Genomewide mexicana ancestry is structured by elevation

We sampled paired sympatric populations from 14 geographically dispersed locations to assess the extent of gene flow between maize and mexicana in Mexico. Maize today is grown across the entire elevational range of its wild teosinte relatives, from sea-level up to 4000 meters [52]. Our sampled sites range from 1547–2600 meters in elevation, which spans a large portion of mexicana’s range and exceeds the upper elevational range for maize’s wild ancestor, parviglumis (Fig 1). For each of 14 maize/mexicana sympatric sample locations, we resequenced 7–15 individuals per subspecies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Sampled sympatric maize/mexicana populations compared to the distribution of teosintes.

(A) Elevational range of teosintes based on historical occurrence data (1842–2016) from [29]. Parviglumis and mexicana overlap at middle elevations (dark green) and maize today is grown across this entire elevational range. (B) Geographic location and elevation of contemporary sympatric maize and mexicana population pairs sampled across 14 sites. Map of Mexico created with Natural Earth data (https://www.naturalearthdata.com).

https://doi.org/10.1371/journal.pgen.1009810.g001

We additionally sequenced 43 individuals from 3 mexicana reference populations, totalling 348 low-coverage genomes (mean ∼1x). Two of these mexicana reference populations are documented to have no adjacent maize agriculture within the past 50 years, while a third higher elevation population (Amecameca) was chosen because it grows above the elevational range of parviglumis, and thus outside of the historical range of maize. For a maize reference population, we added 55 previously published high-coverage genomes from a population grown near Palmar Chico at 983 m [53, 54], well below the elevational range of mexicana. Because parviglumis is known to admix historically with both of our focal subspecies in Mexico, we also included 50 previously published high-coverage parviglumis genomes from the ‘Mound’ population at 1,008 m, also near Palmar Chico [53–55]. Completely allopatric reference populations are not available because maize has been grown at high density throughout Mexico across the full elevational range of both teosintes. A priori, gene flow from maize into mexicana is possible at Amecameca, and historically between maize and teosinte at all locations. We therefore assess for possible gene flow into each reference population below.

Principal components analysis of genetic diversity clearly separates maize and mexicana, with putative admixed individuals from sympatric populations having intermediate values along PC1. Additionally, PC2 provides evidence of gene flow from parviglumis, particularly into lower elevation mexicana populations (S1 Fig), which motivated us to analyse a 3-way admixture scenario in all subsequent analyses.

To estimate genomewide ancestry proportions for each individual, we ran NGSAdmix [56] with K = 3 genetic clusters and genotype likelihoods for all mexicana, maize, and parviglumis individuals. The three genetic clusters clearly map onto mexicana, maize and parviglumis ancestry, with no significant admixture in the lowland maize or parviglumis reference populations. We find minority parviglumis ancestry in the two lower-elevation reference mexicana populations, but no evidence of introgression from parviglumis or maize into the highest elevation population at Amecameca (Fig 2A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of mexicana ancestry by elevation.

(A) Genomewide ancestry estimates (NGSAdmix) for reference maize, mexicana and parviglumis individuals, grouped by sampling location. (B) Genomewide mexicana ancestry estimates (NGSAdmix) for sympatric maize and mexicana individuals (n = 305) along an elevational gradient, colored by sampling location. Lines show best linear model fit for mexicana ancestry by elevation for each subspecies separately.

https://doi.org/10.1371/journal.pgen.1009810.g002

Looking across samples from the 14 sympatric sites, we find a positive association between ancestry proportion and elevation (km), with higher mexicana ancestry at higher elevations in both sympatric maize (β = 0.22, P = 1.01 × 10⁻³¹) and sympatric mexicana (β = 0.32, P = 2.02 × 10⁻³⁸) individuals (Fig 2B).

Increasing mexicana ancestry at higher elevations is consistent with selection favoring mexicana ancestry at higher elevations, but could also be due to purely demographic processes, e.g. a higher density of (wind-dispersed) mexicana pollen at higher elevations, or increased gene flow from non-admixed maize populations at lower elevations. While most populations have admixture proportions well-predicted by their elevation, outlier populations may be the result of recent colonization histories for some locations or adaptation to other environmental niches. Within teosintes, elevation is a major axis of niche separation between parviglumis (the ancestor of maize) and mexicana [50, 57], but genetic differentiation also correlates with soil nutrient content and at least four principal components constructed from climatic variables [33].

Origin and timing of introgression

If mexicana ancestry found in contemporary maize genomes facilitated maize’s colonization of the highlands approximately 6.2 thousands years ago [36], we would expect introgressed ancestry tracts to be short, due to many generations of recombination, and possibly to be derived from an ancient source population common to many present-day maize populations. To test these predictions, we estimated local ancestry across the genome for individuals from each sympatric maize and mexicana population using a hidden Markov model (HMM) based on read counts ([58]; see Materials and methods). For each admixed population, this HMM simultaneously estimates local ancestry and, by optimizing the transition rate between different (hidden) ancestry states, the generations since admixture. We assumed a 3-way admixture scenario in which a founding mexicana population receives a pulse of parviglumis ancestry, then a pulse of maize ancestry.

Admixture between maize and mexicana is generally old, with median estimates of 1014 generations for sympatric maize populations and 509 generations for sympatric mexicana populations (S3 Fig). Parviglumis admixture timing estimates vary substantially across populations (median: 173, range: 29–1006). Because the HMM fits a single-pulse per ancestry to what was almost certainly multiple admixture events over time, we caution against over-interpretation of exact dates. Multiple pulses or ongoing gene flow biases estimates towards the more recent pulse(s) [59, 60] and even old estimates do not exclude the possibility of limited more recent admixture. These single-pulse approximations do, however, provide evidence that a large proportion of the introgression, especially mexicana ancestry into maize, is found on short ancestry tracts and therefore relatively old.

To identify likely source population(s) for introgressed ancestry, we compared F_ST between all sympatric populations using only reads from high-confidence homozygous ancestry tracts (posterior > 0.8) for maize and mexicana ancestry separately. We find that most mexicana ancestry in maize resembles other mexicana ancestry introgressed into other maize populations, rather that mexicana ancestry from the local sympatric mexicana population (Fig 3). This finding is consistent with most introgressed ancestry being drawn from a communal source population, but none of the sympatric mexicana populations have low enough F_ST to tracts introgressed into maize to be a recent source. While we cannot rule out recent introgression from an unsampled source population, the timing of our admixture estimates is more consistent with divergence of mexicana ancestry, once introgressed into a maize background, from its original source population(s) (S3 Fig). Additionally, mexicana ancestry tracts in maize have only slightly reduced genetic diversity (π, S4 Fig), meaning many mexicana haplotypes have introgressed into maize at any given locus, with no evidence of a strong historical bottleneck.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. F_ST between ancestry tracts from different populations.

F_ST between each pair of populations for maize ancestry tracts are shown in the upper left triangle, while F_ST estimates for mexicana ancestry tracts are shown in the lower right triangle. Populations are sorted by subspecies, then elevation. Local sympatric maize-mexicana population pairs are highlighted with a white dot and do not show reduced F_ST relative to other (non-local) maize-mexicana comparisons. Additionally, introgressed mexicana ancestry shows low differentiation between maize populations (creating a light-colored maize block in the left corner of the lower right triangle) and no potential mexicana source populations show especially low F_ST with this block. Light coloring generally across the upper left triangle reflects the low differentiation within maize ancestry, providing little information to distinguish between potential maize ancestry sources.

https://doi.org/10.1371/journal.pgen.1009810.g003

Two lower elevation maize populations are an exception to this general pattern: Ixtlan and Penjamillo. These populations have higher F_ST between their introgressed ancestry tracts and other mexicana tracts in maize (Fig 3), more recent timing of admixture estimates (S3 Fig), and reduced genetic diversity (S4 Fig). These patterns could be caused by small population sizes and more recent independent admixture, although F_ST does not identify a likely mexicana source population. Consistent with this interpretation, we have evidence that local maize at Ixtlan is at least partially descended from recently introduced commercial seed (relayed by local farmers [43]).

The lack of a clear reduction in F_ST for mexicana ancestry tracts between sympatric population pairs, combined with older timing of admixture estimates, indicates that while contemporary hybridization may occur in the field between maize crops and adjacent mexicana populations, this is not the source for the bulk of the introgressed mexicana ancestry segregating in highland maize.

Instead, we propose that the majority of mexicana ancestry in maize derives from admixture over 1000 years ago, possibly from a diverse set of mexicana source populations over a large geographic and temporal span, and the resulting ancestry tracts are now distributed across different contemporary maize populations. These genomewide average F_ST results, however, do not exclude the possibility that adaptively introgressed haplotypes at a particular locus came from one or more distinct, possibly local, source populations. While we also analyzed F_ST within high-confidence maize ancestry tracts, we found that maize ancestry is too homogeneous to make inferences about potential admixture source populations of maize into mexicana (Fig 3 and S4 Fig).

Selection against introgression genomewide

When there is widespread selection against introgressing variants at many loci across the genome, selection will more efficiently remove linked ancestry in regions of the genome with lower recombination rates, which creates a positive relationship between local recombination rate and the proportion of introgressed ancestry [4–13, 61]. To test whether such negative selection is shaping patterns of introgression genomewide in sympatric maize and mexicana, we first divided the genome into quintiles based on the local recombination rates for 1 cM windows. We then ran NGSAdmix on the SNPs within each quintile separately, using K = 3 clusters, to estimate ancestry proportions for each quintile. We used a recombination map from maize [62], which is likely to be correlated with other Zea subspecies at least at the level of genomic quintiles. A limitation of this analysis, however, is that we do not have a recombination map for hybrid populations, which means that e.g. segregating structural inversions will not necessarily show low recombination rates.

Our results from sympatric maize are consistent with selection against mexicana introgression at many loci genomewide, resulting in lower introgressed ancestry in regions of the genome with lower recombination rates (Fig 4A). We find a positive Spearman’s rank correlation between recombination rate quintile and mean introgressed mexicana ancestry proportion (ρ = 1.00, CI₉₅[0.80, 1.00]), reflecting the fact that introgression increases monotonically across quintiles. A similar analysis using f₄ statistics replicates this result (see Materials and methods, S5 and S6 Figs). The higher elevation maize populations show this pattern most starkly; while all individuals have low mexicana ancestry for the lowest recombination rate quintile, some high elevation populations have individuals with over 40% introgressed ancestry for the highest recombination rate quintile (Fig 4B). Using a linear-model fit, we found a significant positive interaction between recombination rate quintile and the slope of ancestry across elevation in sympatric maize (S4 Table). This is again consistent with low-recombination rate regions having a stronger effect of linked selection reducing mexicana ancestry, with higher elevation maize populations either experiencing larger amounts of gene flow or retaining more ancestry due to adaptive processes in high recombination regions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. .

(A) Introgressed ancestry by recombination rate. Inferred average genomewide introgressed ancestry in sympatric maize and mexicana individuals (NGSAdmix K = 3), by recombination rate quintiles. Group mean and 95% confidence interval based on bootstrap percentiles (n = 100) are depicted in black. Introgressed ancestry estimates for each individual are shown as points and points are jittered for better visualization. (B) Slope of mexicana ancestry introgressed into maize populations across elevation for each recombination rate quintile, based on NGSAdmix estimates. Each point is a sympatric maize individual and lines show the best-fit linear model for ancestry by elevation (with shaded 95% confidence interval), estimated separately for each quintile.

https://doi.org/10.1371/journal.pgen.1009810.g004

Because recombination rate is positively correlated with gene density in Zea [63], we also tested the Spearman’s rank correlation between quintiles defined by coding base pairs per cM and their proportion introgressed mexicana ancestry. Again we found evidence supporting pervasive selection against introgression (S8 Fig, ρ = −1.00, CI₉₅[−1.00, −0.90]).

In contrast, sympatric mexicana shows an unexpected negative relationship between recombination rate and introgression, with reduced maize ancestry in the highest recombination rate regions of the genome (ρ = −1.00, CI₉₅[−1.00, −0.90]). Correlations with coding bp per cM and based on f₄ statistics corroborate this pattern (see S6 Fig). One explanation is that some portion of maize alleles are beneficial in a mexicana background. While maize ancestry in general is not predicted to provide adaptive benefits in teosinte, invasive mexicana in Europe shows selective sweeps for maize ancestry at multiple loci that have contributed to its establishment as a noxious weed [64] and we speculate that maize could be a source of alleles adapted to human-modified landscapes.

We repeated these analyses using local ancestry calls as our introgression estimates and found a non-significant Spearman’s rank correlation between mexicana introgression and recombination rates for 1 cM windows in sympatric maize (S9 Fig, ρ = 0.011, CI₉₅[−0.038, 0.061]) and a positive rank correlation between maize introgression and recombination rate in sympatric mexicana (ρ = 0.385, CI₉₅[0.341, 0.428]). Contrasting results between global and local ancestry methods could be a reflection of true evolutionary differences across different time periods; local ancestry methods capture patterns from more recent gene flow that comes in longer ancestry blocks while STRUCTURE-like algorithms (NGSAdmix) and f₄ statistics are based on allele frequencies that collapse information across ancestry blocks of any size, capturing a longer evolutionary time scale. This interpretation would suggest that mexicana has experienced stronger selection against more recent maize gene flow than historical gene flow. However, we caution that local ancestry methods may also have subtle biases in power that are sensitive to local recombination rates and make them less reliable for comparing ancestry patterns across recombination rate quintiles.

Overall, we find support for widespread selection against introgression into maize and mixed results from similar tests of this hypothesis in mexicana.

High introgression peaks shared across populations

To assess adaptive introgression in our sympatric populations, we identified introgression ‘peaks’ where minor ancestry exceeds the genomewide mean by more than 2 standard deviations. We find no strong reduction in average diversity (π) for mexicana ancestry at high introgression peaks (S4 Fig). This maintenance of diversity implies that selection at most peaks has favored multiple mexicana haplotypes, and hard sweeps for recent beneficial mutations on a specific haplotype are rare.

We observe that many high mexicana ancestry peaks are shared across subsets of our 14 maize populations (see e.g. chr4, Fig 5). While most outlier peaks are unique to a single population, many peaks are shared across 7 or more of the populations (S10A Fig). To a lesser extent, we also observe sharing of high-introgression peaks for maize ancestry in sympatric mexicana populations (S10B Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Introgression in maize populations across chromosome 4.

Local introgressed mexicana ancestry frequency for each maize population compared to their genomewide mean. Populations are ordered from high to low elevation (top to bottom). High introgression peaks with more than 2 standard deviations above the population mean introgressed mexicana ancestry are highlighted in blue. Vertical black lines show the previously identified endpoints for a large inversion (Inv4m; coordinates from Fig 3 of [50]). For local ancestry on other chromosomes and for maize introgression into sympatric mexicana, see S11–S29 Figs.

https://doi.org/10.1371/journal.pgen.1009810.g005

High introgression peaks in many independent populations would be very unexpected by chance. However, our sampled populations do not provide independent evidence for adaptive introgression, due to shared gene flow and drift post-admixture (e.g. long-distance human-assisted dispersal of maize seed). To estimate the rate of peak sharing we should expect from demographic processes alone, we simulated 100,000 unlinked loci under a multivariate normal distribution (MVN) parameterized with the empirical ancestry variance-covariance matrix K (see Materials and methods). These simulations preserve the ancestry variance across loci within populations and non-independence in ancestry between populations.

For both sympatric maize and mexicana, every population shares an excess of high introgression peaks with all other populations compared to expectations set by our MVN null model. However, peak sharing is most elevated among high elevation maize populations (with the exception of Cocotilan, see Fig 6). To investigate the origins of population-specific peaks of introgression, we calculated F_ST between homozygous mexicana ancestry in local maize and their paired local mexicana population for these genomic regions. Patterns of F_ST between local sympatric pairs at local introgression peaks differed little from background F_ST (S30 Fig), offering little support for the idea that population-specific peaks arose from recent, locally sourced, adaptive introgression. Instead, patterns in maize are consistent with introgressed mexicana ancestry tracts from old shared admixture being favored by natural selection, and thus rising to high frequency, in a subset of populations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Introgression peaks shared across populations.

Networks for sympatric maize (top) and mexicana (bottom), where each node is a sampled population labelled by location and ordered by elevation. Edges connecting a pair of populations represent the percent of SNPs within shared ancestry peaks (introgressed ancestry > 2 s.d. above each population’s mean ancestry). Sharing between all pairs of populations exceeds expectations based on multivariate-normal simulations that model genomewide covariance in ancestry. The relatively darker thicker lines connecting the high elevation maize populations (except for Cocotilan), indicate that these populations share high introgression peaks at especially high frequencies.

https://doi.org/10.1371/journal.pgen.1009810.g006

This lack of local adaptive introgression is perhaps surprising given the genetic structure in mexicana associated with different ecotypes [49] and evidence for local adaptation within teosinte across elevation [31, 50, 51]. However, mexicana also has substantial standing variation and we find little evidence for hard sweeps, so one possibility is that local maize and local mexicana are adapting to the same environment by different available genetic paths, or even the same causal SNP on a different set of haplotype backgrounds. Older introgressed tracts may also offer more accessible paths for maize adaptation, having already purged some of their linked deleterious variation. Additionally, local exitinction and re-colonization by mexicana is common [37] and may contribute to a lack of local sourcing of adaptive haplotypes from contemporary mexicana populations.

Genomewide scan for selection on introgressed ancestry

We scanned the genome for two types of widespread selection on introgressed ancestry: consistent selection across populations creating an overall excess or deficit of introgression, and fitness trade-offs creating steep clines in mexicana ancestry across elevation. We used our MVN simulated ancestry frequencies to set false-discovery-rates for excess and deficits of mexicana and maize introgression as well as steeper than expected slopes between mexicana ancestry and elevation (see S31 Fig for model fit).

We find several regions with high introgression in both directions that are unlikely to be explained by shared demographic history alone (Fig 7A). These regions of adaptive introgression (< 5% FDR) are spread across the genome and cover a small fraction (<0.5%) of the genome in both subspecies. We additionally find evidence of adaptive parviglumis introgression into each subspecies (S32 Fig). We do not have power to determine if individual genes or regions are barriers to introgression because zero introgressed ancestry is not unusual under our simulated neutral model, given both low genomewide introgression and positive ancestry covariance between admixed populations (Fig 7).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Genomewide scan for selection on introgressed ancestry.

(A) Mean mexicana ancestry introgressed into sympatric maize populations and mean maize ancestry introgressed into sympatric mexicana populations. (B) Slope of mexicana ancestry proportion over a 1 km elevation gain in sympatric maize and mexicana populations. In both (A) and (B) the blue lines show the 5% false discovery rates, set using multi-variate normal simulations. Positions for Inv4m [50] and the mhl1 locus [65] were converted to the maize reference genome v4 coordinates using Assembly Converter (ensembl.gramene.org). Chromosome numbers are placed at the centromere midpoint (approximate centromere positions are from [66]).

https://doi.org/10.1371/journal.pgen.1009810.g007

Additionally, we identify outlier loci across the genome where mexicana ancestry forms steep clines across elevation (Fig 7). Our top candidate for strong associations between introgression and elevation in maize is Inv4m, a large 14 Mb inversion on chromosome 4 previously identified to have introgressed into high elevation maize populations [43–45, 67]. This inversion maintains steep elevational clines within teosintes [50], overlaps QTLs for leaf pigmentation and macrohairs [42], and is associated with increased yield in maize at high elevations and decreased yield at low elevations [67], but has thus far eluded functional characterization of genes within the inversion [67].

Our second strongest association co-localizes with macrohairless1 (mhl1), a locus on chromosome 9 that controls macrohair initiation on the leaf blade [65] and is associated with a major QTL explaining 52% of macrohair variation between high and low elevation teosinte mapping parents [42]. Within teosintes, populations of the lowland ancestor of maize, parviglumis, show convergent soft sweeps at the mhl1 locus not shared by mexicana [32]. Macrohairs are characteristic highland phenotypes in teosinte and maize and are thought to confer adaptive benefits through insect defence and/or thermal insulation [41, 42]. We identified a 3 Mb outlier region within the larger mhl1 QTL which we analyzed further using PCA. We found three genetic clusters along the first principal component, evidence that an inversion polymorphism (hereafter Inv9f) maintains differentiation between maize/parviglumis and mexicana haplotypes across this region (S33 and S34 Figs). Additionally, we found evidence that the mexicana-type allele at the inversion segregates at low frequency within our lowland parviglumis reference population. Based on reduced diversity, the lowland maize/parviglumis-type allele at the inversion is likely derived (S6 Table). Thus mexicana-alleles at Inv9f could have been inherited by maize either through introgression or incomplete lineage sorting before selection pushed them to high frequency in highland populations.

The clinal patterns of admixture that we observe at inversions Inv4m and Inv9f suggest they contribute to elevation-based adaptation in maize, with variation in their fitness impacts even within the historic elevational range of mexicana. While our highest peaks localize with regions previously associated with characteristic highland phenotypes, many additional outlier regions with steep increases in mexicana ancestry across elevation have undiscovered associations with local adaptation to elevation. Additionally, outliers for steep ancestry slopes across elevation in sympatric mexicana suggest that introgression from maize into mexicana may facilitate adaptation in mexicana at the lower end of its elevational range.

Selection at candidate domestication genes

We hypothesized that domestication genes will be barriers to introgression bilaterally between maize and mexicana [43]. While we do not have power to identify individual outlier genes that have low introgression, we can test for enriched overlap between ‘introgression deserts’ and a set of putative domestication genes spread across the genome.

We examined introgression for a sample of 15 well-characterized domestication genes from the literature (see S7 Table), and compared them to the regions of the genome with the lowest 5% introgression of teosinte ancestry into sympatric maize and maize ancestry into sympatric mexicana (‘introgression deserts’). A small but enriched subset of these domestication genes overlap with introgression deserts in sympatric maize (5, P < 0.001) and likewise in sympatric mexicana (5, P = 0.001).

Among these 15 domestication genes, we find that teosinte branched1 (tb1), a key transciption factor that regulates branching vs. apical dominance [68, 69], overlaps introgression deserts in both maize and mexicana, consistent with tb1’s role at the top of the domestication regulatory hierarchy [70]. We also find evidence for reduced introgression into both maize and mexicana at teosinte glume architecture1 (tga1) [71, 72], which is associated with ‘naked’ edible grains.

Another six domestication genes have low introgression in one direction only [73–77] (see S7 Table). Among these, sugary1 (su1) in the starch pathway has low maize ancestry in mexicana but shows a steep increase in introgressed mexicana ancestry proportion with elevation in maize (+0.95 per km, < 5% FDR), which suggests this gene has pleiotropic effects on non-domestication traits in maize, with fitness trade-offs across elevation. Sugary1 mutations modify the sweetness, nutrient content and texture of maize kernels (e.g. sweet corn), but also affect seed germination and emergence at cold temperatures [78], candidate pleiotropic effects that could be more deleterious at higher elevations.

The remaining seven domestication genes do not overlap introgression deserts in either subspecies despite evidence for their roles in domestication: zfl2 (cob rank) [79–81], pbf1 (storage protein synthesis) [82], ba1 (plant architecture) [83], ae1 (starch biosynthesis) [76], ra1 and ra2 (inflorescence architecture) [84, 85] and ZmSh1–5.1+ZmSh1–5.2 (seed shattering) [75]. Despite evidence of introgression at many domestication loci, maize populations retain all of the classic domestication traits, and mexicana populations maintain ‘wild’ forms. Epistasis for domestication traits [47] could help explain this discrepancy if compensatory effects from other loci contribute to maintaining domestication traits in admixed highland maize, or if key domestication alleles segregate at moderate frequencies within mexicana but do not have the same phenotypic effects in a teosinte background.

Selection within the flowering time pathway

Flowering earlier is adaptive in high-elevation environments where days are cooler and there are fewer total growing degree days in a season. We therefore expect an excess of introgressed mexicana ancestry at flowering time genes that may contribute to adaptive early flowering in highland maize. For example, the mexicana allele at High PhosphatidylCholine 1 (HPC1) has recently been shown to reduce days to flowering and confer a fitness benefit in maize at higher elevations [86], and we find that the steepest clinal SNP within HPC1 has a +0.79 mexicana ancestry proportion increase per km (FDR < 10%), adding support for HPC1’s role in adaptive earlier flowering at higher elevations in Mexican maize varieties (S35 Fig).

We tested more broadly for enriched selection within the flowering time pathway using a set of 849 candidate flowering time genes [87, 88] and a more stringent 5% FDR cutoff. No genes from the core flowering time pathway [87] and only 13/806 other candidate flowering time genes [88] (+/- 20kb) overlap outlier regions with steep increases in mexicana introgression with increasing elevation in sympatric maize, which matches expected overlap by chance (1.5%, P = 0.93). Thus steep clinal introgression patterns, indicative of strong fitness trade-offs across elevation, are the exception, not the rule, for flowering-time related genes. While maladaptive flowering times have strong fitness consequences, flowering time is also a highly polygenic trait [89], which may reduce the strength of selection at most individual genes to below what we can detect using steep ancestry clines. It is alternatively possible that mexicana alleles show adaptive benefits across the entire range sampled (moderate to high elevation), but we find that only 11/849 candidate flowering time genes overlap high mean mexicana introgression outliers at a 5% FDR (P = 0.95).

Population sampling

We used maize and mexicana seed accessions sampled from locations across Mexico in 2008 [43] and currently stored at UC Davis. We included 14 maize and 14 mexicana accessions that are paired populations sampled in sympatry from the same locations: Ixtlan*, Amatlan, Penjamillo, Puruandiro*, Nabogame*, El Porvenir*, Santa Clara*, Opopeo*, Xochimilco*, Cocotilan, Tlapala, San Pedro*, Jicaltepec and Tenango del Aire* (see S1 Table). A previous study of crop-wild admixture genotyped different maize and mexicana individuals from 9 of these locations (marked with *), using the Illumina MaizeSNP50 Genotyping BeadChip [43]. In addition, we chose three population accessions to sequence as a mexicana reference panel: Puerta Encantada and Malinalco were chosen because they have no record of contemporary maize agriculture nearby and a third population, Amecameca, was added as a complement to these two reference populations because it grows at a higher elevation, beyond the historical range of parviglumis.

At each sampling location, multiple ears from maternal plants were collected for seed. Population accessions varied in the number of maternal plants with viable seeds. When available, we planted multiple seeds within each ear but only randomly selected one individual for sequencing from the plants that successfully germinated in the greenhouse.

DNA extraction and sequencing

We extracted DNA from leaf tissue and then prepared sequencing DNA libraries using a recently published high-throughput protocol (“Nextera Low Input, Transposase Enabled protocol” [90]) with four main steps: (1) DNA shearing and tagmentation by the Nextera TD enzyme, (2) PCR amplification (Kapa2G Robust PCR kit) and individual sample barcoding (custom 9bp P7 indexing primers) (3) library normalization and pooling, and (4) bead-based clean-up and size-selection of pooled libraries. We sequenced the resulting pooled libraries using multiple lanes on Illumina HiSeq 4000 and Novaseq 6000 machines (paired-end 150 bp reads).

To address low sequencing output from some libraries, we re-sequenced 26 libraries (and merged output) and replaced 53 lower-coverage libraries with a higher-coverage library prepared from another seed grown from the same half-sibling family. We excluded 7 samples from analysis because their final libraries did not yield sufficient sequencing output (<0.05x coverage after filtering reads for mapping quality). We additionally removed one lane of sequencing (58 samples) from the study after determining a labelling error had occurred for that plate.

In total, we obtained whole genome sequences for 348 individuals (1.0x average coverage, range: 0.1–2.4x). Of these samples, 43 are mexicana from the three reference populations, with a total of 34.1x combined coverage. The remaining samples are maize and mexicana from paired sympatric populations, 262 of which have sufficient coverage for local ancestry inference (≥ 0.5x, 6–12 per sympatric population, see S36 Fig). Raw sequencing reads for these low-coverage maize and mexicana genomes are available at NCBI (PRJNA657016).

Reference genome and recombination map

We used version 4 of the B73 maize reference genome [66] (Zea_mays.B73_RefGen_v4.dna.toplevel.fa.gz, downloaded 12.18.2018 from Gramene).

To find local recombination rates, we converted marker coordinates from a published 0.2 cM genetic map [62] to the v4 maize genome using Assembly Converter (ensembl.gramene.org). We removed any markers that mapped to a different chromosome or out of order on the new assembly, and extended the recombination rate estimates for the most distal mapped windows to the ends of each chromosome (S37 Fig). From this map, we used approx() in R (v3.6.2 [91]) to estimate the cM position for any bp position, based on linear interpolation.

Read mapping and filtering

First, we checked read quality using fastQ Screen (v0.14.0 [92]) and trimmed out adapter content from raw sequencing reads using the trimmomatic wrapper for snakemake (0.59.1/bio/trimmomatric/pe) [93]. We mapped trimmed reads to the maize reference genome using bwa mem (v0.7.17 [94]). We then sorted reads using SAMtools (v1.9 [95]), removed duplicates using picardtools (v2.7.1) MarkDuplicates and merged libraries of the same individual sequenced on multiple lanes using SAMtools merge. In all subsequent analyses in the methods below we filtered out reads with low mapping scores (< 30) and bases with low base quality scores (< 20).

High-coverage Tripsacum genome sequencing

In addition to low-coverage genomes for maize and mexicana, we selected a Tripsacum dactyloides individual as an outgroup and sequenced it to high coverage. This individual is an outbred ‘Pete’ cultivar (rootstock acquired from the Tallgrass Prairie Center, Iowa, USA). We extracted genomic DNA from leaf tissue using the E.Z.N.A.^® Plant DNA Kit (Omega Biotek), following manufacturer’s instructions, and then quantified DNA using Qubit (Life Technologies). We prepared a PCR-free Truseq DNA library and sequenced it with an Illumina HiSeq2500 rapid run (paired-end 250 bp reads). We generated a total of 136.53 Gb of sequencing for this individual, available at NCBI (SRR7758238). For the following analyses that use Tripsacum as an outgroup, we randomly subsampled 50% of reads using seqtk, for approximately 30x coverage. We mapped reads to the maize reference using the pipeline described above, and additionally capped base quality scores with the ‘extended BAQ’ model in SAMtools [96], which reduces the influence of bases with lower alignment quality.

Additional genomes from published sources

For a maize reference population, we used 55 previously published high-coverage genomes from a ‘Tuxpeño’ maize population grown at 983 m near Palmar Chico (NCBI: PRJNA616247 [53, 54]).

For a parviglumis reference population, we used 50 previously published high-coverage lowland individuals sampled from the ‘Mound’ population at 1,008 m near Palmar Chico [53–55] (NCBI: PRJNA616247, see S2 Table). These maize and parviglumis reference populations were sampled about 1 km apart from each other but maintain high F_ST and are isolated by differences in flowering time [54]. Both maize and parviglumis reference populations are allopatric to mexicana, growing well below its elevational range. We mapped and filtered reads for these reference maize and parviglumis individuals using the pipeline described above and capped base quality scores using BAQ.

SNP calling

We called SNPs using a combined panel of the 348 low-coverage maize and mexicana genomes sequenced in this study and 105 high-coverage published genomes for reference maize and parviglumis described above. We used ANGSD (v0.932 [97]) to identify variant sites with minor allele frequencies ≥ 5% in the total sample based on read counts (‘angsd -doMajorMinor 2 -minMaf 0.05 -doCounts 1 -doMaf 8’). In addition to mapping and base quality filters (‘-minMapQ 30 -minQ 20’), we capped base qualities using the extended per-Base Alignment Quality algorithm (‘-baq 2’ [96]) and removed sites that did not have at least 150 individuals with data or had sequencing depth exceeding 2.5x the total sample mean depth. To apply this total depth filter, we estimated mean depth (‘angsd -doCounts 1 -doDepth 1 -maxDepth 10000’) for 1000 regions of length 100bp randomly sampled using bedtools (v2.29.0 [98]). In total, we identified 61,612,212 SNPs on the assembled chromosomes. In conjunction with SNP calling, we produced genotype likelihoods for each individual at these variant sites using the SAMtools GL method [95] implemented in ANGSD (‘-GL 1 -doGlf 2’).

Global ancestry inference

To estimate genetic relationships between populations and genomewide ancestry proportions, we used methods specific to low-coverage data that rely on genotype likelihoods, rather than called genotypes. Because these methods are sensitive to SNPs in high linkage disequilibrium (LD), we thinned genotype likelihoods to every 100th SNP (∼4kb spacing) [99]. To first confirm that maize and mexicana ancestry form the major axis of genetic variation in our sample, we estimated the genetic covariance matrix between all individuals using PCAngsd (v0.98.2 [100]) and visualized principal components computed using eigen() in R. We then estimated global ancestry proportions using the same thinned genotype likelihood files as input to NGSAdmix [56], using K = 3 clusters. Clusters clearly mapped onto the three reference groups, which we used to label the three ancestry components as ‘maize’, ‘mexicana’ and ‘parviglumis’.

Local ancestry and timing of admixture

We inferred local ancestry across the genome using a hidden Markov model that is appropriate for low-coverage data because it models genotype uncertainty down to the level of read counts for all admixed individuals (ancestry_hmm [58]). This method relies on allele counts from separate reference populations to estimate allele frequencies for each ancestry. Because some of our reference individuals have too low of coverage to accurately call genotypes, we randomly sampled one read per individual to get unbiased frequency estimates for major and minor alleles at each site (‘angsd -doCounts 1 -dumpCounts 3’). To maximize ancestry-informativeness of sites in this analysis, we identified SNPs in the top 10% tail for any of the three possible configurations of the population branch statistic between maize, mexicana and parviglumis reference populations, thereby enriching for SNPs that distinguish one reference population from the other two. We only considered SNPs with at least 44 reference maize, 12 reference mexicana, and 40 reference parviglumis individuals with sequencing coverage at a site. For these SNPs, we estimated reference population allele frequencies from angsd (‘angsd -doMajorMinor 3 -GL 1 -baq 2 -doMaf 1’), then estimated pairwise F_ST’s using the Hudson 1992 estimator [101] as implemented by Bhatia et al. 2013 [102] to calculate the population branch statistic [103] (PBS equation from pg 8 of the supplement). We then calculated genetic distances between SNPs using the maize recombination map and filtered our enriched variants to have minimum 0.001 cM spacing between adjacent SNPs.

Running ancestry_hmm jointly infers local ancestry for each individual and the timing of admixture. This HMM method assumes a neutral demographic history in which a constant-size population receives a pulse of admixture t generations in the past, and finds the t that maximize the likelihood of the observed read counts and hidden local ancestry state across each admixed individual’s genome. The timing of admixture defines the generations for possible meiotic recombination between ancestry tracts, and therefore scales the transition probabilities between hidden ancestry states. We ran ancestry_hmm under a three-way admixture model: a population is founded 10,000 generations ago by mexicana, then receives a pulse of parviglumis ancestry at t_parv (prior: 1000 generations ago, range: 0–10,000) and a pulse of maize ancestry at t_maize (prior: 100 generations ago; range: 0–10,000). Because these Zea mays subspecies are all annual grasses, generations can equivalently be interpreted as years since admixture. In addition to t_parv and t_maize, the HMM outputs the posterior probabilities for homozygous maize, homozygous mexicana, homozygous parviglumis, and heterozygous ancestry for each individual at every site. We analysed each sympatric maize and mexicana population separately, using the population’s mean NGSAdmix global ancestry estimate as a prior for mixing proportions, an approximate effective population size (Ne) of 10,000 individuals, genetic positions for each SNP based on the maize linkage map, an estimated sequencing base error rate of 3 × 10⁻³, and the three-way admixture model described above. We ran ancestry_hmm with an optional setting to bootstrap 100 random samples of 1,000-SNP genomic blocks to estimate uncertainty around the estimated generations since admixture (t_parv and t_maize). To test the sensitivity of the HMM to our choice of Ne, we re-ran ancestry_hmm with two other Ne’s that differ by an order of magnitude (Ne = 1k, 100k), but did not analyze these results further after finding high correspondence for both local ancestry and timing estimates (S38 Fig).

To get a single point estimate for local ancestry at a site for an individual, we computed a sum of ancestry from the different possible ancestry states, weighted by their posterior probabilities, e.g. mexicana ancestry proportion = P(homozygous mexicana) + ½P(heterozygous mexicana-maize) + ½P(heterozygous mexicana-parviglumis). In addition, for analyses that require ancestry tract positions, we assumed that the estimated ancestry at a focal site extends halfway to the next site with a local ancestry estimate.

Diversity within ancestry

Using local ancestry estimates from the HMM, we identified high-confidence homozygous ancestry tracts (posterior > 0.8). We filtered individual bams for reads that overlap these tracks and used the resulting filtered bams to calculate diversity within maize, parviglumis, and mexicana ancestry, separately. We estimate diversity using the ANGSD/realSFS framework which is appropriate for low-coverage sequence data it takes into account uncertainty in both genotypes and variant sites. We created a concensus fasta sequence for Tripsacum (‘angsd -doFasta 2’) to use as the ancestral state for polarizing the unfolded site frequency spectrum in these analyses.

For each population and ancestry, we estimated the site allele frequencies (‘angsd -doSaf 1 -GL 1’) and subsequently estimated the genomewide site frequency spectrum (SFS). We then used this SFS as a prior to estimate within-ancestry pairwise diversity (π) genomewide from the site allele frequencies (‘realSFS saf2theta’).

For each pair of populations and ancestry, we additionally used realSFS to estimate the two dimensional SFS from the individual population site allele frequencies genomewide. We then used this 2D SFS as a prior to estimate genomewide within-ancestry F_ST between the two populations (‘realSFS fst index -whichFst 1’). This call uses Hudson’s F_ST estimator [101] as parameterized in [102].

Effect of local recombination rate on introgressed ancestry

To estimate the effects of linked selection and recombination rate on genomewide introgression patterns, we compared introgressed ancestry estimates across genomic quintiles. Based on a 0.2 cM-resolution recombination map [62] for maize, we merged adjacent recombination windows into larger 1 cM non-overlapping windows and calculated each window’s mean recombination rate and overlap with coding base pairs (bedr ‘coverage’) [104]. We retrieved coding regions (‘CDS’) using gene annotations from Ensembl (ftp://ftp.ensemblgenomes.org/pub/plants/release-41/gff3/zea_mays/Zea_mays.B73_RefGen_v4.41.chr.gff3.gz, dowloaded 11.6.2018). We sorted windows into quintiles for either recombination rate or coding density (bp/cM). Each quintile covers approximately ⅕ of the genome based on physical bp.

i. NGSAdmix estimates

To estimate ancestry proportions for each recombination rate quintile, we first reduced LD by thinning to 1% of SNPs (every 100th) and ran NGSAdmix 5 times separately (once per quintile) with K = 3 clusters. We assigned ‘maize’, ‘mexicana’ and ‘parviglumis’ labels to the ancestry clusters based on majority assignment to the respective reference panels.

To bootstrap for uncertainty, we re-sampled 1 cM windows with replacement from each quintile 100 times, and re-ran NGSAdmix on the resulting bootstrap SNP sets. Genetic clusters could be unambiguously assigned to maize, parviglumis, and mexicana ancestries in all but 2 bootstrap replicates for the lowest recombination quintile and 3 for the highest coding density quintile, so we dropped this small number of replicates with unclear ancestry assignment from the bootstrap analysis. We calculated 95% percentile bootstrap confidence intervals for the estimated admixture proportions, and the Spearman’s rank correlation between the recombination rate (or coding bp per cM) and admixture proportion ranks for each quintile.

We also tested for a difference in ancestry slopes with elevation across different recombination rate quintiles by fitting a linear model with an elevation by recombination quintile interaction term: mexicana ancestry ∼ elevation + r + elevation*r. Using lm() in R, we fit this model for sympatric maize and sympatric mexicana separately, treating quintiles as a numeric scale 0–4.

ii. f₄ estimates

In a complementary analysis, we used a ratio of f₄ statistics as an alternative method to estimate ancestry proportions by quintile. The f₄ statistic measures shared genetic drift (allelic covariance) between populations in a phylogeny, due to either shared branch lengths or admixture events in the evolutionary history relating these populations. Excess shared drift with one population from a pair of sister populations in the tree is a signature of admixture, analogous to the ABBA-BABA test [105], and a ratio of two f₄ statistics can be used to quantify the admixture proportion. Assuming the basic phylogenetic tree relating our reference populations (((parviglumis, maize), mexicana), Tripsacum) in S5 Fig, we can estimate α, the proportion of ancestry inherited from mexicana in a focal sympatric population, as follows [105, 106]: The denominator of this statistic estimates the branch length leading to parviglumis and maize that separates these sister subspecies from mexicana; the full ratio estimates the proportion of this branch that separates sympatric population X from parviglumis and maize, i.e. the mexicana ancestry in X. Because the f₄ statistic is sensitive to additional unmodeled admixture within the tree, we limited our reference mexicana group to individuals from just one of the three reference populations (Amecameca), which showed no evidence of admixture in our global ancestry analysis (see Fig 2).

For each 1 cM window across the genome, we used ANGSD to calculate ABBA-BABA statistics from observed read counts for the 4 populations in the numerator and denominator of the α estimator separately (‘angsd -doabbababa2 1 -remove_bads 1 -minMapQ 30 -minQ 20 -doCounts 1 -doDepth 1 -maxDepth 10000 -useLast 1 -blockSize 5000000’). From the resulting output files, we summed the negative of the ABBA-BABA numerator (‘Num’) and divided by the total number of included sites (‘nSites’) across all 1 cM windows within a quintile to get the f₄ statistic [107].

We then calculated the Spearman’s rank correlation between the recombination rate quintiles and admixture proportion ranks for these quintiles. We calculated simple bootstrap confidence intervals for our ancestry estimates and correlations by re-sampling 1 cM windows within quintiles with replacement 10,000 times and re-calculating the f₄ ratios and resulting rank correlation across quintiles to construct 95% percentile confidence intervals. We repeated this analysis using quintiles based on coding bp per cM in place of recombination rate (cM/Mbp).

iii. Local ancestry estimates

We also calculated the Spearman’s rank correlation between local recombination rate (or coding bp per cM) and local ancestry proportion at the level of individual 1 cM windows. For each window, we averaged local ancestry estimates from the HMM across all individuals within sympatric maize, and separately, sympatric mexicana. We then calculated simple bootstrap confidence intervals for our local ancestry estimates and local recombination rate (or coding bp per cM) by re-sampling 1 cM windows across the genome with replacement 10,000 times and re-calculating the rank correlation across windows to construct 95% percentile confidence intervals.

Local ancestry simulations

We simulated maize, mexicana and parviglumis ancestry population frequencies marginally using a multivariate-normal null model, e.g. where is the vector of mean mexicana ancestry frequencies genomewide for each sympatric population and K is the marginal mexicana ancestry variance-covariance matrix relating these 14 populations and estimated from the empirical data. The diagonal entries of the K matrix capture the expected variation in local ancestry across the genome within populations due to drift and random sampling. The off-diagonals capture ancestry covariances between populations created by shared gene flow and drift post-admixture: at loci where one population has an excess of introgression, other admixed populations with shared demographic history will also tend to have an excess of introgression.

To construct K, we calculated the covariance in ancestry between each pair of populations i and j using all L loci with local ancestry calls genomewide:

Above, Anc_i,l and Anc_j,l are local ancestry frequencies at a locus l while α_i and α_j are the mean local ancestry frequencies across the genome for populations i and j.

For sympatric maize and sympatric mexicana separately, we calculated the empirical K matrix between populations from all 14 sympatric locations, and then took 100,000 independent draws from their MVN distribution, thereby simulating each focal ancestry marginally for all populations at 100,000 unlinked loci. Because ancestry frequencies are bounded at [0, 1] but normal distributions are not, we truncated any simulated values outside of this range.

Introgression peaks shared between populations

To characterize introgression peak sharing between individual populations, we defined ‘ancestry peaks’ as sites where a population has over 2 standard deviations more introgressed ancestry than the genomewide mean. We counted the number of peaks that are shared between all pairs and combinations of populations. To compare these results to our null model, we also counted the number of introgression peaks shared by populations in our simulated dataset, using the 2 s.d. cutoff set by the empirical data to define peaks.

Because mexicana ancestry shows significant diversity, we additionally characterized diversity for mexicana ancestry peaks introgressed into maize. For all introgressed ancestry outlier regions in a focal maize population, we used ANGSD to estimate pairwise diversity within the population (π) and differentiation (F_ST) between the focal sympatric maize populations and their local sympatric mexicana population. We focused on the mexicana ancestry within peaks by limiting our diversity estimates to only include high-confidence homozygous mexicana ancestry tracts (posterior > 0.8). For these analyses, we pooled information across outlier peaks, but distinguish between introgression peaks exclusive to the focal population and introgression peaks shared between the focal population and at least 3 other sympatric maize populations. We used global estimates of the SFS and 2D SFS as priors to estimate π and F_ST for the subsets of the genome within introgression peaks, and otherwise followed the same methods listed above in ‘Diversity within ancestry’.

Genomewide scan for ancestry outliers

For sympatric maize and mexicana separately, we calculated the mean introgressed ancestry across all individuals at a locus, and fit a linear model using lm() in R to estimate the slope of mexicana ancestry frequencies for sympatric populations across elevation (km): mexicana ancestry ∼ elevation. We then repeated these summary statistics for every locus with an ancestry call in the empirical data and each simulated locus in the MVN simulated data.

We calculated 5% false-discovery-rate (FDR) cutoffs for high and low introgressed ancestry using the Benjamini-Hochberg method [108] and simulation results to estimate the expected frequency of false-positives under our null MVN model (one-tailed tests). We repeated this approach to identify outlier loci with steep positive (or negative) slopes for mexicana ancestry across elevation at a 5% FDR.

Test for reduced introgression at domestication genes

To test whether domestication genes are unusually resistant to introgression, we first defined ‘introgression deserts’ as regions with the lowest 5% genomewide of teosinte introgression into sympatric maize (or, separately, maize introgression into sympatric mexicana). We then looked up v4 coordinates on Ensembl.org for genes associated with maize domestication in the literature (S7 Table), and used bedtools ‘intersect’ to identify which of these genes ±20 kb overlap introgression deserts. To test for significance, we randomly shuffled the gene positions across the genome (bedtools ‘shuffle’) 1000 times and re-calculated overlap with introgression deserts for each permuted data set.

Test for selection within the flowering time pathway

We identified a list of 48 core flowering time pathway genes from the literature [87], and a broader list of 905 flowering time candidate genes [87, 88]. From the combined set, we included 849 total genes (43 core pathway) which we were able to localize on assembled autosomes of the v4 reference genome using MaizeGDB gene cross-reference files [109]. We counted the number of genes ±20 kb that intersected with outlier regions for steep increases in mexicana introgression with elevation (and, separately, high mexicana introgression) in sympatric maize populations (< 5% FDR) using bedtools ‘intersect’, then tested for significance by repeating this analysis with 1000 randomly shuffled gene positions.

Analysis pipeline and data visualization

We constructed and ran bioinformatics pipelines using snakemake (v.5.17.0 [110]) within a python conda environment (v3.6). We analyzed and visualized data in R (v3.6.2 [91]) using the following major packages: tidyverse (v1.3.0 [111]), viridis (v0.5.1 [112]), bedr (v1.0.7 [104]), boot (v.1.3_25 [113, 114]), gridExtra (v2.3 [115]), ggupset (v0.3.0 [116]) and tidygraph (1.2.0 [117]). All scripts can be found on our gitHub repository, https://github.com/ecalfee/hilo, which also includes a full list of software and versions (see envs/environment.yaml).