardo Pablo. Instituto Nacional de Tecnología Agropecuaria (INTA). Instituto de Recursos Biológicos; Argentina
Fil: Cappa, Eduardo Pablo. Consejo Nacional de Investigaciones Científicas y Técnicas; Argentina
Fil: Chen, Charles. Oklahoma State University. Department of Biochemistry and Molecular Biology; Estados Unidos
Fil: Benowicz, Andy. Forest Stewardship and Trade Branch. Alberta Forestry and Parks; Canadá
Fil: El-Kassaby, Yousry. University of British Columbia. Faculty of Forestry. Department of Forest and Conservation Sciences; Canadá
Fil: Thomas, Barb R. University of Alberta. Department of Renewable Resources; Canadá
Fuente
3rd Biennial Conference of the North American Forest Genetics Society (NAFGS), Québec City, Canada, June 15th - 19th, 2026.
Materia
Machine Learning
Genomics
Forest Trees
Aprendizaje Automático
Genómica
Árboles Forestales
Capacidad Predictiva
Modelos Mixtos de Rasgos Unicos y Múltiples
Predicción Genómica
Abeto Blanco
BLUP de un Solo Paso
Método Kernel
Efecto Genético No-Aditivo
Predictive Ability
Single and Multiple Trait Mixed Models
Genomic Prediction
Tree Breeding
White Spruce
Single Step GBLUP
Kernel Methods
Non-Additive Genetic Effects
Nivel de accesibilidad
acceso abierto
Condiciones de uso
http://creativecommons.org/licenses/by-nc-sa/4.0/
Repositorio
INTA Digital (INTA)
Institución
Instituto Nacional de Tecnología Agropecuaria
OAI Identificador
oai:localhost:20.500.12123/27940

id INTADig_4f0a945b32c98c393fae08bb65ca6fde
oai_identifier_str oai:localhost:20.500.12123/27940
network_acronym_str INTADig
repository_id_str l
network_name_str INTA Digital (INTA)
spelling Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruceCappa, Eduardo PabloChen, CharlesBenowicz, AndyEl-Kassaby, Yousry A.Thomas, Barb R.Machine LearningGenomicsForest TreesAprendizaje AutomáticoGenómicaÁrboles ForestalesCapacidad PredictivaModelos Mixtos de Rasgos Unicos y MúltiplesPredicción GenómicaAbeto BlancoBLUP de un Solo PasoMétodo KernelEfecto Genético No-AditivoPredictive AbilitySingle and Multiple Trait Mixed ModelsGenomic PredictionTree BreedingWhite SpruceSingle Step GBLUPKernel MethodsNon-Additive Genetic EffectsPósterBackground: Genomic prediction in forest tree breeding has traditionally relied on parametric linear models such as genomic best linear unbiased prediction (GBLUP), which assume additive genetic effects and linear genotype–phenotype relationships. Although effective for additive traits, these models may be suboptimal for complex traits influenced by dominance and epistasis, particularly in long-lived tree species where productivity, defense, and climate-related traits are often polygenic. Machine-learning (ML) kernel methods, including Gaussian and arc-cosine kernels, offer a flexible alternative by implicitly capturing non-linear and interaction genetic effects, yet their integration into single-step GBLUP (ssGBLUP) is limited in forest trees. Methods: We evaluated single- and multi-trait ssGBLUP models using linear and non-linear kernels for 30 traits in white spruce (Picea glauca (Moench) Voss). Phenotypic data were obtained from an open-pollinated progeny test in Alberta, Canada (up to 3,936 trees). A subset of 612 trees was genotyped using genotyping-by-sequencing, yielding 211,061 SNPs after quality control. Standard ssGBLUP using VanRaden’s genomic relationship matrix served as a benchmark and was compared with two ML-based extensions: (i) a Gaussian kernel using bandwidth averaging and (ii) an arc-cosine kernel (depth 1–20). Predictive ability was assessed by ten-fold cross-validation using Pearson correlations between predicted and reference breeding values. Models partitioning additive, dominance, and epistatic variance were also fitted. Results: Genetic variance partitioning showed marked heterogeneity. Additive variance predominated in 21 of 30 traits. Several defense traits exhibited strong dominance (>50% of total variance), while physiological and climate-related traits (e.g., stomatal conductance, water-use efficiency, intercellular CO₂ concentration) showed high epistatic contributions (>25%). Predictive ability varied by trait and kernel. The arc-cosine kernel achieved the highest accuracy in 29 of 30 traits, with gains >23% for physiological traits and improvements for some secondary metabolites. Moderate gains (5–10%) were observed for defense traits and wood density, whereas growth traits showed minimal improvement. Diameter at breast height was the only trait where the linear kernel performed best. The Gaussian kernel produced smaller but positive gains. In multi-trait analyses, performance was similar across kernels, although integration improved accuracy for low-heritability traits such as drought resistance (15.9–18.7%). Conclusions: Non-linear kernels improved genomic prediction in white spruce, with the arc-cosine kernel outperforming linear and Gaussian models for traits with complex non-addative architectures, while linear models remained competitive for core productivity traits. Integrating non-linear kernels into ssGBLUP can enhance prediction accuracy and accelerate genetic gain in climate-resilient breeding programs.Instituto de Recursos BiológicosFil: Cappa, Eduardo Pablo. Instituto Nacional de Tecnología Agropecuaria (INTA). Instituto de Recursos Biológicos; ArgentinaFil: Cappa, Eduardo Pablo. Consejo Nacional de Investigaciones Científicas y Técnicas; ArgentinaFil: Chen, Charles. Oklahoma State University. Department of Biochemistry and Molecular Biology; Estados UnidosFil: Benowicz, Andy. Forest Stewardship and Trade Branch. Alberta Forestry and Parks; CanadáFil: El-Kassaby, Yousry. University of British Columbia. Faculty of Forestry. Department of Forest and Conservation Sciences; CanadáFil: Thomas, Barb R. University of Alberta. Department of Renewable Resources; CanadáNorth American Forest Genetics Society2026-09-30T10:18:15Z2026-09-30T10:18:15Z2026-06info:eu-repo/semantics/conferenceObjectinfo:eu-repo/semantics/publishedVersionhttp://purl.org/coar/resource_type/c_5794info:ar-repo/semantics/documentoDeConferenciaapplication/pdfhttp://hdl.handle.net/20.500.12123/27940Cappa, E.P., Chen, C., Benowicz, A., Benowicz, A., El-Kassaby, Y.A., &, Thomas, B.R. (2026, 15th - 19th June). Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce [Sesión de póster]. 3rd Biennial Conference of the North American Forest Genetics Society (NAFGS). Québec City, Canada.3rd Biennial Conference of the North American Forest Genetics Society (NAFGS), Québec City, Canada, June 15th - 19th, 2026.reponame:INTA Digital (INTA)instname:Instituto Nacional de Tecnología Agropecuariaenginfo:eu-repo/semantics/openAccesshttp://creativecommons.org/licenses/by-nc-sa/4.0/Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)2026-11-05T11:51:39Zoai:localhost:20.500.12123/27940instacron:INTAInstitucionalhttp://repositorio.inta.gob.ar/Organismo científico-tecnológicoNo correspondehttp://repositorio.inta.gob.ar/oai/requesttripaldi.nicolas@inta.gob.arArgentinaNo correspondeNo correspondeNo correspondeopendoar:l2026-11-05 11:51:39.433INTA Digital (INTA) - Instituto Nacional de Tecnología Agropecuariafalse
dc.title.none.fl_str_mv Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
title Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
spellingShingle Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
Cappa, Eduardo Pablo
Machine Learning
Genomics
Forest Trees
Aprendizaje Automático
Genómica
Árboles Forestales
Capacidad Predictiva
Modelos Mixtos de Rasgos Unicos y Múltiples
Predicción Genómica
Abeto Blanco
BLUP de un Solo Paso
Método Kernel
Efecto Genético No-Aditivo
Predictive Ability
Single and Multiple Trait Mixed Models
Genomic Prediction
Tree Breeding
White Spruce
Single Step GBLUP
Kernel Methods
Non-Additive Genetic Effects
title_short Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
title_full Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
title_fullStr Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
title_full_unstemmed Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
title_sort Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce
dc.creator.none.fl_str_mv Cappa, Eduardo Pablo
Chen, Charles
Benowicz, Andy
El-Kassaby, Yousry A.
Thomas, Barb R.
author Cappa, Eduardo Pablo
author_facet Cappa, Eduardo Pablo
Chen, Charles
Benowicz, Andy
El-Kassaby, Yousry A.
Thomas, Barb R.
author_role author
author2 Chen, Charles
Benowicz, Andy
El-Kassaby, Yousry A.
Thomas, Barb R.
author2_role author
author
author
author
dc.subject.none.fl_str_mv Machine Learning
Genomics
Forest Trees
Aprendizaje Automático
Genómica
Árboles Forestales
Capacidad Predictiva
Modelos Mixtos de Rasgos Unicos y Múltiples
Predicción Genómica
Abeto Blanco
BLUP de un Solo Paso
Método Kernel
Efecto Genético No-Aditivo
Predictive Ability
Single and Multiple Trait Mixed Models
Genomic Prediction
Tree Breeding
White Spruce
Single Step GBLUP
Kernel Methods
Non-Additive Genetic Effects
topic Machine Learning
Genomics
Forest Trees
Aprendizaje Automático
Genómica
Árboles Forestales
Capacidad Predictiva
Modelos Mixtos de Rasgos Unicos y Múltiples
Predicción Genómica
Abeto Blanco
BLUP de un Solo Paso
Método Kernel
Efecto Genético No-Aditivo
Predictive Ability
Single and Multiple Trait Mixed Models
Genomic Prediction
Tree Breeding
White Spruce
Single Step GBLUP
Kernel Methods
Non-Additive Genetic Effects
dc.description.none.fl_txt_mv Póster
Background: Genomic prediction in forest tree breeding has traditionally relied on parametric linear models such as genomic best linear unbiased prediction (GBLUP), which assume additive genetic effects and linear genotype–phenotype relationships. Although effective for additive traits, these models may be suboptimal for complex traits influenced by dominance and epistasis, particularly in long-lived tree species where productivity, defense, and climate-related traits are often polygenic. Machine-learning (ML) kernel methods, including Gaussian and arc-cosine kernels, offer a flexible alternative by implicitly capturing non-linear and interaction genetic effects, yet their integration into single-step GBLUP (ssGBLUP) is limited in forest trees. Methods: We evaluated single- and multi-trait ssGBLUP models using linear and non-linear kernels for 30 traits in white spruce (Picea glauca (Moench) Voss). Phenotypic data were obtained from an open-pollinated progeny test in Alberta, Canada (up to 3,936 trees). A subset of 612 trees was genotyped using genotyping-by-sequencing, yielding 211,061 SNPs after quality control. Standard ssGBLUP using VanRaden’s genomic relationship matrix served as a benchmark and was compared with two ML-based extensions: (i) a Gaussian kernel using bandwidth averaging and (ii) an arc-cosine kernel (depth 1–20). Predictive ability was assessed by ten-fold cross-validation using Pearson correlations between predicted and reference breeding values. Models partitioning additive, dominance, and epistatic variance were also fitted. Results: Genetic variance partitioning showed marked heterogeneity. Additive variance predominated in 21 of 30 traits. Several defense traits exhibited strong dominance (>50% of total variance), while physiological and climate-related traits (e.g., stomatal conductance, water-use efficiency, intercellular CO₂ concentration) showed high epistatic contributions (>25%). Predictive ability varied by trait and kernel. The arc-cosine kernel achieved the highest accuracy in 29 of 30 traits, with gains >23% for physiological traits and improvements for some secondary metabolites. Moderate gains (5–10%) were observed for defense traits and wood density, whereas growth traits showed minimal improvement. Diameter at breast height was the only trait where the linear kernel performed best. The Gaussian kernel produced smaller but positive gains. In multi-trait analyses, performance was similar across kernels, although integration improved accuracy for low-heritability traits such as drought resistance (15.9–18.7%). Conclusions: Non-linear kernels improved genomic prediction in white spruce, with the arc-cosine kernel outperforming linear and Gaussian models for traits with complex non-addative architectures, while linear models remained competitive for core productivity traits. Integrating non-linear kernels into ssGBLUP can enhance prediction accuracy and accelerate genetic gain in climate-resilient breeding programs.
Instituto de Recursos Biológicos
Fil: Cappa, Eduardo Pablo. Instituto Nacional de Tecnología Agropecuaria (INTA). Instituto de Recursos Biológicos; Argentina
Fil: Cappa, Eduardo Pablo. Consejo Nacional de Investigaciones Científicas y Técnicas; Argentina
Fil: Chen, Charles. Oklahoma State University. Department of Biochemistry and Molecular Biology; Estados Unidos
Fil: Benowicz, Andy. Forest Stewardship and Trade Branch. Alberta Forestry and Parks; Canadá
Fil: El-Kassaby, Yousry. University of British Columbia. Faculty of Forestry. Department of Forest and Conservation Sciences; Canadá
Fil: Thomas, Barb R. University of Alberta. Department of Renewable Resources; Canadá
description Póster
publishDate 2026
dc.date.none.fl_str_mv 2026-09-30T10:18:15Z
2026-09-30T10:18:15Z
2026-06
dc.type.none.fl_str_mv info:eu-repo/semantics/conferenceObject
info:eu-repo/semantics/publishedVersion
http://purl.org/coar/resource_type/c_5794
info:ar-repo/semantics/documentoDeConferencia
format conferenceObject
status_str publishedVersion
dc.identifier.none.fl_str_mv http://hdl.handle.net/20.500.12123/27940
Cappa, E.P., Chen, C., Benowicz, A., Benowicz, A., El-Kassaby, Y.A., &, Thomas, B.R. (2026, 15th - 19th June). Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce [Sesión de póster]. 3rd Biennial Conference of the North American Forest Genetics Society (NAFGS). Québec City, Canada.
url http://hdl.handle.net/20.500.12123/27940
identifier_str_mv Cappa, E.P., Chen, C., Benowicz, A., Benowicz, A., El-Kassaby, Y.A., &, Thomas, B.R. (2026, 15th - 19th June). Machine learning-augmented single- and multi-trait single-step GBLUP enhanced genomic prediction in white spruce [Sesión de póster]. 3rd Biennial Conference of the North American Forest Genetics Society (NAFGS). Québec City, Canada.
dc.language.none.fl_str_mv eng
language eng
dc.rights.none.fl_str_mv info:eu-repo/semantics/openAccess
http://creativecommons.org/licenses/by-nc-sa/4.0/
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)
eu_rights_str_mv openAccess
rights_invalid_str_mv http://creativecommons.org/licenses/by-nc-sa/4.0/
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)
dc.format.none.fl_str_mv application/pdf
dc.publisher.none.fl_str_mv North American Forest Genetics Society
publisher.none.fl_str_mv North American Forest Genetics Society
dc.source.none.fl_str_mv 3rd Biennial Conference of the North American Forest Genetics Society (NAFGS), Québec City, Canada, June 15th - 19th, 2026.
reponame:INTA Digital (INTA)
instname:Instituto Nacional de Tecnología Agropecuaria
reponame_str INTA Digital (INTA)
collection INTA Digital (INTA)
instname_str Instituto Nacional de Tecnología Agropecuaria
repository.name.fl_str_mv INTA Digital (INTA) - Instituto Nacional de Tecnología Agropecuaria
repository.mail.fl_str_mv tripaldi.nicolas@inta.gob.ar
_version_ 1881032528471523328
score 13.334379