MultiGS: A Comprehensive and User-Friendly Genomic Prediction Platform Integrating Statistical, Machine Learning, and Deep Learning Models for Breeders

MultiGS: A Comprehensive and User-Friendly Genomic Prediction Platform Integrating Statistical, Machine Learning, and Deep Learning Models for Breeders

ES评分 0

DOI 10.20900/cbgg20260004
刊名
CBGG
年,卷(期) 2026, 8(1)
作者
作者单位

Crop Development Centre, Department of Plant Sciences, University of Saskatchewan, 51 Campus Drive, Saskatoon, SK S7N 5A8, CanadaOttawa Research and Development Centre, Agriculture and Agri-Food Canada, 960 Carling Avenue, Ottawa, ON K1A 0C6, Canada

摘要
Genomic selection (GS) is a core strategy in modern breeding programs, yet the rapid expansion of statistical, machine learning (ML), and deep learning (DL) models has made systematic evaluation and practical deployment increasingly challenging. To address these issues, we developed MultiGS, a unified and user-friendly framework that integrates linear, ML, DL, hybrid, and ensemble GS models within a standardized and computationally efficient workflow. MultiGS is implemented through two complementary pipelines: MultiGS-R, a Java/R pipeline implementing 12 statistical and ML models, and MultiGS-P, a Python pipeline integrating 17 models including five linear models, three ML approaches, and nine recently developed DL architectures. We benchmarked MultiGS using wheat, maize, and flax datasets representing contrasting prediction scenarios. Wheat and maize were evaluated using random training–test splits within the same population, reflecting suitable conditions for assessing model capacity and scalability. Under these scenarios, several DL, hybrid, and ensemble models achieved prediction accuracies comparable to or exceeding those of RR-BLUP and GBLUP. In contrast, the flax dataset represented a true across-population prediction scenario with limited training set size and strong population structure. In this challenging context, classical linear models provided stable baselines, while a subset of DL architectures, particularly graph-based models and BLUP-integrated hybrids, demonstrated comparatively improved generalization across populations. Comparisons with previously published DL tools showed that models implemented in MultiGS achieved comparable or improved prediction accuracies while requiring lower computational cost, enabling routine retraining and large-scale evaluation. Collectively, MultiGS supports scenario-specific model selection and provides a practical platform for deploying genomic prediction under realistic breeding conditions. The software is freely available on GitHub (https://github.com/AAFC-ORDC-Crop-Bioinformatics/MultiGS).
Abstract
Genomic selection (GS) is a core strategy in modern breeding programs, yet the rapid expansion of statistical, machine learning (ML), and deep learning (DL) models has made systematic evaluation and practical deployment increasingly challenging. To address these issues, we developed MultiGS, a unified and user-friendly framework that integrates linear, ML, DL, hybrid, and ensemble GS models within a standardized and computationally efficient workflow. MultiGS is implemented through two complementary pipelines: MultiGS-R, a Java/R pipeline implementing 12 statistical and ML models, and MultiGS-P, a Python pipeline integrating 17 models including five linear models, three ML approaches, and nine recently developed DL architectures. We benchmarked MultiGS using wheat, maize, and flax datasets representing contrasting prediction scenarios. Wheat and maize were evaluated using random training–test splits within the same population, reflecting suitable conditions for assessing model capacity and scalability. Under these scenarios, several DL, hybrid, and ensemble models achieved prediction accuracies comparable to or exceeding those of RR-BLUP and GBLUP. In contrast, the flax dataset represented a true across-population prediction scenario with limited training set size and strong population structure. In this challenging context, classical linear models provided stable baselines, while a subset of DL architectures, particularly graph-based models and BLUP-integrated hybrids, demonstrated comparatively improved generalization across populations. Comparisons with previously published DL tools showed that models implemented in MultiGS achieved comparable or improved prediction accuracies while requiring lower computational cost, enabling routine retraining and large-scale evaluation. Collectively, MultiGS supports scenario-specific model selection and provides a practical platform for deploying genomic prediction under realistic breeding conditions. The software is freely available on GitHub (https://github.com/AAFC-ORDC-Crop-Bioinformatics/MultiGS).
关键词
genomic selection; haplotype; principal component; prediction accuracy; cross-validation; across-population prediction; machine learning; deep learning; breeding
KeyWord
genomic selection; haplotype; principal component; prediction accuracy; cross-validation; across-population prediction; machine learning; deep learning; breeding
基金项目
页码 -
  • 参考文献
  • 相关文献
  • 引用本文

Bunyamin Tar’an, Chunfang Zheng, Frank M. You, John Joseph Zagariah Daniel, Pingchuan Li, Sylvie Cloutier. MultiGS: A Comprehensive and User-Friendly Genomic Prediction Platform Integrating Statistical, Machine Learning, and Deep Learning Models for Breeders [J]. Crop Breeding, Genetics and Genomics. 2026; 8; (1). - .

  • 文献评论

相关学者

相关机构