Selection of SNP panel for multiplex amplicon sequencing

To identify the SNPs that could be used for genotyping of M. oryzae from Bangladesh by multiplex amplicon sequencing (Floodlight Genomics, https://floodlightgenomics.com), we filtered 15,871 SNPs identified in Islam and colleagues (2016) [10] using the following criteria: (i) they had to be polymorphic among wheat blast strains alone to make the markers useful for diagnostics; (ii) the minor allele was >30%; (iii) SNPs were located on long exonic sequences (>1,500 bp without interrupting intron); and (iv) long exons to contain only 2 to 4 SNPs. The last two criteria were to make sure that the assay will focus on SNPs surrounded by well-conserved stretches among wheat blast strains with an aim to reduce amplification failures due to polymorphism in the primer binding sites. The above criteria reduced the available genomic regions to 102 loci. We designed 102 PCR primer pairs to amplify approximately 200 bp amplicon for each gene containing 100 bp flanking regions on each side of the SNP locus for multiplex amplicon sequencing. Trial sequencing runs with these primer pairs resulted in 84 primer pairs that consistently produced amplicons that can be sequenced to identify the correct nucleotides in those SNP loci in Bangladeshi wheat blast isolates [32]. We used these as a panel of 84 SNPs to discriminate between the wheat blast clonal lineage of M. oryzae in Bangladesh from other genotypes. Initial analyses and the use of the 84 SNP panel was reported in [32], and the details of 84 SNPs including gene names and primer sequences along with the dataset for benchmarking are available in Tembo and colleagues (2021) [33].

Phylogenetic placement of Magnaporthe oryzae wheat-infecting isolates from the Zambian wheat blast outbreak (2018 to 2020) using a set 84 SNPs

To establish the genetic makeup and the phylogenetic placement of the wheat-infecting lineage that caused a wheat blast outbreak in Zambia (2018 to 2020), we analyzed a set of 84 SNPs, which were designed to distinguish between the pandemic clonal lineage of the wheat blast fungus (Magnaporthe oryzae) that reached Southeast Asia in 2016 from other M. oryzae genotypes [10]. The dataset included 237 M. oryzae isolates from 13 different grasses that were genotyped using multiplex amplicon sequencing (“MonsterPlex”) [33,34]. MonsterPlex is a highly accurate and cost-effective genotyping method to rapidly genotype field isolates [32,33]. The Monsterplex genotypes were extracted from the genotyping matrices [33,34] as previously described [35]. To complement the MonsterPlex dataset and increase the geographic breadth of M. oryzae isolates, we extracted the 84 SNPs from 351 publicly available M. oryzae genomes. The joint dataset consisted of 537 worldwide distributed M. oryzae isolates (Figs 1 and S1 and S1 Table).

After removing adapters with AdapterRemoval v2 [36], we mapped the Illumina-derived short reads to the M. oryzae B71 reference genome [37] using BWA mem [38]. To identify the genomic location of the 84 diagnostic SNPs in the M. oryzae B71 reference genome, we aligned each diagnostic SNP surrounded by 100 bp surrounding regions to the B71 reference genome using blastn [39]. Afterwards, for each of the per isolate mapped BAM files, we retrieved the genotypes at each of the 84 genomic locations using samtools mpileup [40]. We concatenated all SNPs in a multi-fasta-like file that was used as input for phylogenetic analyses.

We built a Neighbor-Joining tree that includes a total of 537 M. oryzae isolates using the R package phangorn [41]. We displayed a tree topology that corresponds to the optimal tree drawn from 1,000 bootstrap replicates (Figs 1 and S1).

We further estimated the accuracy of the genotyping method by comparing SNP data acquired from 51 isolates using MonsterPlex to the SNPs extracted from matching genome sequences. We found that SNPs from both acquisition methods were identical whenever a SNP call was made (100%, 4,251/4,251 sites). In total, only 33 sites had gaps with missing data from MonsterPlex (0.77%, 33/4,284 sites) (S1 Table).

To show that the set of 84 SNPs are informative, we compared the genetic (Hamming) distances between each pair of blast isolates using the set of 84 SNPs and the genome-wide SNPs. Our analysis revealed a correlation coefficient of 0.82 between the two sets of pairwise distances (S2C Fig). To show that this correlation is robust and much higher than expected by chance, we repeated the calculation of pairwise distances for both datasets (84 SNPs and genome-wide SNPs) randomly subsampling a subset of the isolates (10%) with and without isolate names permutation with 100 repetitions. This analysis revealed a median correlation of pairwise distances of 0.82 and 0.001 for the resampling with permuted distances and sample pairs (S2C Fig). We repeated the analysis using only pairwise distances among wheat-infecting isolates and obtained a correlation coefficient of 0.9 (S2D Fig), which shows that the set of 84 SNPs accurately reflect the genetic diversity of the wheat blast fungus.

Processing of short reads and variant calling

Our phylogenetic analyses based on 84 SNPs (Figs 1 and S1) confirmed our previous analyses, which showed that the emergence of wheat blast in Zambia and Bangladesh was caused by the same pandemic lineage of M. oryzae that traces its origin in South America [32,35,42]. Consequently, from here on, we analyzed a set of 71 wheat-infecting M. oryzae genomes that belong to the wheat-infecting genetic lineage. This set included isolates from South America (N = 37), the outbreak in Bangladesh (2016 to 2018) (N = 21), and isolates from the outbreak in Zambia (2018 to 2020) (N = 13) that were exclusively sequenced for this study [43,44] (S2 Table).

We removed adapters from the short reads from the set of 71 M. oryzae genomes using AdapterRemoval v2 [36]. To avoid reference bias, a common problem in population genetic analysis when comparing genetically similar populations, we mapped the short reads to the rice-infecting 70–15 reference genome [45] using BWA mem2 [46,47] (S2 Table). We used samtools [40] to discard non-mapped reads and sambamba [48] to sort and mark PCR optical duplicates.

As a first step to call variants, we used HaplotypeCaller from GATK [49,50] to generate genomic haplotype calls per individual using the duplicate-marked BAM files as input. Subsequently, we used CombineGVCFs, GenotypeGVCFs, and SelectVariants from GATK [49] to combine the individual genomic VCFs, call genotypes, and filter SNPs, respectively. In order to select high-quality SNPs to be included in our phylogenetic and population genetic analyses, we filtered SNPs using Quality-by-Depth (QD), which is one of the per-SNP summary statistics generated by GATK. Using M. oryzae genomes sequenced by two different sequencing technologies, we have previously shown that QD is the summary statistic that generates the best true positive and true negative rates [19]. Following this approach, using GATK VariantFiltration [49], we kept SNPs with QD values no further away than one standard deviation unit away from the median of the distribution of QD values [19]. Finally, we created a new VCF file keeping only non-missing positions using bcftools [40].

Population structure analyses

To assess the population structure of the 71 M. oryzae isolates, we used genetic distances coupled with dimensionality reduction methods. First, we calculated pairwise Hamming distances using Plink V.1.9 [51]. These distances were used as input for PCA using the function prcomp from the R package stats [52] (Fig 3A). Additionally, we used f3-outgroup statistics [53] to establish the pairwise relatedness between M. oryzae samples (X and Y) after divergence from an outgroup: f3(X, Y; outgroup). We used the rice-infecting M. oryzae 70–15 as an outgroup [45]. We calculated z-scores for every possible pairwise sample comparison included in the f3-statistics test (N = 2,485). Subsequently, we carried out hierarchical clustering using the function hclust from the R package stats [52]. As input, we used a distance matrix generated from the f3-statistics-derived f3 values (S3 Fig).

Distinguishing clonality from outcrossing

To distinguish clonality from outcrossing in the B71 pandemic lineage and other genetic groups identified in our population structure analyses, we used patterns of LD decay. While sexual reproduction (outcrossing) will generate patterns of LD decay that are driven by meiotic recombination, LD is not expected to decay in asexual non-recombining populations, i.e., the whole genome will be in LD. We analyzed LD decay patterns in the B71 lineage, the PY0925 lineage and treated the rest of Brazilians M. oryzae isolates as one group. To compute standard measures of LD, we used VCFtools [54]. Since VCFtools is designed to handle diploid organisms, we transformed the haploid VCF files into “phased double haploid” VCF files. Afterwards, we computed pairwise LD as r² and Lewontin’s D and D’. To evaluate LD decay, for each of the genetic groups, we calculated the average of each LD measure (r², Lewontin’s D and D’) in bins of physical genomic distance (Figs 2B and S4). To quantify the significance of LD decay, we fitted an exponential decay model using nonlinear least squares. In order to compare the patterns of LD decay between the clonal lineages and the Brazilian group, we downsample the number of SNPs in the Brazilian group to the number of SNPs segregating in the B71 clonal lineage.

To perform a qualitative comparison between the observed patterns of LD decay with the expectations of LD decay in idealized populations, we performed forward simulations using FFPopSim [55]. We simulated genomes that consisted of 300 equidistant SNPs. We first sought to ascertain the effect of the probability of sexual reproduction per generation on the patterns of linkage disequilibrium decay. Each simulation was carried out for 100 generations keeping the population size, the crossover probability, and the mutation rate constant, but changing the probability of sexual reproduction per generation (0, 10⁻³, 10⁻², 10⁻¹, and 1) (S5 Fig). Additionally, we investigated the effect of the population size on the patterns of LD decay. To this purpose, we simulated genomes that consisted of 200 equidistant SNPs. Each simulation was carried out for 100 generations keeping the crossover probability, the mutation rate, and the probability of sexual reproduction per generation constant, but changing the population size parameter (10², 10³, 10⁴, 10⁵) (S6 Fig).

Finally, to detect recombination events in each of the genetic groups, we performed the four-gamete test [56] using RminCutter [57]. To be able to compare the number of violations of the four-gamete test among genetic groups, we normalized the number of violations of the four-gamete test by the number of segregating SNPs per genetic group (S7 Fig).

Phylogenetic analyses, estimation of evolutionary rates, and divergence times

To carry out phylogenetic analyses, we used only the non-recombining genetic groups (clonal lineages) B71 and PY0925 (the latter was used as an outgroup) and included exclusively positions with no-missing data (full information). First, we generated a maximum likelihood (ML) phylogeny using RAxML-NG with a GTR+G substitution model and 1,000 bootstrap replicates [58]. Then, we used the best ML tree as input for ClonalFrameML [16], a software that detects putative recombination events and takes those into account in subsequent phylogenetic-based inferences. For every isolate, we calculate the percentage of total SNPs masked by ClonalFrameML (S8 Fig). Additionally, we used pairwise Hamming distances to evaluate the levels of intra- and inter-outbreak genetic variation before and after ClonalFrameML filtering (S9 Fig).

Since the LD decay analyses revealed that the B71 pandemic lineage is a non-recombining clonal lineage, we hypothesized that the SNPs marked as putatively affected by recombination are preferentially located in genomic regions affected by structural variants, e.g., presence/absence variants. Such variants will generate phylogenetic discordances due to differential reference bias among the B71 isolates. To test this hypothesis, we created full-genome alignments of the B71 and the 70–15 reference genomes using Minimap2 [59] and visualized the output with AliTV [60]. Then, we overlapped the visual output with the SNPs putatively affected by recombination that were previously identified by ClonalFrameML (S10 Fig).

To test for the existence of a phylogenetic temporal signal (i.e., a positive correlation between sampling dates and genetic divergence), we used the recombination-corrected tree generated by ClonalFrameML (Fig 2C). Using this tree, and without constraining the terminal branch lengths to their sampling times, we measured tip-to-root patristic distances between all isolates of the B71 clonal lineage and the outgroup clonal lineage PY0925 with the R package ape [61]. We calculated the Pearson’s correlation coefficient between root-to-tip distances and sampling dates, and fitted a linear model using them as response and linear predictor, respectively (distances ~ sampling dates) (Fig 2C). We calculated 95% confidence intervals of the correlation signal between root-to-tip distance and collection dates by sampling with replacement and recalculating the correlation coefficient 1,000 times (Figs 2C and S11). Additionally, to demonstrate that the obtained correlation coefficient was higher than expected by chance, we performed 1,000 permutation tests, where the collection dates were randomly assigned to the wheat blast isolates (S11 Fig).

To estimate the evolutionary rate and generate a dated phylogeny, where the divergence dates of all common ancestors are estimated, we used two different approaches. First, we used BactDating, a Bayesian framework that estimates evolutionary rates and divergence times using as an input a precomputed phylogenetic reconstruction [62]. As input for BactDating, we used the recombination-corrected tree generated by ClonalFrameML. The tree was loaded into BactDating using the function loadCFML, which permits the direct use of the output of ClonalFrameML as input for BactDating without the need of correcting for invariant sites (Fig 2C). In our second approach, we used BEAST2, a Bayesian approach, which in contrast to BactDating, co-estimates the evolutionary rate and the phylogenetic reconstruction [63]. From the alignment of the concatenated SNPs, we masked those that ClonalFramML marked as putatively recombining and used the masked alignment as input for the BEAST2 analyses. We carried out a Bayesian tip-dated phylogenetic analyses using the isolates’ collection dates as calibration points. Since, for low sequence divergence (<10%), different substitution models tend to generate very similar sequence distance estimates [64], to simplify the calculation of parameters, we used the HKY substitution model instead of more complex models such as GTR. To reduce the effect of demographic history assumptions, and to calculate the dynamics of the population size through time, we selected a Coalescent Extended Bayesian Skyline approach [65]. The evolutionary rate was co-estimated using a strict clock model and a wide uniform distribution as prior (1E-10 - 1E-3 substitutions/site/year), which permits a broad exploration of the MCMC chains. To test the robustness of our evolutionary rate estimation to changes in substitution and clock models, we repeated the analysis using GTR in combination with a strict clock model and HYK in combination with a random local clock model [66]. To have genome-scaled results, the invariant sites were explicitly consider in the model by adding a “constantSiteWeigths” tag in the XML configuration file. Using the CIPRES Science Gateway [67], we ran four independent MCMC chains, each of which had a length of 20,000,000 with logs every 1,000 iterations. We combined the outputs with the LogCombiner and used TreeAnotator [63,68] to calculate the Maximum Credibility Tree as well as dating and support values for each node (Figs 2C and S12 and S3 and S4 Tables).

Determination of mating types

To assign the mating type for each isolate, we used two approaches. First, we created a fasta file containing the nucleotides codifying for the two mating type loci: MAT1-1-1 (GenBank: BAC65091.1) and MAT1-2-1 (GenBank: BAC65094.1) and used it as a reference genome. Using bwa-mem2 [46,47], we mapped each isolate and used samtools depth [40] to calculate the breadth of coverage for each locus as a proxy for the mating type assignment.

Identification of AVR-Rmg8 effector variants and generation of the Avr-Rmg8 family tree

We used a mapping approach to identify Avr-Rmg8 family members in all 71 wheat blast lineage genomes. We mapped short reads sequencing data of each isolate to the B71 reference genome assembly [37] using BWA mem2 [38] and extracted the consensus sequence of the AVR-Rmg8 locus from the output alignment files using SAMtools v.1.9 and BCFtools v.1.9 [40] for each isolate. This led to the identification of five AVR-Rmg8 variants in 71 sequences.

Generation of the Avr-Rmg8 family tree

To infer the AVR-Rmg8 family tree, we generated a multiple sequence alignment of 108 amino acid sequences of all identified variants using MUSCLE [69]. We then removed pseudogenes and duplicated sequences to generate a nonredundant dataset and generated a maximum-likelihood tree with 1,000 bootstrap replications in MEGA7 [70].

Leaf-drop and spray infection assay

To evaluate the response of Rmg8 against wheat blast isolates from Zambia, we carried out leaf drop and spray inoculations. An Oryza-infecting isolate from French Guyana, Guy11 was used as a negative control as this isolate does not infect wheat, while a Triticum-infecting isolate PY6047 from Brazil was used as a positive control it lacks the virulent allele of AVR-Rmg8 [71]. Another Triticum-infecting Bangladesh isolate BTJP4-16 that carries an avirulent allele of AVR-Rmg8 was also included. Wheat cultivars, S-615 (+Rmg8) and fielder (-Rmg8) were grown for 14 days in 9-cm diameter plastic plant pots or seed trays. Detached leaves from two weeks old seedlings were inoculated with fungal conidial suspension diluted to a final concentration of 1 × 10⁵ conidia mL-1 using drop inoculation method. The inoculated detached leaves were incubated in a growth room at 24°C with a 12 h light period. Alternatively, plants were inoculated by spray infection using an artist’s airbrush (Badger, United States of America) as previously described in [72]. After spray inoculation, the plants were covered in polythene bags and incubated in a negative pressure glasshouse with a 12 h light and dark cycle. Disease severity was scored after 5 to 6 days by evaluating lesion color and count or color and size for spray infection or drop inoculation, respectively. Lesions were classified as small black/brown non-spreading spots for resistant and large gray-spreading for susceptible. Each infection experiment was carried out three times.

Generation of crosses and fertility status analysis

Generation of genetic crosses was carried as previously described in [73]. A subset of Zambian isolates ZMW20_04, ZMW18_06, ZMW20_07, ZMW18_10, ZMW20_15, and ZMW20_16 (MAT-1-2) were tested against two finger millet tester isolates from Tanzania, T15 (MAT-1-1) or T26 (MAT-1-2), one from Kenya K1(MAT-1-1), and one from Ethiopia E12 (MAT-1-1). Two isolates were cultured opposite each other on oatmeal agar plate and incubated at 24°C for 7 to 10 days. The cultures were then transferred to a 20°C incubator until flask-shaped perithecia appeared at the crossing point. A Leica DFC360 FX microscope (Leica, Wetzlar, Germany) was used to visualize and image the formation of perithecia.

Isolation of azoxystrobin-resistant Magnaporthe oryzae strains

Isolation of azoxystrobin-resistant Magnaporthe strains was carried out by exposure of spores of the fungus to azoxystrobin at 100 g ml^-1. Four million conidia of the Zambian wheat-infecting isolate ZMW20-14 were recovered from CM grown agar cultures and plated on the surface of six 12 cm² square Minimal medium agar plates (MM; [74]) with azoxystrobin at 100 g ml^-1. Plates were incubated for 2 weeks and then overlaid with MM plus 100 μg ml^-1 azoxystrobin and grown for a further week. Five colonies were isolated from these plates and colonies were purified by subculture to MM with 100 μg ml^-1 azoxystrobin for a further week before genotyping. PCR competent genomic DNA was isolated from a 4 mm² plug of mycelium from the purified azoxystrobin-resistant colonies (named AZ1-AZ5) with disruption using an automated tissue homogenizer and cell lyser. To extract the mycelial plug, it was placed in 400 l of extraction buffer (1 M KCl, 10 mM Tris-HCl (pH 8), 5 mM EDTA (pH 8) together with a 5 mm stainless steel bead (Qiagen)) and macerated at 1,200 rpm for 2 min in the Geno/grinder followed by a 10-min centrifugation (17,000 x g) in a bench top centrifuge at room temperature, and 300 l of the supernatant was transferred to a new tube and nucleic acids were precipitated with 300 l of isopropanol followed by a 10-min centrifugation at room temperature at 17,000 x g. Nucleic acid pellets were air dried for 10 min and then resuspended in 50 μl of TE with 50 μg ml^-1 RNAse by vortexing and then incubated for 10 min at 37°C, and 1 l of the DNA was used in a 50 l PCR reaction with the enzyme Q5 polymerase (New England Biolabs) and the primers Cytb-f AGTCCTAGTGTAATGGAAGC and Cytb-r ATCTTCAACGTGTTTAGCACC (annealing temperature 61.5°C). Amplicons were sequenced following gel purification in both directions using the same primers used to generate them to score for the presence or absence of the strobilurin resistance conferring mutation [26].