The number of mapped SNPs that were polymorphic across 330 genotypes varied from 857 on chromosome 9 to 1,679 on chromosome 1 (Table 1), with an average of 1,252 per chromosome. Average minor allele frequency was 0.192. The physical length of each chromosome varied from 22,947 kb on chromosome 9 to 43,230 kb on chromosome 1 (Supplementary Figure S1), with a total genome size of 373 Mb. As shown in Table 1, the average map distance among adjacent markers (inter-marker interval) varied from 20 to 31 kb, (i.e., there was at least one polymorphic SNP within every 20–31 kb physical interval). The number of polymorphic SNPs within each type of rice out of a total of 15,020 was 13,205 for indica, 11,808 for japonica, 13,601 for NERICA and 10,275 for ARICA (Table 1). Residual heterozygosity per genotype varied from 0.6 to 20.1%, with an overall average of 1.4% (Supplementary Table S1). Approximately 95% of the genotypes were considered genetically pure with <3.2% residual heterozygosity, which is the expected average residual heterozygosity for lines extracted from S₅ or later generations. About 4% (13 genotypes) had residual heterozygosity ranging from 5.5 to 20.1%, which may indicate the need for further selfing/purifying and the possibility of error during seed handling or increase in these genotypes.

Kinship coefficients between pairs of genotypes varied from 0 to 1.96 (on a scale of 0–2), with an average of 1.05 (Supplementary Table S3). Overall, nearly 26 and 66% of the pairs of 330 genotypes had kinship values of ≤0.25 and >0.50, respectively. The proportion of highly related genotypes with kinship values greater than 0.50 varied across the four types of rice, with those belonging to the indica showing the highest (96.0%) and NERICAs with the lowest (65.4%) related pairs of genotypes (Figure 1A). Genetic distance between pairs of the 330 genotypes varied from 0.012 to 0.583, with an overall average of 0.277 (Supplementary Table S4). The 54,285 pairwise comparisons are summarized in Figure 1B, where most pairs had distances ranging from 0.1 to 0.2, but 941 pairs of genotypes (1.7%) had a genetic distance ≤ 0.05, and 88 pairs differed by ≤2% of the total number of alleles used in the analyses.

To identify the most similar pairs of genotypes within each group of rice, we computed genetic distance matrices among pairs of genotypes belonging to the indica, japonica, NERICA, and ARICA genotypes, which varied from 0.016 to 0.623, from 0.020 to 0.692, from 0.014 to 0.644, and from 0.075 to 0.763, respectively (Supplementary Table S4). Genetic distance for 9% of the pairs of 85 NERICAs was ≤0.05, a larger percentage of highly similar pairs compared with the 1.0% within pairs of indica, 3.1% within pairs of japonica, and none within pairs of ARICA (Figure 1B and Supplementary Table S4). The largest proportion of pairs of genotypes with genetic distance values >0.400 was observed within pairs of NERICAs (35.1% of the pairs) as compared with japonica (17.4% of the pairs), ARICAs (18.2% of the pair), and indica (5.6% of the pairs). Genetic distance between pairs of the 19 upland NERICAs varied from 0.017 to 0.215, with NERICA16 and NERICA18 and NERICA8 and NERICA9 differing by ∼2% of scored alleles. Genetic distance among pairs of 66 lowland NERICAs varied from 0.014 to 0.307, with 34 pairs of lowland NERICAs differing by ≤ 2% of alleles used for analyses (Supplementary Table S4).

The cluster analysis generated from the genetic distance matrix grouped the 330 genotypes into two major groups (Figure 2A and Supplementary Figure S2), which agrees with the upland and lowland ecologies (Figure 2B). The first group consisted of 81 genotypes that belong to upland (75) and lowland (6) ecologies. The second group had 7 upland and 243 lowland genotypes. Six genotypes from the upland and 7 genotypes from the lowland ecologies were mis-grouped based on the SNP data (Figure 2B and Supplementary Table S1). The indica, lowland ARICA and lowland NERICA genotypes were clearly separated from the japonica and upland NERICA and upland ARICA genotypes (Figure 2C). The cluster analysis also has shown the clear separation of the groups predicted based on the model-based population structure at K = 2 and K = 3 (Figure 2D), which is described in detail below. We observed groups of genotypes within indica, japonica, or NERICA that appeared to be highly similar. For each type of rice, we used the sphericity index curve in deciding the most informative number of genotypes to be retained and the redundant ones to be removed (Perrier et al., 2003). Using the sphericity index curve, 1 japonica, 11 indica, and 13 NERICAs (all marked in red in Figure 3) showed ≥98% similarity with one or more genotypes and were considered redundant (Supplementary Table S1). We then compared the genotypic data of the original 330 genotypes with the 304 non-redundant genotypes for polymorphism, allele frequency, genotype frequency, and genetic distance. Of the 15,020 SNPs that were polymorphic across the 330 genotypes, 99.8% of the markers (14,985 SNPs) were remained polymorphic within the 304 genotypes after filtering. Both allele and genotype frequencies were identical in both data sets (data not shown). The exclusion of 26 redundant genotypes reduced the number of pairs of genotypes that differed by <2% of the alleles from 88 (Supplementary Table S4) to just 22 (Supplementary Table S5).

In the model-based population structure analysis, the log probability of the data [LnP(D)] and ad hoc statistics ΔK both suggest the presence of two or three possible groups or sub-populations (Figure 4A) that was consistent with the rice growing ecologies (Figure 4B and Supplementary Table S1). At K = 2, a total of 77 genotypes belonging to the upland ecology form group 1, along with 5 lowland genotypes, while the second group consists of 245 genotypes adapted to lowland ecology along with 4 upland genotypes. At K = 3, the first group remained the same as the group membership obtained at K = 2. The first group consisted of 70% of japonica genotypes and 23% upland NERICAs derived from interspecific hybridization between O. glaberrima × O. sativa. The second group consisted of 178 lowland and 4 upland genotypes, of which 81% are indica and 11% are lowland NERICAs. The third group had 67 lowland genotypes of which 69% were interspecific lowland NERICAs and 30% indica (Supplementary Table S1). Results from PCA (Figure 5) demonstrated the presence of two distinct groups corresponding to the lowland and upland ecology, and a subgroup of the same genotypes forming the third STRUCTURE group obtained at K = 3 (Supplementary Table S1).

Single nucleotide polymorphisms diversity statistics indicate that NERICA genotypes were more diverse (higher polymorphism and nucleotide diversity) than ARICA, japonica, and indica genotypes (Table 2). When the analyses were done by dividing NERICAs into two ecologies, however, both polymorphism and nucleotide diversity in the lowland and upland NERICA genotypes were much lower than the indica and japonica genotypes. Groups predicted based on cluster and STRUCTURE analyses followed the pattern based on ecology of most genotypes in the group. Thus, genotypes belonging to the first cluster and STRUCTURE groups, which were predominantly upland japonica, showed lower polymorphism and nucleotide diversity than those genotypes in the second group, which were mostly lowland indica and NERICA. At K = 3, the third group, which primarily consisted of the lowland NERICAs, had the lowest polymorphism and nucleotide diversity as compared with the other two groups; however, this was probably due in part to a smaller sample size. Most of the molecular variation was observed between ecologies and groups predicted based on STRUCTURE and cluster analyses; very little was partitioned between rice types (Table 3).

The partitioning of the overall molecular variance into different hierarchical levels revealed that differences in types of rice, ecologies, and groups predicted based on cluster analysis and the model-based population structure accounted for 41.2%, 57.6–68.9%, and 69.9–76.0% of the total variation, respectively (Table 3). A permutation tests indicated that the proportion of variance attributable for all hierarchical levels were highly significant (p < 0.001). To investigate the extent of genetic differentiation among hierarchical levels, we compared F_ST values between pairs of the different types of rice, ecologies, and predicted groups based on both cluster and the model-based STRUCTURE analyses (Supplementary Table S6). F_ST values were highly variable among the predicted groups based on STRUCTURE at K = 3, with group 1 (upland ecology) showing the highest divergence as compared with both group 2 (0.759) and group 3 (0.834); groups 2 and 3 showed less divergence (F_ST 0.227). We found little differentiation between irrigated and rainfed lowland ecologies (0.015), and between indica and ARICA (0.030); moderate differentiation between NERICA and ARICA (0.075), and between indica and NERICA (0.112). There was very great genetic differentiation between indica and japonica (0.624), between japonica and NERICA (0.449), between japonica and ARICA (0.588), between the lowland and upland ecologies (0.688). In all multivariate analyses, we observed clear population structure between the lowland (irrigated and rainfed lowland) and upland ecologies. To identify SNP markers that contributed to the major genetic differentiation between the lowland and upland ecologies, we used the hierarchical island model to identify loci that may have undergone selection. We found out that approximately 8% of the SNPs (1,197 of the 15,020 SNPs) were significantly (p < 0.05) different between the lowland and upland ecologies (Supplementary Table S7). The 1,197 SNPs were sufficient to clearly separate the three groups (Figure 6) in the same way as predicted based on the model-based STRUCTURE at K = 3 (Figure 4).

Rice is a selfing species with an outcrossing rates ranging from 2 to 5% (Semon et al., 2005). As a result, most genotypes used in the current study were expected to display less than the average residual heterozygosity expected in S₅ generation, which was seen with ∼95% of the genotypes. Only about 4% (13 of 330 genotypes) had residual heterozygosity exceeding 5%, which is higher than expected in the absence of human error, which includes ARICA18 (5.9%), NERICA-L38 (7.4%), and NERICA-L54 (16.3%), eight indica genotypes, and two japonica genotypes (Supplementary Table S1). Residual heterozygosity is common in many rice varieties and has been reported to be useful as sources of genetic variation in parental lines used in developing new segregating populations (Belefant-Miller et al., 2012). However, the availability of genetically pure seeds is one of the important quality control criteria in breeding and seed production, which directly affects grain quality for commercialization (Semagn et al., 2012; Ertiro et al., 2015). Even a small proportion of contamination in seed sources could lead to variation in agronomic performance, reducing the quality of the product for marketing. Rouging off-types and voluntary plants in seed production plots minimizes such issues but incurs additional effort and cost. Maintenance of genetic purity of parental seed sources by reducing residual heterozygosity to less than 5% is an efficient and cost-effective method to maintain quality seed production. Genotypes with residual heterozygosity >5% should be purified if they will be used as trait donors and genome-wide association studies.

Kinship values estimated between pairs of genotypes using either known pedigree relationships or genome-wide molecular markers are useful to understand the extent of relatedness (Eu-ahsunthornwattana et al., 2014). Kinship values close to zero indicate unrelated germplasm, while those close to 0.5 or higher (which around 66% of the pairs of genotypes in this study) refer to full sibs or highly similar germplasm (Dodds et al., 2015). Groups of closely related parents tend to bring redundant genetic value to a breeding program, which is evident in NERICAs due to the repeated use of four O. glaberrima and eight O. sativa parents in developing 66 lowland, and 19 upland NERICAs. In fact, some of the parents, such as TOG5681 and IR64 were extensively used in 63 and 71 crosses, respectively (Supplementary Table S1).

We observed highly variable genetic distances among pairs of genotypes belonging to the ARICA, NERICA, indica, and japonica groups. None of the ARICAs used in the present study shared the same pedigree, which is clearly evident from the huge genetic difference (range 0.075–0.763) observed among pairs of these varieties. The 19 upland NERICAs were derived from interspecific crosses involving a common O. glaberrima CG14 and three O. sativa (WAB56-50, WAB56-104, and WAB181) parents (Supplementary Table S1). Highly similar NERICA8 vs. NERICA9 and NERICA16 vs. NERICA18 are pairs of sister lines extracted from WAB56-104/CG14//2^∗WAB56-104 and CG14/WAB181-18//2^∗WAB181-18, respectively. Using 102 microsatellite markers on 18 upland NERICAs, we previously reported wider genetic variation among most pairs of upland NERICAs except NERICA8 and NERICA9 that were found to be nearly identical at the molecular level (Semagn et al., 2006). The 66 lowland NERICAs were developed from interspecific crosses involving three O. glaberrima (TOG 5681, TOG 5674, and TOG 5675) and five O. sativa (IR64, IR28, IR31785-58, IR1529-680, and IR31851-96) parents. TOG 5681 and IR64 were used as parents in 63 and 60 of the lowland NERICAs, respectively (Supplementary Table S1), which resulted in multiple sister lowland NERICAs with a common pedigree, the same level of inbreeding, and the same adaptation. Twelve of the 66 lowland NERICAs (NERICA-S44, NERICA-L17, NERICA-L14, NERICA-L15, NERICA-L36, NERICA-L03, NERICA-L33, NERICA-L16, NERICA-L59, NERICA-S-19, NERICA-L08, and NERICA-L49), differ from at least one other NERICA by <2% of scored alleles, and have been considered redundant based on the maximum length subtree method (Supplementary Table S1). Despite the common parentages, these lines exhibited large phenotypic variability, which formed the basis for their releases as different varieties. Thus, these lines are evidently different for some key genes, and may even be used as nearly isogenic pairs to determine genomic regions causing this phenotypic diversity, once the differences are confirmed in further phenotypic studies.

Several studies reported clear population structure in Asian rice based on subspecies/ecotypes (Glaszmann, 1987; Zhang et al., 1992; Garris et al., 2005; Zhao et al., 2011). Our results from cluster analysis (Figure 2), the model-based population structure (Figure 4), and PCA (Figure 5) revealed clear population stratification consistent with ecological adaptation and ecotypes. Cluster analysis clearly separated the upland genotypes from the lowland genotypes. All japonica genotypes in our study are upland types, and most (75%) clustered together with all 19 upland NERICAs derived from three japonica parents. The 66 lowland NERICAs derived from five indica parents clustered together with the intraspecific lowland ARICA and lowland indica genotypes. The model-based structure at K = 3 and PCA further divided the lowland genotypes into two subgroups corresponding indica (group 2) and most interspecific lowland NERICA (group 3) genotypes. This is likely because while the NERICA are predominantly O. sativa (indica) background following backcrossing, they contain a low proportion of O. glaberrima alleles from the initial cross, which was tracked in a previous study using 60 microsatellite markers (Ndjiondjop et al., 2008). Depending on the number of backcrosses used, the average remaining O. glaberrima genome contribution varied from 7.2 to 8.5%. The lowland NERICAs with the smallest introgression from the O. glaberrima parent were the most similar to the lowland indica compared to their interspecific sister lines.

The three main rice ecologies across west and central Africa are upland, the rainfed lowlands, and the irrigated lowlands (Maclean et al., 2002) which, respectively, account for ∼44, 31, and 12% of the total rice production area in the region. These ecologies differ significantly in terms of the severity of drought, soil fertility, soil acidity, iron toxicity, diseases and pests, and desirable agronomic traits (i.e., maturity, plant height, and yield potential). Upland rice is grown in well-drained soils with variable topographies ranging from sloping lands with high runoff to low-lying valleys and flat lands. These areas are characterized by poor soil physical and chemical properties, including low soil fertility and high soil acidity, and erratic precipitation. Most upland rice varieties are thus characterized by early maturity, deep roots system, and a higher tolerance to drought and acidic soils; however, they have low yield potential (on average about 1 t ha^-1) and tend to lodge under high-levels of fertilizer and supplemental irrigation.

Rainfed lowland systems are more robust than upland systems, with good potential for intensification. Rice yields in rainfed lowlands are substantially higher than those in the rainfed uplands, but nevertheless still on average about 2 t ha^-1; however, there is a possibility of higher yield potential for these genotypes with increased external inputs, including supplemental irrigation. Most cultivars grown in the irrigated lowland ecology have a short stature to avoid lodging under high fertilizer input and can produce over 5 t ha^-1. Although rice genotypes have clear differences in morphological, phenological and other traits for adaptation to upland, irrigated lowland and rainfed lowland ecologies, we only observed moderate genetic differentiation (F_ST = 0.015) between the rainfed lowland and irrigated lowland ecologies as compared with the very great genetic differentiation between irrigated and upland (0.680) and rainfed and upland (0.70) ecologies (Supplementary Table S6).

Indica genotypes are predominantly adapted to the lowland tropical and subtropical regions, while japonica rice is more adapted to the upland and highland temperate and tropical regions. In the present study, we observed similar levels of polymorphism and diversity within indica and japonica (Table 2), which is in some disagreement with other studies (Zhang et al., 1992; Garris et al., 2005) who reported higher genetic variation and haplotypes within the indica than japonica groups. The present study found 710 SNPs with F_ST values ranging from 0.904 and 0.941 and were sufficient to clearly differentiate the indica and japonica genotypes (except nine genotypes). Previous studies were also able to assign most rice lines into their respective subspecies groups using molecular markers (Morishima and Oka, 1981; Glaszmann, 1987; Zhang et al., 1992). The high genetic differentiation between the indica and japonica, and also between upland and lowland ecologies may also be partly due to the autogamous breeding system, which plays a significant role in structuring the genetic variation within and among hierarchical groups or populations. Generally, outcrossing promotes gene flow, which results to higher polymorphism and greater genetic diversity, while selfing restricts gene flow and leads to more genetic differentiation among populations (Hamrick and Godt, 1997). The large amount of genetic differentiation attributable to differences among groups (ranging from 41.1 to 76.0%) observed in the present study (Table 3) agrees with previous studies.