Integrative analysis of seed morphology, geographic origin, and genetic structure in Medicago with implications for breeding and conservation.

Background: Seed morphology and color are critical agronomic traits in Medicago spp., reflecting adaptations to diverse environments and influencing seedling establishment and vigor. Understanding the interplay between seed traits, geographic origin, and genetic diversity is crucial for effective germplasm conservation and breeding. This study presents a comprehensive analysis of these factors in a diverse collection of Medicago accessions, leveraging machine learning to illuminate these complex relationships.

Results: We analyzed seed size, shape, and color data from 318 Medicago accessions representing 29 species/subspecies from 31 countries. Machine learning models, including Neural Boost, Bootstrap Forest, and Support Vector Machines, effectively classified accessions based on seed traits and geographic origin, achieving up to 80% accuracy. Seed size was accurately predicted (R-squared > 0.80) using a combination of species, geographic origin, and shape descriptors. Hierarchical clustering of 189 M. sativa accessions based on 8,565 SNP markers revealed 20 distinct genetic clusters, indicating substantial population structure. A machine learning-based genome-wide association (GWA) analysis identified SNPs on chromosomes 1, 6, and 8 with high importance for predicting geographic origin. Notably, the most significant SNPs were located in or near genes involved in stress response and genome stability, suggesting their potential role in local adaptation. Finally, we successfully imputed missing M. sativa SNP genotypes using multiple machine learning approaches, achieving over 70% accuracy overall and over 80% for individual nucleotides (A, T, C, G), enhancing the utility of genomic datasets with missing data.

Conclusions: Our integrated analysis of phenotypic, genetic, and geographic data, coupled with a machine learning-based GWAS approach, provides valuable insights into the diverse patterns within Medicago spp. We demonstrate the power of machine learning for germplasm characterization, trait prediction, and imputation of missing genomic data. These findings have significant implications for seed trait improvement, germplasm management, and understanding adaptation in Medicago and other diverse crop species. The identified candidate genes associated with geographic origin provide a foundation for future investigations into the functional mechanisms of local adaptation. Furthermore, our imputation method offers a valuable data for maximizing the utility of genomic resources in Medicago and other species.