The study panel comprised of 131 cassava varieties collected from different geographical areas (three agro-ecological zones) of Sierra Leone (Figure 1; Supplementary Table S1). The varieties were collected using the multistage sampling approach. Districts considered as the cassava belt areas within the country were selected and three communities per district were also selected at random based on the list of communities used by Statistics Sierra Leone during 2018. The varieties were collected from farmers’ fields with the aid of International Institute of Tropical Agriculture (IITA) cassava descriptors, farmers’ knowledge, preferences and special attributes of the varieties. The different collection sites were located using global positioning system (GPS). A minimum of 10 km separated the two nearest sites of collection. The planting materials collected were stem cuttings.

The experiment was conducted at the Biological Sciences Experimental Farm, Njala University (08˚14'S, 12˚1'W), southern Sierra Leone. The experiment was laid out in an augmented randomized complete block design with three replicates. The stem cuttings of each variety were cut into 20 cm long each and planted in holes made on the crest of single ridge plots measuring 10 m². The cuttings were planted at a planting distance of 1 m × 1 m between and within rows giving a plant population of 10,000 plants ha⁻¹. The field trial management was done whenever necessary, following the technical recommendations and standard agricultural practices for cassava [26] . Phenotypic data on fresh storage root yield and African cassava mosaic disease (ACMD) severity were recorded. Detailed trait description and ontologies are available

(https://cassavabase.org/search/traits). The fresh storage root yield was recorded

in t·ha⁻¹ at·harvest (eight months after planting). The ACMD severity score was done based on a visual assessment of the relative area of plant surface affected by the mosaic virus disease using a five-ordinal scale of 1 - 5 at 6 months after planting. The 1 - 5 disease rating scales used to record the proportion of a plant surface in a plot affected by ACMD represented the following: 1 for no visible virus symptom, 2 for mild symptoms on few leaves but no leaf distortion, 3 for low symptoms of the mosaic virus on leaves, 4 for the severe mosaic on most leaves and leaf distortion, and 5 for severe mosaic and bleaching with severe leaf distortion and stunting.

The phenotypic data were analyzed using the one-step linear mixed model that utilizes G-matrix to compute the best linear unbiased predictor (BLUP) values of each genotype for a trait from the best fit model; and the average information criterion (AIC) in restricted maximum likelihood (REML) algorithm [27] was done in the ASReml-R version 4 package [28] . Accordingly, the genetic variance was partitioned into the additive genetic effect (i.e. the proportion that associated with a covariance structure proportional to genetic relationships derived from the molecular markers) and the non-additive genetic effect. The non-additive genetic variance is explained by individual identity rather than the genomic relationship matrix [29] [30] . Broad sense heritability (H²) estimates for fresh root yield and ACMD were estimated from phenotypic variance (σ²p) and the genotypic variance (σ²g). The BLUP values of the genotypes for the traits extracted from the best fit model were used as input for the GWAS model.

Total genomic deoxyribonucleic acid (DNA) was isolated from lyophilized young and fully expanded healthy leaves of each varieties studied. The DNA was extracted using the CTAB procedure with slight modification [31] . The quality and concentration of the DNA were assessed using agarose gel and nanodrop, respectively [32] . High-throughput genotyping was done using 96 plex DArTseq protocol, and SNPs were called using DArTSoft as described by Kilian et al. [33] . The raw HapMap file generated was first converted to a Variant call format (VCF) and filtered for missing value and polymorphic SNPs using quality control criteria of low sequence depth < 5; SNP markers with missing values > 20%; genotype quality < 20; minor allele frequency (MAF) < 0.05 and heterozygosity > 50 [15] [34] . Of the 61,214 SNPs subjected to the filtering quality criteria, 12,500 good-quality SNPs were retained for further analyses.

Different population genetic analysis techniques were used to explore the structure and level of genetic diversity in the germplasm. The SNP distribution and density were determined using the “Cmplot” function implemented in the CMplot R package [35] . The SNPlay open website was used to estimate the rate of transition and transversion across the retained SNP. Summary statistics on the minor allele frequency (MAF), polymorphism information content (PIC), the observed and the expected heterozygosity were estimated based on the function “--freq” and “--hardy” using PLINK V1.90 [36] .

The genetic relationship among the cassava germplasm was explored using the principal component analysis (PCA) in FactorMiner R package V3.3 [37] . For the PCA, the origin of the varieties was used as factor.

The cluster samples were put into populations using structure software version 2.3.3 [38] [39] . The structure simulations were done using the admixture model with a burn in period of 20,000 iterations and a Markov chain Monte Carlo (MCMC) set at 20,000. The simulations were repeated 3 times for K-values of 1 to 10. The optimal subpopulation model was determined using the following procedures: 1) by applying the informal pointers (i.e. geographical origin) proposed by Pritchard et al. [38] and Falush et al. [39] ; 2) by considering ΔK, a second order rate change with respect to K, as defined in Evanno et al. [40] , and implemented in STRUCTURE HARVESTER V0.6.94 [41] . Structure population was then plotted using barplot function implemented in R. The phylogeny tree was constructed using ape version 5.0 implemented in R [42] . The exploratory Discriminant Analysis of Principal Components (DAPC) was applied using the adegenet package V2.1.8 [40] . The Admixture was performed through the Bayesian Information Criterion (BIC). A hierarchical cluster was constructed utilizing a kinship relation matrix implemented in GAPIT V3 [41] .

The GWAS analysis was done using the compressed mixed linear model (CMLM) implemented in the GAPIT R package V3 and the Manhattan and the QQ plots were visualized in CMplot [35] . For the GAPIT analysis, population structure (Q) and the relationships among individuals were accounted for through a principal component (PC) analysis and a kinship (K) matrix generated from marker data, respectively [36] . For each trait, the optimal number of PCs/covariates included in the GWAS models was determined through model selection using the Bayesian information criterion (BIC), with a maximum of four tested PCs. The GWAS study was done using the following formula: Y = X β + W α + Q v + Z u + ε ; where Y was treated as the observed vector of BLUP; β as the fixed effect vector (p ×  1) other than molecular markers effects and population structure; α as the fixed effect vector of the molecular markers; ν as the fixed effect vector from the population structure; u as the random effect vector from the polygenic background effect; X, W, and Z are the incidence matrixes from the associated β, α, ν, and u parameters; and ε as the residual effect vector. The significance threshold for the marker-trait associations (MTA) was set to p = 0.05 after assigning the associated probability (P) to each marker using the Bonferroni threshold correction (−log10(P)). The percentage of variation explained by the associated marker (R²) was determined using a step regression implemented in GAPIT.

Circular Manhattan and Quantile-Quantile (QQ) plots were generated by plotting the negative logarithms (−log10) of the P-values against their expected P-values to fit the appropriateness of the GWAS model with the null hypothesis of no association and to determine how well the models accounted for population structure using CMplot [35] . The Manhattan plot was created for visualization of GWAS on the entire genome, and zoom mapping was performed on a particular chromosome after identifying a significant SNP marker.

Broad-sense heritability estimates was high, 0.99 for African cassava mosaic disease and intermediate, 0.51 for fresh root yield. The phenotypic value for ACMD ranged from 1.0 to 5.0 with an average of 1.76. The phenotypic value for the fresh root yield ranged from 0.20 to 21.40 t·ha⁻¹ with an average of 7.08 t·ha⁻¹ ( Table 1). The best linear unbiased predictor (BLUP) values of African cassava mosaic disease (ACMD) and fresh storage root yield (FSRY) varied among 131 cassava varieties studied (Supplementary Table S2).

The DArT genotyping of 131 cassava varieties detected the highest number of SNPs (1434) mapped on chromosome 1 and the lowest of 457 on chromosome 13 (Supplementary Figure S1(A)). Transition SNPs (60.76%, 4,470 SNPs) were more frequent than transversions (39.24%, 2887 SNPs) (Supplementary Figure S1(B)). The observed heterozygosity value ranged from 0.0 to 0.802, with an average of 0.233 (Supplementary Figure S1(C)). The expected heterozygosity value ranged from 0.023 to 0.5, with an average of 0.263 (Supplementary Figure S1(D)). The minor allele frequency ranged from 0.012 to 0.500, with a mean of 0.184 (Supplementary Figure S1(E)). The polymorphic information content (PIC) ranged from 0.02 to 0.38, with an average value of 0.22 (Supplementary Figure S1(F)).

The population structure of a diverse panel of 131 cassava varieties was investigated on the basis of a 1K method of model-based Bayesian clustering using 12,500 SNP markers. The population structure analysis revealed the existence of five distinct subpopulations in the cassava panel, which was found consistent with the results of the phylogeny tree analysis through the kinship matrix, and

ACMD = African cassava mosaic virus, FSRY = fresh root yield (FRY) of cassava.

PCA (Figure 1). The highest membership was recorded in group 1 (36 varieties) and the lowest reported in group 5 (12 varieties).

Group 1 had 36 varieties comprising 34 landraces (collected from farmers and maintained at the Njala Agricultural Research Center, Njala, Sierra Leone) and two improved varieties. One of the two improved varieties, SLICASS6 (SC6_6) is a released variety. Group 2 consisted of 30 varieties, mainly comprising 29 local landraces and one improved variety (IP6). Group 3 had 28 varieties, comprising of 23 local and five improved varieties. Group 4 had 25 varieties, comprising 21 local and four improved varieties. SLICASS6 (SC6_1) and SLICASS4 (SC4_2) are released varieties. Groups 5 had 12 members, comprising 10 local and two improved varieties (Supplementary Table S3). The Cluster membership displayed through the phylogeny tree was in perfect alignment with the Discriminant analysis of principal components (DAPC) cluster membership (Supplementary Table S3). The genetic relationship based on the principal component analysis revealed that the first two PCs account for 85.8% of the total variation (Figure 2). Both the local and improved varieties were distributed along PC1 and the PC2 (Figure 2).

Two SNP loci exhibited significant association with the reaction to African cassava mosaic disease infection (Table 2, Figure 3). The two SNP loci (chr2_10463640 and chr2_11944909) that associated with ACMD had marker effects of 2.47 and 2.93, respectively, and explained 22.6% of the total phenotypic variance (Table 2). These SNPs were mapped on chromosome 2 at 10463640 and 11944909 bp physical positions, respectively. The Quantile–Quantile (QQ) plot corroborated with reducing −log10 (p-value) towards the expected level for the ACMD (Figure 3). The LOD values for SNPs chr2_10463640 and chr2_11944909 were 4.83 and 4.06 with minor allele frequency (MAF) of 0.012 and 0.018, respectively.

Three SNPs significantly associated with fresh storage root yield (t·ha⁻¹), based

p represents the analysis of variance probability value associated with the variation across variants, SNP = single nucleotide polymorphism, MAF = minor allele frequency, LOD = logarithm of odds.

on p-values (p ≤ 0.005) across the whole cassava genome scan (Table 2, Figure 4). Of the three significant markers, one SNP locus (chr18_1103992) had marker effect of −2.33, and explained 10.1% of the total phenotypic variation. The other two SNPs chr1_31172650 and chr15_10392402 exhibited marker effects of 6.08 and 3.00, respectively, and explained 21.4% of the total phenotypic variation. The SNPs chr1_31172650, chr15_10392402 and chr18_1103992 were mapped on chromosomes 1, 15 and 18 at 31172650, 10392402 and 1103992 bp physical positions, respectively. Evidence of the SNP association was also found in the quantile-quantile (QQ) plot of the observed p-values of the association analysis for

the FSRY (Figure 4). The LOD values for SNPs chr1_31172650, chr15_10392402 and chr18_1103992 were 4.28, 4.11 and 4.33 with minor allele frequency (MAF) ranging from 0.012, 0.065 and 0.223, respectively.

The natural variation among the cassava accessions for traits was very informative. High broad-sense heritability of 0.99 for African cassava mosaic disease (ACMD) and intermediate broad-sense heritability of 0.51 for fresh root yield

demonstrated substantial genetic variation in the studied traits among the different varieties. These findings indicated that the studied traits are amenable to genetic improvement through selection [43] . Furthermore, the observed natural genetic variation in the studied cassava germplasm signifies their relevance for genetic studies.

A good understanding of the existing population structure within cassava breeding population is necessary for determination of its effects on the ability of GWAS to infer marker-trait association. In this study, all the three clustering methods utilized, DAPC, Kinship relationship matrix and structure, revealed five sub-populations that are imperative for preventing sham associations in GWAS [44] . Thus, the sample size, marker density and diversity demonstrated that the cassava breeding panel used for this study is sufficiently powered to capture allelic variations for ACMD and FSRY studied.

The average minor allele frequency of the risk allele tested for varieties was greater than 10% indicating its significance in detecting genetic effects in the studied populations. The findings support the view that loci with high minor allele frequency have a higher power to detect weak genetic effects compared with those with lower minor allele frequency values [45] [46] . The numbers and patterns of SNP mutations in this study indicate a bias in chloroplast genome evolution in varieties of cassava. The diversity and patterns of SNP mutations in the cassava varieties were possibly due to the function of genes as previously suggested by Cao et al. [47] .

The whole-genome scan for phenotypic and allelic variation in African cassava mosaic disease resistance and fresh root yield identified genome regions on four chromosomes (chromosomes 1, 2, 15, and 18) with significant −log10 values. Both Q matrix (population structure) and K matrix (Admixture) were considered covariates in a mixed linear model for the association analysis to reduce false-positive associations. The Q–Q plots for tolerance to African cassava mosaic disease and fresh root yield showed no inflation of p-values indicating that the structure of relationships was well accounted for in the GWAS analysis. These findings are consistent with the view that traits with no inflation of p-values show that the structural relationship is adequate for GWAS analysis [15] [48] . Genome-wide association mapping has been used to explore the elite alleles of many agronomic traits such as green mite infestation [49] , cassava brown streak disease [50] , cassava mosaic disease [51] and provitamin A and dry matter content [52] in cassava (Manihot esculenta). The genetic basis of ACMD resistance has been reported to be caused by a single major gene, CMD2 locus, detected on chromosome 12 [53] and chromosome 14 [54] . This gene was discovered using bi-parental linkage mapping [55] and GWAS [51] . In the present study, we detected two additional loci on chromosome 2 that indicate the existence of the ACMD resistance gene.

The phenotypic effect values of the favorable alleles of the studied traits were assessed and inferred to positively affect ACMD; and positively and negatively affect FSRY. Based on the stringent criterion of −log10, two significant markers trait associations exhibiting 1.46 e^−0.5 and 8.68 e^−0.5 were identified for ACMD and three ranging between 4.71 e^−0.5 and 7.71 e^−0.5 were identified for FSRY. The information on SNP variants from this study would fast-track the application of genomics-informed selection decisions in breeding cassava for resistance to African cassava mosaic disease and higher root yield. Similar studies revealing the great potential of GWAS in contributing to genomics-informed selection decisions have been reported for some root and tuber crops such as cassava [56] , water yam [44] , white yam [15] and potatoes [57] .