Genetic diversity and population structure of temperate maize (Zea mays L.) inbred lines from a public breeding program in Argentina Authors: Ignacio Torrent, Roberto D. Lorea, Darío Micheli, Gustavo Rubén Rodríguez, Julio G. Velazco, Daniel A. Presello, María Laura Federico Published: 2026 ABSTRACT BackgroundUnderstanding and effectively leveraging the genetic diversity present in breeding programs is decisive for increasing genetic gain and rapidly responding to changes in breeding objectives. In this work, a set of 472 representative lines from the Argentine National Institute of Agricultural Technology (INTA) temperate maize (Zea mays L.) breeding program was genotyped with a mid-density DArTag panel to assess genetic diversity, degree of linkage disequilibrium (LD) and population structure. In addition, a breeding core set that captured the program’s genetic diversity was selected to facilitate future functional studies.ResultsAfter curation of the genotypic matrix, 461 lines and 2,199 single nucleotide polymorphisms (SNPs) remained in the analysis. SNPs were highly informative, with an average polymorphic information content of 0.32. Inbred lines exhibited low heterozygosity (1.6% average) and kinship coefficients indicated low redundancy among them. Rogers’ genetic distances (RD) presented values greater than 0.35 in 92.6% of the comparisons, showing high allelic diversity. LD decay was estimated to be 158,609 bp at r2 half-decay distance. Two levels of population structure were detected using Bayesian inference at k = 2 and k = 12, mainly pinpointing Argentine Flint (AF) and Iowa Stiff Stalk Synthetic (BSSS) origins. These results were consistent with both Principal Component Analysis and cluster analysis based on RD. Additionally, a breeding core set (n = 138) that retained 100% of alleles (allele coverage, CV = 1) present in this panel of INTA breeding lines was selected using Core Hunter 3. Core set representativeness was validated through multiple quality metrics, including a low mean difference percentage (MD=0.11%), a high coincidence rate of range (CR = 100%), and appropriate variance difference percentage (VD = 30%) and variable rate of coefficient of variation (VR = 114%).ConclusionsImplementation of mid-density SNP genotyping enabled the characterization of a public panel of temperate maize inbred lines, revealing an intricate population structure with high genetic diversity and low redundancy. These findings provide a foundational framework for germplasm reorganization and the potential development of genomic prediction models. The selected multipurpose breeding core set effectively preserved the program’s allelic diversity and represents a versatile genomic resource for future functional studies.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12870-026-08935-7. Background Maize (Zea mays L.) is the leading global staple cereal with more than 1 billion metric tonnes produced by more than 145 countries per year [1, 2]. Argentina ranks among the top maize producers, with significant contributions to the global market [3]. It is also a major global maize exporter, contributing to global food security by supplying external demands of other countries [4, 5]. The maize breeding program at the National Institute of Agricultural Technology (INTA) Pergamino Agricultural Experiment Station (EEA Pergamino) is a long-standing public temperate maize breeding program in Argentina. Over more than 70 years of continuous work, the program has developed, selected, and released inbred lines adapted to the temperate productive region of the country [6]. Notably, these public inbred lines have been widely used in early flint-based hybrids and have contributed to the base germplasm of private breeding programs in Argentina. INTA EEA Pergamino inbred lines have been developed using germplasm from diverse sources, including native Argentine Flint (AF) germplasm and, to a lesser extent, exotic germplasm from other regions of the world [7–9]. AF germplasm is characterized by its hard and colored endosperm, which makes it suitable for the dry milling industry, and by its tolerance to biotic and abiotic factors [6]. AF inbred lines exhibit good combining ability with North American yellow dent materials, from both the Reid Yellow Dent (RYD) and Lancaster Sure Crop (LSC) heterotic groups, a pattern defined as “AF by US Dent” [6]. In addition, other heterotic patterns (such as “Flint by Flint”) were explored, which combined with the integration of a variety of germplasm sources (including selfing of commercial hybrids) generated a complex pattern of relationships among lines in the breeding program. However, the exploration of these heterotic patterns remained only partial and was not systematically deepened. Instead, breeders recurrently developed inbred lines using the entire available germplasm. Therefore, a concerted effort to formally characterize the population structure and define distinct heterotic groups remains a critical challenge. Population structure in plant breeding programs is strongly shaped by the existence of diverse genetic origins and the strategic crossing of specific parental lines, which generates kinship relationships among breeding lines [10, 11]. This structure is even more complex in breeding programs that exploit heterosis, since the development of heterotic groups is crucial for developing commercial hybrids [12]. Several studies have indicated that predictive ability within and between families can vary based on the genetic population structure, making its study of paramount importance when implementing genomic selection (GS) strategies in a breeding program [13–15]. The population structure of a set of inbred lines can be analyzed through the implementation of different multivariate techniques that rely on SNP genotyping data [16]. Among these, Pritchard et al. [10] developed a clustering model based on a Bayesian approach, implemented in the STRUCTURE software. This method assumes an admixture model in which there are k subpopulations, each characterized by a specific allele frequency at each locus. Using multilocus marker data, individual inbred lines are assigned probabilistically to a single population or partitioned proportionally among multiple ancestral groups (Q-values) to maximize Hardy–Weinberg and linkage equilibrium for all loci [10]. Complementarily, Evanno et al. [17] proposed an ad hoc statistic, ΔK, to identify the number of subpopulations that best fits the analyzed population. Another method for inferring population structure is Principal Component Analysis (PCA), which partitions the total variance of multilocus marker data into a limited number of new uncorrelated variables, called principal components (PCs). These PCs can be used to construct two- or three-dimensional scatter plots, where the geometric distances between inbred lines can identify possible groups or subpopulations [16]. Lastly, a third technique for evaluating population structure is the use of hierarchical clustering methods based on genetic distances, among which the Unweighted Pair Group Method with Arithmetic Mean (UPGMA) is the most widely used [16]. These clustering methods enable the construction of dendrograms that help identify groupings of lines and assess the genetic diversity of a given germplasm panel. These complementary multivariate approaches have been widely applied to characterize genetic diversity and population structure in temperate and tropical maize improvement programs worldwide, thereby supporting more informed breeding decisions [18–23]. Genome-wide SNP genotyping also facilitates the selection of breeding core collections or core sets, which are subsets of lines (accessions) designed to capture the genetic diversity of a larger collection [21, 24]. Since breeding core sets maximize variation while minimizing redundancy [25], breeders can efficiently characterize and utilize germplasm without evaluating every accession [26]. Breeding core sets can be selected for different objectives, including multipurpose subsets that maximize representativeness [27] and trait-customized subsets that represent the range of genotypes/phenotypes maximizing the diversity of a trait of interest, such as salt tolerance [28–30]. Advances in high-throughput genotyping and phenotyping have driven the development of algorithms for efficient core set selection. Among these, Core Hunter 3 [31] stands out as a versatile tool that integrates genotypic and phenotypic data to optimize representativeness and diversity simultaneously through multi-parameter optimization of metrics such as genetic distance and allelic richness. Core Hunter 3 can select multipurpose subsets by minimizing the average distance between each accession of the breeding panel and the nearest entry (A-NE) in the core set to maximize diversity and representation [31]. Since 1992 INTA EEA Pergamino has partnered with private seed companies through the INTA-Semilleros CVT collaborative agreement, which aims to strengthen the maize breeding program by supporting the development and improvement of inbred lines [6]. In turn, these lines are made available to these private breeding programs for the creation of new breeding populations, the introgression of important traits, or their direct use in commercial hybrids. As part of this agreement, several attempts have been made to characterize the genetic diversity and population structure of INTA EEA Pergamino temperate germplasm. These previous studies examined different subsets of inbred lines —each representing only a portion of the program’s current diversity—and employed different marker technologies [8, 9, 32]. First, a panel of 111 inbred lines was characterized using 74 simple sequence repeats (SSRs) and pedigree records [8, 9]. This characterization revealed the presence of four subpopulations, identifying two subpopulations with mainly dent combining ability and two others with mainly flint combining ability [9]. In addition, a second (overlapping) panel of 216 inbred lines was characterized using 36,665 SNP markers [32]. This latter characterization revealed broad genetic variability in the breeding program, identifying a primary level of population structure consisting of three subpopulations, from which a secondary level of nine subpopulations derived. Briefly, at k = 9 AF lines were assigned to five distinct subpopulations that differentiated from dent and semi-dent clusters and two additional groups that comprised lines developed from AF by dent crosses and lines developed from self-pollination of commercial hybrids [32]. The present study implemented a standardized mid-density SNP panel, which offers a practical and cost-effective advantage for routine breeding applications compared to the higher-density platform used in previous research. This approach enabled the evaluation of a much larger and comprehensive collection of inbred lines, bringing together a panel of 472 released and advanced breeding lines. The study had two main objectives: (i) to assess the genetic diversity, extent of linkage disequilibrium (LD) and population structure of this breeding panel and (ii) to select a multipurpose breeding core set that captures the genetic diversity and representativeness of the program. The molecular characterization of the breeding panel will provide insights for reorganizing and better exploiting the existing and working germplasm while the multipurpose core set will facilitate future functional studies. Methods Plant materials Four hundred and seventy maize inbred lines developed by INTA EEA Pergamino for adaptation to the temperate environments of Argentina were used in this study. These inbred lines represent the current genetic diversity of this maize breeding program. In addition, two US dent inbred lines (B73 and B100) were included as reference materials for a total number of 472 inbred lines. SNP genotyping and data filtering Four fresh leaf tissue discs were collected from two-week-old seedlings of each line and placed in 96-well plates for subsequent freeze-drying. Freeze-dried samples were sent to Intertek Australia for DNA extraction and SNP genotyping using the Maize DArTag 3.3 K EiB (2.0) panel from the CGIAR-CIMMYT Excellence in Breeding (EiB) platform [33]. This mid-density panel consists of 3,305 SNPs distributed across the 10 maize chromosomes, with an average marker density of one SNP every 0.72 Mbp. Curation of the genotypic matrix was performed with the ASRgenomics v1.1.4 R package [5, 34]. Eight hundred and seventeen SNPs with more than 20% missing data and 283 SNPs with minor allele frequencies (MAF) lower than 0.05 were discarded, as well as 11 lines exhibiting more than 15% heterozygosity. Both reference lines (B73 and B100) successfully passed these quality filters and were retained for downstream analyses. Six SNPs that could not be assigned to a specific chromosomal location on the physical map were also removed. In summary, 461 inbred lines and 2199 SNPs remained in the genotypic matrix for diversity analysis of the breeding panel (Table S1). Allele frequencies, MAF, observed heterozygosity (Ho), expected heterozygosity (He), Nei´s genetic diversity (H), polymorphic information content (PIC), and missing data were estimated for the filtered matrix with snpReady R package [35]. The SNP density plot was generated using the rMVP R package [36] using a 1 Mbp sliding window. Linkage disequilibrium Linkage disequilibrium (LD) was estimated by calculating the squared Pearson correlation coefficient (r2) values for all pairwise comparisons of SNP markers within each chromosome [37]. This analysis was performed in TASSEL v5.2.93 [38] using a 50 SNP sliding window, no inter-marker distance filtering was used. LD decay was plotted against the physical distance (bp) between pairs of linked SNPs across the ten maize chromosomes using R Studio [29]. The average LD decay distance was evaluated at different r2 thresholds (r2 = 0.1, r2 = 0.2, half-decay). Half-decay was defined as the physical distance at which r2 drops to half of its maximum value. Population structure and diversity analysis The genetic population structure of the breeding panel was first assessed using the Bayesian clustering method implemented in STRUCTURE v2.3.4 [10]. For this analysis, an admixture model was chosen, assuming that each inbred line may have complex ancestry and could have inherited fractions of its genome from ancestors belonging to different subpopulations or clusters (k). STRUCTURE was run under the assumption of standard unlinked loci and correlated allele frequencies [39]. The analysis was performed for k values from one to 15, with 10 independent runs per k, each consisting of 10,000 burn-in iterations followed by 10,000 Markov Chain Monte Carlo (MCMC) iterations. This iteration scheme was selected to ensure adequate parameter convergence while maintaining computational efficiency. For each repetition, the natural logarithm of the probability of the data (LnP(D)) was obtained. The number of k that best fitted the breeding panel was determined through the method proposed by [17], which uses the ad hoc statistic ΔK. The estimation of ΔK was conducted with the CLUMPAK tool [40]. Once the optimal number of k for the breeding panel was determined, a final STRUCTURE run was performed for that k value with an extended run period of 100,000 iterations and 100,000 subsequent MCMC iterations to obtain membership coefficients (Q). Lines with Q-values lower than a threshold level (Q < 0.6 for k = 2, Q < 0.5 for k = 12) were assigned to a Mixed cluster [19, 23, 41, 42]. Population structure was further examined by PCA using TASSEL v5.2.93 software [38]. Scatterplots of the first four PCs were generated to identify clustering patterns among individuals. To study the genetic diversity of the breeding panel, a distance matrix was generated by calculating Rogers´ genetic distances (RD) [43] for each pair of lines based on SNP allele frequencies. From this matrix, a UPGMA dendrogram was constructed and plotted in MEGA v11 [44] to explore clustering patterns and evaluate their correspondence with the population structure methods described above. No bootstrap support was calculated, as the analysis was intended for exploratory visualization rather than a formal assessment of branch stability. A relative kinship (K) matrix among inbred lines was obtained using ASRgenomics v1.1.4 R package [5, 34] according to VanRaden´s method 2 [45]. This K matrix is constructed as the sum of centered and variance-standardized SNP genotypes, which measures the degree of relatedness among pairs of lines relative to the average similarity of the current genotyped population. Selection of a breeding core set Core Hunter 3 [31] was used to select a multipurpose breeding core set optimizing the Accession-To-Nearest-Entry (A-NE) objective and applying a 30% selection intensity (n = 138). The core set included 19 lines manually pre-selected based on their historical importance as key founders and testers of INTA EEA Pergamino temperate maize breeding program. The filtered genotypic matrix (461 lines, 2,199 SNPs) and a phenotypic matrix composed of Best Linear Unbiased Predictors (BLUPs) of three phenotypic traits (plant height -PH-, ear insertion height -EH- and days to flowering -DTF) were used to calculate Modified Rogers (MRD) and Gower distances, required for core set selection. BLUPs were obtained by fitting linear mixed models with the lme4 R package via the META-R interface [46]. In these models, genotype, environment, and their interaction were considered as random effects to account for multi-environmental variance. Phenotypic data was collected during the 2019/2020 growing season in three environments: two planting dates at Fontezuela (Buenos Aires, Argentina) and one in Las Vertientes, (Córdoba, Argentina). Field trials followed a randomized complete block design with two replicates. Selection and evaluation of the multipurpose core set were performed using the coreHunterData(), sampleCore(), and evaluateCore() functions. Allele coverage (CV) was used to quantify the percentage of breeding panel alleles retained in the selected core [31]. Four metrics were used to evaluate the quality of the selected core set: mean difference percentage (MD), variance difference percentage (VD), coincidence rate of range (CR) and variable rate of coefficient of variation (VR) as described in [47]. A core set is considered representative of the whole collection if MD is lower than 20%, VD is small enough, CR is greater than 80% and VR is large enough [47]. Nei's genetic diversity (H), observed heterozygosity (Ho), expected heterozygosity (He), and inbreeding coefficient (F) were estimated using the popgen() function in the snpReady R package [35]. Membership assignment of each line to different subpopulations was based on the probability determined at k = 12 (Q > 0.5). The Chi-square goodness-of-fit test was performed using the chisq.test() function to compare subpopulation frequencies (k = 12) between the core set and the original breeding panel (p < 0.01). LD and K for the core subset (n = 138) were calculated as previously stated. Results SNP marker statistics The Maize DArTAg 3.3 K EiB (2.0) panel was used to genotype 472 public inbred lines. After filtering the SNP data, 2,199 informative SNPs were retained for subsequent analyses, with an average MAF of 0.32 and an average PIC of 0.32 (Fig. S1). Of these markers, a total of 250, 219, 246, 235, 279, 195, 201, 234, 184, and 156 SNPs were mapped on chromosomes 1 to 10, respectively, with an average of 219.9 markers per chromosome and an average density of one SNP every 1.09 Mbp (Fig. 1).Fig. 1SNP Density Plot. Chromosome-wide SNP density plot displaying the number of SNPs within 1 Mb window size. The horizontal axis shows the chromosome length (Mb); different colors depict SNP density SNP Density Plot. Chromosome-wide SNP density plot displaying the number of SNPs within 1 Mb window size. The horizontal axis shows the chromosome length (Mb); different colors depict SNP density The level of heterozygosity of the 472 inbred lines, calculated with the selected 2,199 SNPs, showed an average of 2.1%, ranging from 0.3% to 39.1%. As expected for inbred material, 92.5% of the lines exhibited heterozygosity levels below 5%. After filtering lines that exceeded the heterozygosity threshold, the resulting genotypic matrix (Table S1) consisted of 461 individuals and 2,199 markers with an average of 2.52% missing data and 1.6% heterozygosity, based on which all downstream analyses were performed (Table S2; Table S3). LD Decay LD across the genome for the entire breeding panel, measured as r2, presented a range between 0 and 1 and declined steadily with increasing physical distance between SNP markers (Fig. 2; Table S4). LD decay was estimated to be 158,609 bp at r2 half-decay distance (Fig. 2; Table S5). The average LD between SNPs located at distances lower than 1,000 bp was 0.52. At larger distances between SNPs, LD decayed rapidly reaching an average of r2 = 0.2 at 210,549 bp (Table S5).Fig. 2Genome-wide LD decay in the 461-breeding panel. Scatter plot of LD estimates of intrachromosomal pairwise loci against physical distance (Mbp). LD decay distance at r2 half-decay (blue horizontal line) was estimated to be 158,609 bp Genome-wide LD decay in the 461-breeding panel. Scatter plot of LD estimates of intrachromosomal pairwise loci against physical distance (Mbp). LD decay distance at r2 half-decay (blue horizontal line) was estimated to be 158,609 bp The average LD decay distance was generally consistent across the ten chromosomes (Table S5), but relatively faster for chromosome 2 (71,006 bp) and slower for chromosome 8 (414,797 bp). At the mean distance between markers (1.09 Mbp), the observed LD was 0.072. Population structure of the breeding panel The average values of LnP(D) resulting from the clustering method based on the admixture model and correlated allele frequencies implemented in STRUCTURE steadily increased with the increase in the number of k up to k = 12, a value above which LnP(D) showed a sharp decline, with the largest increase observed between k = 1 and k = 2 (Fig. 3A). The ad hoc statistic ΔK also reached its maximum at k = 12, with a second peak of lower value at k = 2 (Fig. 3A). Taken together, these results suggest a primary structure level of two subpopulations, and a secondary, more detailed structure of twelve subpopulations derived from the first (Fig. 3B-C). The assignment of inbred lines to subpopulations based on increasing thresholds of Q value (0.5, 0.6, 0.7, or 0.8) resulted in a higher proportion of unassigned lines. A threshold of Q = 0.6 was used for k = 2, but it was relaxed to Q = 0.5 for k = 12 since this larger number of subpopulations resulted in fragmented membership probabilities (Table S6). Using these thresholds, the proportion of inbred lines assigned to a subpopulation was 83.5% for k = 2 and 74.6% for k = 12 (Fig. 3B-C, Table S6).Fig. 3Population Structure of INTA EEA Pergamino Temperate Maize Breeding Program based on 2,199 SNPs. A. Plot of LnP(D) and ad hoc statistic ΔK estimated for k ranging from 1 to 14. B. and C. Population structure for the 461 inbred lines at k = 2 and k = 12, respectively. Each inbred line is represented by a single vertical line that is partitioned into k-colored segments. Subpopulation names are shown on the x axis Population Structure of INTA EEA Pergamino Temperate Maize Breeding Program based on 2,199 SNPs. A. Plot of LnP(D) and ad hoc statistic ΔK estimated for k ranging from 1 to 14. B. and C. Population structure for the 461 inbred lines at k = 2 and k = 12, respectively. Each inbred line is represented by a single vertical line that is partitioned into k-colored segments. Subpopulation names are shown on the x axis The assignment of inbred lines to two subpopulations (k = 2) revealed a first group comprising 286 lines predominantly originated from flint or semi-flint germplasm, with the most relevant sources being composites, synthetic populations, and local varieties representing AF germplasm, Argentine commercial hybrids of flint or semi-flint type, and combinations of these materials with other exotic sources such as US or Canadian dented germplasm. This first group will be referred to as AF. The second group comprised a total of 99 inbred lines with fundamental ancestry from the BSSS heterotic group (Iowa Stiff Stalk Synthetics). This second group will be referred to as BSSS. The remaining 76 inbred lines were classified as Mixed, showing Q values between 0.4 and 0.6 for both AF and BSSS, consistent with their genetic backgrounds, which in many cases involved combinations of germplasm from both subpopulations. The assignment of inbred lines to twelve subpopulations (k = 12) further explained the genetic structure of the breeding panel, with different groups belonging to the previously described subpopulations (Table S6). Seven subpopulations (1–7) comprised inbred lines previously assigned to the AF subpopulation. Two subpopulations (8–9) were formed by inbred lines previously assigned to the BSSS subpopulation. Finally, the remaining three subpopulations (10–12) had members previously assigned to the Mixed group. The assignment of each inbred line to different subpopulations defined by the STRUCTURE analysis for k = 2 and k = 12 is detailed in Table S6. A brief description of the twelve subpopulations follows. The first and largest subpopulation (AF-I) with 143 lines, was predominantly of AF ancestry. Major contributions came from Hard Red Synthetic, Cuarentin Synthetic, Synthetic 34 (including crosses with tropical lines from INTA Leales), the Argentine-Caribbean Composite, Composite I, and Composite II [9]. Founder lines of the breeding program derived from local varieties such as P21, P1338, ZN6, A1, and LP582 were also grouped here, with LP582 contributing multiple derivatives. A second subpopulation (AF-P465) includes P465, a line derived from a local flint variety, and 15 P465 derivatives. The third subpopulation (AF-LP561) of 25 lines is centered on LP561, a line developed from the synthetic population Flint R4PC selected for resistance to the Mal de Río Cuarto virus. Most of the remaining lines are derived from LP561 through various crosses. The fourth subpopulation (AF-B106) with 9 members, includes derivatives of B106 (BS26 origin, LSC ancestry) obtained from crosses with LP561 and the commercial hybrid Dekalb DK615. The fifth subpopulation (AF-CD) with 12 member lines, originated from a synthetic population derived from AF lines crossed with Canadian dent (CD) materials. The sixth subpopulation (AF-LP580) with 16 lines, is centered on LP580 and lines derived from the commercial hybrid Cargill 280. The seventh subpopulation (AF-HET) showed no common ancestry among its 7 members, consistent with the observed range of Q values (0.53–0.60) and low Fst value (0.074). Two subpopulations (8th-9th) were associated with BSSS: BSSS-I (53 lines) and BSSS-II (24 lines). BSSS-I includes lines derived from BS13 and B110 (developed at Iowa State University from BS13), improved by recurrent selection for combining ability with AF testers. It also includes derivatives of B110, crossed with commercial hybrids (Nidera, Dekalb) or LP561. BSSS-II contains B73, its derivatives, and lines developed by selfing commercial hybrids (Dekalb, Nidera). Three subpopulations were associated with the previously described Mixed group (k = 2). The tenth subpopulation, SDS, consisted of 15 lines mainly derived from semi-dent synthetic populations that included early semi-dent and selfing of Pioneer hybrids. The eleventh subpopulation, B113-AH (16 lines) traced back to B113 (from BS11), crossed with commercial hybrids AX707 and AX599 (Nidera). The twelfth and final subpopulation, centered on B100 and named after it, contains eight members: B100 and seven lines derived from crosses with AX707, with Q values ranging 0.51 to 0.66. A total of 117 lines (26.2%) remained in the Mixed group at k = 12, which comprised materials of diverse origins, including derivatives of commercial hybrids, flint by dent crosses, and lines developed from composites or synthetic populations associated with the other subgroups. Principal component analysis The first four PCs (PC1–PC4) explained 6.8%, 3.5%, 2.9%, and 2.6% of the variation among inbred lines for the 2,199 SNPs, respectively, and together captured 15.8% of the total variation (Fig. 4). PC1 divided inbred lines into a clustering pattern consistent with that observed in STRUCTURE at k = 2, with lines belonging to the BSSS subpopulation at one extreme, those included in the AF subpopulation at the opposite extreme, and intermediate positions for the lines included in the Mixed group (Fig. 4A). The high correspondence between PC1 and the results obtained with STRUCTURE at k = 2 was supported by a Pearson correlation of 0.99 (Fig. S2) between this PC and the AF Q-value (equivalent to one minus the BSSS Q-value).Fig. 4Principal component (PC) scatter plots of 461 inbred lines based on 2,199 SNPs. Different colors indicate the assignment to subpopulations estimated by STRUCTURE analysis at k = 2 (A, B) and k = 12 (C, D) Principal component (PC) scatter plots of 461 inbred lines based on 2,199 SNPs. Different colors indicate the assignment to subpopulations estimated by STRUCTURE analysis at k = 2 (A, B) and k = 12 (C, D) The scatter plot of PC1 and PC2, which together explained 10.3% of the variation, facilitated the separation of certain subgroups associated with the AF subpopulation. These included inbred lines assigned to AF-LP580, AF-CD, AF-I, and AF-P465 (with some overlap between the latter two), as well as subpopulations AF-LP561 and AF-B106, which showed a high degree of overlap (Fig. 4C). Regarding groups associated with the BSSS subpopulation, PC3 and PC4 distinguished the BSSS-II, SDS, B113-AH, and B100 groups from the rest of the panel (Fig. 4D). Similarly, the scatter plot of PC1 and PC4, which jointly explained 9.4% of the variation, separated individuals assigned to BSSS-I from the remaining lines (Fig. S3). In contrast, the separation of AF-LP561, AF-B106, and AF-HET subpopulations across the first four PCs was less distinct compared to the other subpopulations. As expected, inbred lines assigned to the Mixed group at k = 12 were generally located in intermediate positions between subpopulations (Fig. 4). Cluster analysis An UPGMA dendrogram based on RD among the 461 inbred lines was generated and colored according to STRUCTURE results at k = 2 (Fig. 5A) and k = 12 (Fig. 5B). Two main branches were identified. One branch clustered AF-I, AF-P465, AF-LP561, and AF-B106, while the other branch further subdivided into two clusters: the first included AF-CD and AF-LP580 (AF at k = 2) along with SDS (Mixed at k = 2), and the second distinguished B100 and B113-AH (Mixed at k = 2) from BSSS-I and BSSS-II (BSSS at k = 2). Most inbred lines clustered consistently within their respective subpopulations, whereas AF-HET lines were scattered across the dendrogram. As also observed in the PCA, lines assigned to the Mixed group at k = 12 were generally located at intermediate positions between subpopulations. Fig. 5UPGMA dendrograms of 461 inbred lines based on Rogers’ genetic distances. Different colors indicate the assignment to subpopulations estimated by STRUCTURE analysis at k = 2 (A) and k = 12 (B). The dendrogram was generated without bootstrap support UPGMA dendrograms of 461 inbred lines based on Rogers’ genetic distances. Different colors indicate the assignment to subpopulations estimated by STRUCTURE analysis at k = 2 (A) and k = 12 (B). The dendrogram was generated without bootstrap support Genetic distance and relative kinship RD among inbred lines ranged from 0 to 0.53, with an average of 0.40 and 92.6% of comparisons above 0.35 (Fig. 6A). K coefficients ranged from –0.53 to 2.34 and showed a continuous decline in frequency as K increased (Fig. 6B). More than half of the comparisons (57.6%) showed below population average relatedness (K < 0), suggesting low redundancy among lines in the breeding panel (Fig. 6B).Fig. 6Genetic Diversity and Relative Kinship. Frequency distribution of pairwise (A) Rogers’ genetic distances (RD) and (B) relative kinship coefficients (K) for 461 inbred lines based on 2,199 SNPs Genetic Diversity and Relative Kinship. Frequency distribution of pairwise (A) Rogers’ genetic distances (RD) and (B) relative kinship coefficients (K) for 461 inbred lines based on 2,199 SNPs Analysis of average RD values among the twelve subpopulations (k = 12) revealed the greatest genetic distances (0.44) between AF-I and BSSS-I, AF-I and BSSS-II, AF-P465 and B100, AF-P465 and B113-AH, and between AF-B106 and BSSS-II (Table S7). The smallest RD (0.32) was observed between AF-LP561 and AF-B106. Average kinship (K) values across subpopulations ranged from –0.19 to 0.39, with the highest between AF-LP561 and AF-B106. Within subpopulations, mean K values ranged from 0.18 in AF-I to 1.3 in AF-LP580 (Table S7). Breeding core set selection A multipurpose core set that maximizes the genetic variability and representativeness of INTA EEA Pergamino temperate maize breeding program was selected using genotypic and phenotypic information (Table S1; Table S8). This selected core set (n = 138, Table S9) retained 100% of the alleles present in the breeding panel (allele coverage, CV = 1), showing average values of expected and observed heterozygosity (He = 0.40, Ho = 0.018), inbreeding coefficient (F = 0.96), genetic variability (additive genetic variance, Va = 882.75; dominance genetic variance, Vd = 378.81) and population genetic structure (Chi-square test, p = 0.90, α = 0.01) similar to those of the complete breeding panel (Fig. 7A; Table S10). The drop in LD for the multipurpose core set was estimated at 146,743 bp at r2 half-decay distance (Fig. S4; Table S11). LD decay distance and frequency of r2 values per chromosome exhibited similar patterns for the multipurpose core set compared to the breeding panel (Fig. S5; Fig. S6). In addition, this selected core set maintained a similar distribution of PH (Chi-square test, p = 0.52, α = 0.01), EH (Chi-square test, p = 0.30, α = 0.01) and DTF BLUPs (Chi-square test, p = 0.18, α = 0.01) relative to the breeding panel (Fig. 7B). The high quality of the selected core set was also reflected by low MD and VD (0.11 and 30%, respectively) and high CR and VR (100% and 114%, respectively). Descriptive statistics of the evaluated phenotypic traits for the breeding panel and the selected core set can be found on Table S12.Fig. 7Multipurpose core set representativeness assessment. A. Distribution of the 138 inbred lines selected in the core set (CS, filled dots) in the PC1 vs PC2 scatter plot of the complete breeding panel (BP, empty dots) genotyped with 2,199 SNPs. Colors are assigned based on subpopulations estimated by STRUCTURE analysis at k = 12. B. Frequency distribution of phenotypic traits (PH, EH, DTF) in BP (pink) and CS (light blue). Chi square (X2) test results are shown on the upper left corner of each figure. PH: plant height. EH: ear insertion height. DTF: days to flowering Multipurpose core set representativeness assessment. A. Distribution of the 138 inbred lines selected in the core set (CS, filled dots) in the PC1 vs PC2 scatter plot of the complete breeding panel (BP, empty dots) genotyped with 2,199 SNPs. Colors are assigned based on subpopulations estimated by STRUCTURE analysis at k = 12. B. Frequency distribution of phenotypic traits (PH, EH, DTF) in BP (pink) and CS (light blue). Chi square (X2) test results are shown on the upper left corner of each figure. PH: plant height. EH: ear insertion height. DTF: days to flowering Discussion In this study, the genetic population structure of a panel of inbred lines representing the Temperate Maize Breeding Program at INTA EEA Pergamino was comprehensively characterized using three complementary multivariate techniques. These approaches facilitated the identification of distinct subpopulations and provided quantitative evidence supporting their existence. Consistent with previous reports on maize population structure [19, 41, 48, 49], the log-likelihood of the data (LnP(D)) derived from STRUCTURE analyses did not definitively indicate the optimal number of subpopulations, as it increased monotonically with the number of clusters, up to k = 12 in our analysis. Conversely, the ΔK statistic proposed by [17] effectively identified two hierarchical levels of population structure. A primary level composed of two subpopulations, indicated by the second highest ΔK peak and a second level composed of twelve groups, where ΔK reached its maximum. This multilevel structure provided a comprehensive understanding of the breeding program complexity, with the first level reflecting the exploitation of the “AF by US Dent” heterotic pattern [6, 7] whereas the second level of twelve subpopulations captured the extensive diversity arising from the integration of multiple germplasm sources through structured crossing schemes within and between major groups. Similarly, Badu-Apraku et al. [19] reported the identification of a first ΔK peak reflecting broad population structure in a panel of maize lines adapted to sub-Saharan Africa and a second ΔK peak that subdivided the lines into smaller groups, better revealing their relationships and diversity. The threshold for assigning individuals to subpopulations at k = 2 was set at Q = 0.6, in accordance with values commonly employed in molecular characterization studies [19, 41, 48, 49]. Relaxation of this threshold to Q = 0.5 for k = 12 followed the recommendations of Van Inghelandt et al. [42] who argued for lower stringency in high-diversity contexts. If a higher Q threshold such as 0.6 had been applied at k = 12, a greater proportion of individuals (74.6%) would have been classified as Mixed, making the results more difficult to interpret. A threshold of Q = 0.5 has been also used in other studies [22, 23]. Results from PCA and cluster analysis corroborated the STRUCTURE findings. The primary division into two subpopulations identified at k = 2 was reflected by the first PC (PC1), which exhibited an almost perfect correlation with the STRUCTURE assignment, reinforcing a clear differentiation between the AF and BSSS germplasm. The second PC (PC2) predominantly captured differences within the AF cluster, driven by the internal diversity of the local germplasm sources that originated it, the development of heterotic subgroups based on “Flint by Flint” crosses during the initial phase of the breeding program, and the targeted introgression of foreign germplasm, among other factors. The separation of AF-I and AF-P465 subpopulations appears aligned with the exploitation of this heterotic pattern, with testers used in the development of this heterotic pattern [7], such as LP122-2 (AF-I) and LP612 (AF-P465), emerging as anchor lines with high membership values for these groups (Q > 0.98). The influence of external germplasm introduction was also apparent, particularly in subpopulations like AF-LP561 and AF-B106, which combined local germplasm with northern dent sources such as LSC, and the AF-LP580 subpopulation, which constitutes an example of the utilization of commercial hybrids with desirable traits as a source of germplasm. The inclusion of the third (PC3) and fourth PCs (PC4) facilitated the discrimination of inbred lines assigned to the twelve subpopulations, except for subpopulations AF-LP561 and AF-B106, that appear together, and AF-HET. The inability to discriminate all the subpopulations could be due at least in part to the unexplained variance beyond the first four PCs (15.8%). Furthermore, the low proportion of variation explained by these PCs indicates a highly intricate genetic structure of the breeding panel, consistent with prior observations by [19], who reported PCA’s limitations in resolving secondary structure levels. The dendrogram constructed using the UPGMA method from RD effectively grouped inbred lines in the twelve STRUCTURE-defined subpopulations, except for those inbred lines belonging to the AF-HET subpopulation. Consistent with its role as an exploratory visualization tool, the dendrogram was presented without bootstrap support; therefore, branch patterns were interpreted as indicative of broad clustering trends rather than definitive phylogenetic relationships. Interestingly, the dendrogram’s main branches indicated the presence of three large clusters, not perfectly mirroring STRUCTURE’s primary structure, but illustrating a gradient from AF to BSSS, with AF-I at one end and BSSS-I at the other. Subpopulations containing inbred lines with predominantly flint and semi-flint endosperm clustered on one branch, whereas the other branch contained a first large group of subpopulations containing inbred lines with intermediate, semi-flint, and semi-dent phenotypes and another group that clustered subpopulations with dent and semi-dent endosperms. Overall, the current study showed strong agreement with previous molecular characterization studies conducted on smaller and partially overlapping panels from this breeding program. Olmos et al. [9] molecularly characterized a panel of 111 lines with 74 SSR markers identifying four subpopulations. Their inbred line assignments closely matched those observed in the present study for the subpopulations AF-I, AF-P465, SDS, and BSSS-I, which they referred to as ACDS, P465, LP299-2, and BS13-BSSS, respectively. Later, Torrent evaluated a more extensive panel of 218 lines using the Illumina MaizeSNP50 beadchip, revealing a population structure consisting of nine subpopulations [32]. Of these, five correspond to the AF-P465, AF-LP580, AF-CD, SDS, and BSSS-I subpopulations identified in our study. Inbred lines belonging to the remaining four subpopulations previously described were not resolved in our study and clustered together in the AF-I subpopulation. Differences between these results and those of previous studies may be attributed to several factors, including the larger number of genotypes evaluated, variations in the proportion of representative genotypes within each subpopulation, differences in genotyping density, and/or the type of molecular markers employed. The development of heterotic groups in maize relies on long-term selection efforts. In North American temperate maize, heterotic groups originated nearly a century ago following pioneering research in heterosis and inbreeding [50]. In Argentina, however, breeding efforts at INTA EEA Pergamino primarily focused on generating broad-based composites and improving them via recurrent selection, with limited emphasis on developing formal heterotic groups [8, 9]. The importance of establishing heterotic groups has increased significantly only in recent decades, with continuous exploration and integration of new germplasm sources. This complex breeding history and strategy is clearly reflected in the intricate genetic structure revealed in this study, showing a general AF vs. BSSS pattern but without a complete assignment of all inbred lines to either of these two groups. Similarly, tropical maize breeding programs where no clear heterotic patterns are observed were also originally focused on the development of broad-based populations resulting in inbred lines with mixed genetic backgrounds [19, 20]. For these breeding programs, molecular analyses facilitated the design of breeding strategies for future refining of heterotic groups. Although the 461 inbred lines were developed through multiple generations of self-fertilization, residual heterozygosity was detected, suggesting incomplete fixation in a subset of the 461 lines (1.7% average, 0%−15% range). Among the 212,060 pairwise comparisons for the 461 lines, 92.5% showed RD greater than 0.35, while 96.8% had K coefficients below 0.5. This suggests, on one hand, the potential for most inbred lines to provide genetic variability through new alleles to the breeding program and, on the other hand, low redundancy among them. Additionally, the average RD within groups provided insight into the internal genetic divergence of each subpopulation. The analysis of K coefficients revealed a complex familial background among inbred lines of the breeding panel, consistent with the program’s history as previously discussed. Generally, low K coefficients were observed between inbred lines from different subpopulations, showing that the population structure analysis effectively separated inbred lines minimizing relatedness between subgroups. Conversely, K relationships were observed among inbred lines assigned to the same subpopulation. The highest average K coefficient was observed in the AF-LP580 subpopulation, composed of nine full-sib lines derived from the same biparental cross. The lowest average K coefficient was observed in the AF-I subpopulation, the largest and most diverse, as its lines have diverse origins and often lacked pedigree relationships. As expected, an inverse relationship was observed between RD and K coefficient within subpopulations, consistent with findings reported by other authors [41]. The intricate population structure and complex kinship revealed in this study offer valuable insights for the future implementation of GS in INTA EEA Pergamino breeding program. Windhausen et al. [51] demonstrated that predictive ability in structured populations is heavily influenced by differences in the mean performance between subpopulations, rather than solely by within-family genetic variation. Furthermore, research using both simulations [13] and real maize datasets [14, 15] indicated that predictive accuracy can vary substantially when evaluated across the entire population versus within specific clusters. In the context of INTA’s breeding panel, characterized by a gradient from AF to BSSS and 12 distinct subpopulations, these findings suggest that GS models should be validated not only at the whole-program level but also within specific genetic groups. This approach is essential to avoid overestimating accuracy due to population structure (spurious correlations) and to ensure effective genetic gain within each target heterotic group. Previous studies have evaluated the decay of LD in maize across different germplasm collections using various types and densities of molecular markers [48, 49, 52]. Consistent with our findings, previous research has shown that LD decay distances vary among individual maize chromosomes [53]. In addition, Yan et al. [49] found the average LD distance to be higher in temperate (10–100 Kb) compared to tropical and subtropical germplasm (5–10 Kb). This was proposed to result from more intensive recombination and a higher frequency of rare alleles in tropical germplasm [49, 53]. The LD decay in our breeding panel (210 Kb at r2 = 0.2) could be considered slow. However, this estimate should be interpreted with caution, as LD decay values can be influenced by several factors, including marker density, specific filtering criteria, population size, and internal genetic structure [54]. The extent of LD in a population depends on the effective population size (Ne) [55], smaller Ne results in higher kinship and increased LD. Although kinship relationships among most pairs of lines were near zero, some subpopulations, such as AF-LP580, B100, B106, and SDS, displayed significant internal relatedness. This intrapopulation consanguinity could contribute to higher levels of LD. Slower LD decay in our breeding panel could also be the result of its complex population structure, especially in lines with mixed heterotic backgrounds [20]. In this study, selection of a multipurpose breeding core set aimed at retaining maximum diversity while securing representation of the twelve subpopulations previously described. The choice of Core Hunter 3 as the selection method was based on the advantages it provides over other strategies [56]. These included the possibility of simultaneously evaluating genotypic and phenotypic data, applying different selection objectives to maximize diversity parameters, pre-selecting relevant accessions and choosing the final size of the core set [31]. Additionally, Core Hunter 3 has been successfully used to select core sets in a wide range of species such as grapevine [57], eucalyptus [58], potato [59, 60], wheat [26, 28], walnut [61], and maize [29, 30, 56]. Diversity parameters and statistical analyses indicated that the selected breeding core set (n = 138) preserved the allelic diversity (100%), LD decay pattern, phenotypic diversity for three evaluated phenological traits (MD lower than 20% and CR higher than 80%) and population structure representativeness of the complete panel of INTA breeding lines. While the inclusion of 19 manually pre-selected lines—based on their historical importance—may have influenced the final composition of the subset, all quality metrics confirmed that the overall representativeness of the program’s diversity was maintained. Conclusions The implementation of mid-density SNP genotyping provided a detailed characterization of the genetic diversity, LD, and population structure of INTA EEA Pergamino temperate maize breeding panel. This approach revealed a highly intricate population structure characterized by high genetic diversity and low redundancy, and facilitated the classification of inbred lines with unknown genetic background. Taken together, these findings provide a foundational framework for the strategic reorganization of INTA EEA Pergamino maize germplasm. The comprehensive understanding of the population structure provides a basis for optimizing future crossing schemes within defined heterotic pools. Furthermore, this characterization could support the development and validation of genomic prediction models tailored to the program’s unique genetic architecture. The selection of a multipurpose core set (n = 138) ensures the preservation of the program's total allelic variability in a manageable subset. This core set constitutes a versatile genomic resource specifically optimized for future functional studies—such as GWAS—and high-throughput phenotyping efforts. Ultimately, the insights gained in this study have the potential to contribute to the overall improvement of Argentine maize germplasm, especially through the adoption of INTA EEA Pergamino inbred lines by private breeding programs through INTA-Semilleros CVT. Supplementary Information Supplementary Material 1. Supplementary Material 1. Supplementary Material 2 Supplementary Material 2