Utilizing Multispectral Time-Series Data for GWAS in Spring Wheat: A Comparison of Conventional and Machine-Learning Approaches
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3148374Utgivelsesdato
2024Metadata
Vis full innførselSamlinger
- Master’s theses (BioVit) [386]
Sammendrag
Plant breeding is the practice of breeding plants with desirable traits, for example, resistance to local pathogens, adaptation to environmental stressors, and increased yield. With the recent development of high-throughput phenotyping using remote sensing technologies and multispectral sensors, researchers can derive metrics (i.e., bands and vegetation indices) that correlate with important agronomic traits like yield and maturity time to study these traits in a high-throughput manner. There lies great potential in multispectral data for dissecting the genetic architecture of spring wheat (Triticum aestivum), and previous studies in a variety of crops (wheat, maize, and rice) have identified both known and novel genetic regions influencing agronomic traits.
This thesis explored the integration of multispectral time-series with Genome-Wide Association Studies (GWAS) in spring wheat. Given that spectral time-series are not typically used in association studies, we also implemented a machine learning (ML) alternative based on genetic algorithms and support vector regression (GA-SVR) for comparison with conventional GWAS using mixed linear models (MLM). Both GWAS approaches were employed on agronomic traits data and spectral phenotypes derived from the multispectral time-series.
Our results demonstrated that spectral phenotypes had moderate to large heritability, and correlated well with certain traits such as grain yield. We identified several spectral Quantitative Trait Loci (QTL) containing or near genes, like Rht-B1 and Vrn-A1, that overlapped with agronomic QTL related to the function of these genes. We also identified temporal genetic patterns, where certain spectral regions were significant only in specific time-periods (e.g., grain filling). In addition, the MLM detected few QTL associated with yield and protein content that overlap with spectral QTL. This suggests that the data or model may not be adequate for identifying spectral QTL related to complex physiological traits. On the other hand, the GA-SVR method produced a greater number of significant markers and QTL compared to MLM. It also found more spectral QTL overlapping with regions identified for yield and protein content, indicating GA-SVR may be suitable for capturing and dissecting more complex traits. However, results should be interpreted with care, as confounding variables were not accounted for in the GA-SVR pipeline.
In conclusion, this thesis demonstrates the potential usefulness of multispectral time-series for GWAS, because we identified temporal differences in significant spectral QTL, as well as basic genes related to height and maturity time. The GA-SVR method provided an alternative to MLM, and it is interesting to see that it returns different genomic regions, indicating that it might be capturing different information compared to MLM, and thus potentially overcoming some of the weaknesses of MLM. However, the effects of confounding variables in the GA-SVR pipeline remains unknown and requires further investigation.