Inv4m haplotypes are strongly diverged and associated with local adaptation in agronomic traits

Romero-Navarro et al [33] identified Inv4m as a large-effect QTL in a multi-environment trial of landrace hybrids grown in Mexico, and observed that Inv4m frequencies were highly differentiated between highland, subtropical and lowland tropical populations. We repeated this analysis to study the geographic distribution of Inv4m in more detail.

We determined the Inv4m genotype of 4845 maize plants from the SeeD-maize GWAS panel using published unimputed genotype-by-sequencing (GBS) data. Of these lines, 707 were homozygous for the inverted minor haplotype, and 351 were heterozygous. Of the 585 plants carrying at least one allele of the minor haplotype and with complete with geographic information, all but 7 were from Mexico, with the majority collected from the central highlands (Fig 1A). Therefore, to assess evidence of local adaptation, we modeled the association of Inv4m genotype with elevation among the 1757 Mexican plants. In Mexico, 1186 and 381 plants were homozygous for the alternate non-inverted and inverted haplotypes, respectively, and 190 were heterozygous, a distribution that significantly differs from Hardy-Weinberg expectations (D = -252.0, p = 1.91e-197). Genotypes at Inv4m were strongly associated with elevation, as previously reported [33] (Fig 1B).

Download:

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

Fig 1. Association of Inv4m genotype with environmental factors and agronomic traits.

A. Geographic locations for each of the 1757 Mexican plants genotyped by GBS, colored by their imputed genotypes at Inv4m. B. Association of Inv4m and elevation. Each point shows the mean frequency of the “High” allele at the Inv4m locus among plants from landraces collected in each 100m bin. The ribbon shows a loess fit (± 2SE) to the logit-transformed frequencies weighted by the number of landraces in each elevation bin. Bins with fewer than 10 landraces were excluded (those with elevation >2700m). C F_ST between 4 landraces from Mexican highlands and 5 landraces from Mexican lowlands [36] across an ≈ 130Mb region of chromosome 4. Points show average F_ST across 1Mb regions with a step size of 100Kb. D. Average pairwise nucleotide diversity estimates across the same regions within each population and between the two populations. The boundaries of Inv4m from [31] are denoted by vertical lines.

https://doi.org/10.1371/journal.pgen.1009213.g001

While useful for genotyping individuals for the major haplotypes at Inv4m, this GBS dataset was not appropriate for quantifying genetic divergence within or among haplotypes because data are only available for variable sites, and these sites may be impacted by ascertainment bias against the rarer Inv4m haplotype. We therefore used previously published whole-genome-sequencing data from 4 highland and 5 lowland maize landraces previously shown to carry alternate alleles at Inv4m to calculate nucleotide diversity across the Inv4m locus [36]. These data showed that the Inv4m locus is considerably more diverged between the highland and lowland populations than flanking regions along chromosome 4 (Fig 1C), with average F_ST approximately 0.38 across the locus. Based on within and between-population pairwise nucleotide diversity statistics (π) the highland allele is more diverse than the lowland allele inside Inv4m, and more diverse than much of the surrounding genomic regions along chromosome 4 (Fig 1D). Therefore, the two haplotypes are relatively old and have shared very little gene flow for thousands of generations.

The near-fixation of alternative alleles of Inv4m along the elevation gradient in Mexico strongly supports the idea that Inv4m is under selection and contributes to local adaptation. But this doesn’t itself help identify why, i.e. what traits are controlled by Inv4m and are responsible for this selection? Romero-Navarro et al [33] identified strong associations of SNPs inside Inv4m with flowering time in at least some of their field trials. However, none of the individual SNP markers reported in this study were perfectly associated with the Inv4m genotype, and the association with one trait does not preclude it having additional affects on other adaptation-related phenotypes. Using data from the same SNP markers and field trials, Gates et al [37] identified associations between SNPs inside Inv4m and the plasticity to elevation for five of six agronomic traits: Days-to-Anthesis, the Anthesis-Silking Interval (ASI), Grain Weight-per-hectare, Bare cob weight, and Field Weight (but not Plant Height).

We refined these observations by re-running the genetic association models reported by Gates et al [37] to extract estimates of the effect of the inverted allele of Inv4m as a function of the elevation of the field trial. For each trait, except for Plant Height, the inverted (highland) haplotype of Inv4m was associated with trait variation in the direction consistent with local adaptation: earlier flowering, reduced ASI, and increased yield components in highland trials, while the opposite effect was observed in lowland trials (Fig 2). The highland haplotype was weakly associated with greater plant height across elevations, but the relationship was not significant (p = 0.11 for the main effect). These results are consistent with this single locus causing antagonistic pleiotropy [38] across elevation environments, which would explain the strong divergence in allele frequencies across elevation. The relationship between Inv4m genotype and these traits was not simply an indirect effect on flowering time; each relationship remained qualitatively the same even after accounting for the the effect of Days-to-Anthesis separately within each trial. The inversion appears to independently control both flowering and yield traits which is consistent with the hypothesis that Inv4m contains multiple important variants that each contribute to phenotypic differences between lowland and highland maize populations [33, 37].

Download:

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

Fig 2. Association of Inv4m genotype with agronomic traits in field trials changes as a function of elevation.

We modeled each trait as a function of Inv4m genotype, trial elevation, and tester line, with controls for main effects and responses to elevation of the genomic background. Black lines and ribbons show estimates of the effect of the highland allele of Inv4m as a function of trial elevation ± 2SE, based on conditional F-tests at the REML solutions of the random effect variance components. Green lines show estimates of the Inv4m effect in a model that additionally included effects of Days-to-Anthesis on the focal trait within each trial.

https://doi.org/10.1371/journal.pgen.1009213.g002

However, while compelling, due to the strong population structure in this diversity panel, we caution that the associations of Inv4m with the agronomic traits is still preliminary. Inv4m genotypes are highly correlated to overall genetic ancestry (as measured by PC1 calculated using all other chromosomes except chromosome 4) within Mexico. We corrected for ancestry using genome-wide kinship (again excluding chromosome 4), but if this correction was incomplete, it may have lead to false-positive associations with Inv4m. An alternative explanation for the phenotypic associations above is that each trait has a polygenic basis, with small-effect loci distributed throughout the genome, each of which has subtle allele-frequency differences across elevations. Therefore, any associations of traits with Inv4m in this population could be driven by the combination of many loci genome-wide all correlated with this dominant axis of genomic divergence. To confirm that Inv4m in fact causes important phenotypic effects we must experimentally break the association between Inv4m and the rest of the genome through experimental crosses, as we describe next.

Introgression of highland alleles of Inv4m regulate diverse phenotypes

We generated two sets of segregating families for alternative alleles of the Inv4m locus by crossing the reference line B73, which carries a lowland haplotype of Inv4m, to two highland landrace donors (called PT and Mi21). The two families were used as replicates of the effect of the highland haplotype of Inv4m to ensure that Inv4m-associated effects that we identified were not specific to a single highland landrace. Progeny were backcrossed to B73 for 5 generations and then selfed to generate a population of plants that were largely homozygous B73 across the genome, but segregated for the lowland and highland haplotypes at the Inv4m locus (Fig 3). We grew a total of 360 plants in growth chambers simulating two temperature environments (warm: 32C/22C and cold: 22C/11C), genotyped each plant using a diagnostic CAPS marker for Inv4m, and selected 81 plants homozygous for either the High or Low haplotypes for further analysis.

Download:

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

Fig 3. Plant families segregating for Inv4m.

A) Graphical overview of crossing scheme and experimental design. Red represents the high haplotype of Inv4m, green represents the landrace genome, and gold represents the B73 genome. We used the same crossing scheme design for both donor landraces. B) Genotypes called from RNA sequencing data across the maize genome. In each panel, horizontal rows of each chromosome represents the genotypes of a single plant from either the PT-derived or Mi21-derived families, with blue representing landrace alleles and grey the B73 background. Within each family, plants are sorted first by maternal plant (2 per family) and then by genotype at Inv4m. Vertical black lines on chromosome 4 delineate the location of Inv4m.

https://doi.org/10.1371/journal.pgen.1009213.g003

Phenotype measurements.

We sampled multiple tissues (Table 1) from the 81 selected plants at either the V1 or V3 developmental stage to measure the effect of Inv4m on genome-wide gene expression (S1 Fig). We selected nine total tissues (S2 Fig) with the goal of capturing as much variation in expression profiles as possible in maize seedlings, given the experimental constraints on how large we could let the plants grow in our growth chambers, based on the maize gene expression atlas [39]. Of the 432 tissue samples collected (nine tissues × two temperature treatments × two segregating families × two Inv4m arrangements × three biological replicates × two experimental replicates), we excluded 53 samples with fewer than 100,000 reads, leaving a total of 379 samples. We detected a total of 23,428 genes with 10-or-more counts in at least a third of the samples (with an average of 17,016 genes per tissue) for a total of more than 306,000 gene expression traits. We used multi-dimensional scaling (MDS) analysis to visualize the overall transcriptome variation among samples (S3 Fig). Samples clustered predominantly by tissue, with an additional slight separation by temperature treatment, but no visible separation by family or genotype at the Inv4m locus. This was expected because only ∼ 7% of the genomes differed among samples. Among tissues, all leaf samples except the S2 leaf base formed one major cluster, the two root tissues formed a second cluster, and the stem and SAM and S2 leaf base tissues formed separate individual clusters in the MDS plot.

Download:

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

Table 1. Developmental stage and description of tissues sampled for gene expression analysis.

https://doi.org/10.1371/journal.pgen.1009213.t001

In addition to gene expression, we quantified the effect of Inv4m on seedling emergence in each temperature treatment separately. Emergence times in the cold chamber averaged approximately 9 days, and plants with the High haplotype of Inv4m emerged 0.75 and 0.35 days earlier on average in the PT and Mi21 families, respectively (S1 Table). In the warm chamber, the average emergence time was approximately 4 days, and the High Inv4m haplotype had no detectable effect (S1 Table).

Quantification of residual landrace alleles in BC₅S₁ plants.

As a preliminary step in identifying Inv4m-associated traits in this population, we characterized each plant for the extent of residual landrace genome that remained after the five generations of backcrossing to B73. This residual landrace genome could bias both our expression results due to mapping biases, and confound our analysis of the effect of Inv4m if these residual regions contain functional variants between each landrace and B73.

We used RNAseq reads to identify residual landrace alleles at every expressed gene in each of the BC₅S₁ plants. We observed residual alleles in large regions flanking the Inv4m locus in plants from all families. In the PT segregating family, residual PT alleles were observed in a 57Mb window surrounding Inv4m. In the Mi21 family, residual Mi21 alleles were observed in an 18Mb region surrounding Inv4m (Fig 4A). The PT family also segregates for a large paracentric region of residual donor alleles on chromosome 5 and a small region on chromosome 2, and the Mi21 family segregates for a large paracentric residual region on chromosome 3 (Fig 3B). Beyond these large contiguous blocks, we identified another 821 and 52 genes in the PT and Mi21 families, respectively, that harbored high-confidence SNPs in the RNAseq data, yet were not contiguous with any of the large residual introgression regions (Fig 3). It is unlikely that there was sufficient recombination in the BC₅ populations to generate these independent blocks; rather these genes likely have moved genomic coordinates in the landraces relative to B73, and actually reside inside one of the large introgression regions [40, 41]. None of the residual genes had genotypes that were perfectly correlated with genotypes at the Inv4m haplotype, so we excluded them all from further analysis (S4 Fig).

Download:

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

Fig 4. Sequence and expression divergence between landrace donors is greater outside of the Inv4m boundaries than within.

A) Boxplots show the correlation between genetic and expression divergence between the two donor populations for genes inside Inv4m or outside Inv4m but in the region of shared introgression. Correlations were measured separately in each of the 18 tissue:temperature combinations. B) Boxplots show the proportion of shared SNPs per gene for genes inside or outside Inv4m. C) Diagram of PT and Mi21 introgressions containing Inv4m on chromosome 4.

https://doi.org/10.1371/journal.pgen.1009213.g004

We also used the RNAseq reads that did not map to the B73 genome to search for genes that may be present in the High haplotype of Inv4m but not in the Low haplotype, and thus may have been missed by our analysis. We assembled all un-mapped reads from samples carrying the PT or Mi21 haplotype at Inv4m using Trinity, and searched for de-novo transcripts with evidence for expression only in the samples carrying the High haplotype. We found 772 candidate transcripts. However, all were very lowly expressed. Only five had estimated transcript counts of at least 10 summed across all PT or Mi21 samples, and none had estimated transcript counts of at least 10 in two or more samples per family. Therefore, we found no evidence that the High haplotype of Inv4m carries high-expressed genes that are not present in the Low haplotype.

In the segregating families there were a total of 7,236 genes with PT alleles, and 4,095 genes with Mi21 alleles, of which 355 were within Inv4m. Of all genes from the landrace donor which were highly correlated with the Inv4m genotype (ρ > 0.5), 14% and 29% resided inside Inv4m in the PT and Mi21 families respectively. Therefore, many associations of traits with the Inv4m haplotype that we observe in one segregating family may be caused by functional variants inherited from the PT or Mi21 landraces in regions flaking the inversion, rather than functional variants inside the inversion itself. Variants specific to either the PT or Mi21 allele of Inv4m cannot be responsible for the adaptive value of the inversion. However, since population level variation in the regions flanking Inv4m was not correlated with Inv4m genotype (Fig 1A), Inv4m-associated effects that replicate in our two segregating families were likely caused by alleles located within Inv4m and which were common to at least these two samples of the highland haplotype.

We tested the likelihood that gene-regulatory variation was due to Inv4m, or the shared residual introgression flanking Inv4m. We quantified the genetic and expression divergence between families using plants that were homozygous either for the PT or Mi21 Inv4m alleles. The proportion of shared SNPs between PT and Mi21 families within Inv4m was significantly higher than in the flanking introgressed regions (Fig 4B, t = 10.937, df = 319.93, p-value < 2.2e-16). The same pattern occurred for gene expression variation attributed to landrace genotypes for each gene: the landrace genotype effect on genes inside Inv4m were more correlated than landrace genotype effects in flanking regions (Fig 4A, t = 5.1793, df = 33.162, p-value = 1.073e-05) across the 18 tissue:temperature combinations. Therefore, gene regulatory effects associated with Inv4m that replicate across the two donor populations are more likely to be caused by functional variation inside Inv4m rather than in residual landrace DNA present in each population (Fig 4). We also quantified the correlation between genetic dissimilarity and expression divergence between Inv4m sources (Mi21 and PT) across the shared introgressed region containing the inversion. We found that genetic divergence between Mi21 and PT was higher outside of the inversion that within the breakpoints, and that expression divergence was higher outside of the inversion boundary. We calculated expression divergence as the average log-fold change in expression averaged across tissue and temperature combinations between Mi21 and PT (S5 Fig). These two separate measures of divergence were significantly positively correlated (r = 0.80, p = 2.2e-16, df = 76).

Effects of Inv4m on genome-wide gene expression.

We measured the effect of the highland allele on each gene within each segregating family for all 18 tissue:temperature conditions to quantify Inv4m-associated gene expression effects, and used multivariate adaptive shrinkage [42] to increase precision and power by sharing information across the 18 conditions. Then, we classified Inv4m-effects as being successfully replicated if the sign of the effect was the same in both segregating families, and was significant (as defined as a local false sign rate (lfsr less that 5%) in both segregating families.

We excluded genes in genomic regions with residual landrace alleles to avoid biased effect estimates. 11,842 and 12,482 genes were associated with Inv4m in the PT and Mi21 segregating families, respectively, in at least one tissue:temperature condition using a 5% local false sign rate (lfsr) threshold for significance (Table 1), with 1,645-5,532 and 145-6,208 genes per condition in the PT and Mi21 families respectively. Of these genes, 39-607 per tissue replicated across both donor families, totaling 8–41% of the differentially expressed genes in the PT family, and 11–38% of the differentially expressed genes in the Mi21 family.

In contrast, we found little evidence for strong genotype-environment interaction effects for Inv4m on gene expression. We detected 46-435 genes per tissue in the PT family and 127-1408 genes per tissue in the Mi21 family with significant genotype-treatment interactions at a 5% lfsr threshold. However only 1-38 genes per tissue replicated between the two families (Table 2), suggesting that the majority may be allele-specific and therefore we were not confident that they were representative of the high haplotype of Inv4m in general.

Download:

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

Table 2. Number of differentially expressed genes in each population and their overlap using a local false sign rate threshold of 5% across tissues and temperature treatments.

https://doi.org/10.1371/journal.pgen.1009213.t002

Inv4m-regulated genes are enriched for diverse biological processes

The list of consensus Inv4m-regulated genes are reporters of the effect of Inv4m, and were distributed across the genome with no visible clustering by chromosome. To summarize these genome-wide responses to Inv4m genotype and explore whether Inv4m regulates one, or many, separate processes, we tested for enrichments of Gene Ontology (GO) terms within the sets of Inv4m-regulated reporter genes. For each tissue:treatment experiment, we separately quantified GO term enrichments among sets of reporter genes that were up-regulated, down-regulated, or regulated in either direction by Inv4m. We identified a total of 564 enriched categories overall, with 0-114 signifiantly enriched categories per tissue:temperature treatment at a 5% FDR. These GO terms provide candidate descriptors of the global effects of Inv4m. However, GO terms are highly redundant, so to reduce the number of categories and combine results across all tissue:temperature treatments, we collapsed terms into clusters by semantic similarity [43] and selected the most-enriched term across all tissue:temperature treatments in each cluster to label the sets of effects. After this filtering step, we ended with a list of twenty-one representative GO terms describing Inv4m effects across the three main GO ontologies: two cellular component terms, one molecular function term, and eighteen biological processes terms (Table 3).

Download:

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

Table 3. Representative Gene ontology (GO) terms for Inv4m-regulated genes.

A universal enrichment analysis was conducted on each tissue and temperature and directional (up-regulated, down-regulated, or both) combination for Inv4m-regulated genes. Terms were then ranked by enrichment score and grouped by a semantic similarity score of higher than 0.5. The top term in each semantic similarity group selected to represent the group.

https://doi.org/10.1371/journal.pgen.1009213.t003

Among the 564 enriched GO terms, a set of 8 terms involving photosynthesis stood out as being the most strongly enriched among the candidate Inv4m-regulated genes. These categories are largely over-lapping in their gene content, so we selected the term: GO:0009534: Chloroplast thylakoid from the Cellular Component ontology as a representative label for this Inv4m effect. Genes annotated with this term tended to be up-regulated by Inv4m in the basal (developmentally immature) section of the S2 leaf in the cool temperature treatment, but down-regulated in this tissue in the warm temperature treatment (S6 Fig). Inv4m-associated effects on these genes were present in one or the other of the families in other tissues, but with inconsistent effect directions so we cannot conclude that these were driven by Inv4m itself rather than by other linked regions of residual genomic introgression present in each family (S6 Fig). These results suggest that Inv4m may regulate some aspect of chloroplast development or activity. To test for a change in chloroplast activity, we inspected the expression of chloroplast-expressed genes. Genes expressed in the chloroplast are inconsistently represented by our RNAseq libraries because their transcripts are not generally polyadenylated [44]. However we were able to quantify expression of 39 chloroplast genes, 8 of which were differentially expressed in at least one tissue. Overall, plants with the high Inv4m haplotype had higher expression of chlorophyllic genes in the plastid genome in the tip of the S2 leaf the cool temperature, and lower average expression in the warm temperature. This mirrors the patterns we observed for chloroplast-related nuclear genes, but only in the more developmentally mature region of this leaf (S6 Fig).

Since the phenotypic association of Inv4m with flowering time was so strong, we collected a list of genes with known roles in regulating flowering from two sources [45, 46] and used this list as another term to test for Inv4m-effects. We removed genes that were within any landrace introgression block, and listed any gene regulated by Inv4m that occurs in either of the flowering gene lists (S3 Table). While several genes from this list were regulated by Inv4m, as a whole, this category was not enriched for Inv4m effects relative to the rest of the genome (p = 0.4).

Identifying candidate causal alleles within Inv4m

As a final step, we attempted to use the RNAseq data to identify candidate genes for the effect of Inv4m itself. The causal variants responsible for Inv4m’s effect must lie within or at the boundaries of Inv4m. And if the causal variants operate by altering the expression of Inv4m genes, we may be able to detect these by inspecting for Inv4m-effects on Inv4m-located genes. Furthermore, if the expression of these genes were responsible for the effects of the Inv4m locus, we would expect that the expression of these genes would be a significant predictor of at least a subset of the Inv4m-regulated reported genes (genome-wide), even after accounting for the genotype at the Inv4m locus.

To identify candidate genes within Inv4m, we measured the Inv4m genotype effect (i.e. cis effect) on each of the annotated genes located inside the Inv4m locus, and the correlation between the expression of these Inv4m-genes and all Inv4m-regulated genes outside of the locus (a total of 4642 unique genes, with a range of 89-713 genes per tissue; Fig 5). We repeated this analysis in each of the 18 tissue:temperature treatments. For cis-genotype effects to be counted, we required the effects to be significant (lfsr < 5%) and in the same direction between the two donor populations.

Download:

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

Fig 5. Scatter plot of candidate gene scans within Inv4m.

Each point represents one of the 160 genes inside Inv4m that were both differentially expressed and correlated with at least one Inv4m-regulated gene. The x-axis represents the proportion of Inv4m-regulated genes correlated with the Inv4m-gene (p-value< 0.05). The y-axis represents the proportion of tissues in which the Inv4m-gene is differentially expressed (lfsr< 0.05, and regulated in the same direction in both populations.) Red points are correlated with more than 6% of Inv4m-regulated genes on average, or are differentially expressed in more than 90% of the tissues in which it was expressed. Black points are transcription factors.

https://doi.org/10.1371/journal.pgen.1009213.g005

Overall, of 355 annotated genes within the boundaries of Inv4m, 224 were expressed in both donors. Of those, 182 were differentially expressed in the same direction in both donors in at least one condition, and 160 of those were significantly correlated with at least one Inv4m-regulated gene located on other chromosomes used to measure the global Inv4m effect, even after accounting for the effect of Inv4m itself (Fig 5). Of these, 6 genes in particular stood out as being correlated with a large number (> 6%) of the reporter genes: Zm00001d051885, Zm00001d051929, Zm00001d051995, Zm00001d052075, Zm00001d052101, Zm00001d052224, and 17 were differentially expressed between Inv4m alleles in 90% or more of the tissue:temperature combinations. Finally, 7 of the 160 candidate Inv4m genes were transcription factors. One of the seven transcription factors, Zm00001d052054, was differentially expressed in the same direction in every tissue, and is predicted to be a sequence-specific DNA binding transcription factor (S2 Table). We did not observe an enrichment of Inv4m-related genes near the inversion breakpoints, which would have implicated the breakpoints themselves as functional variants for Inv4m.

The screening steps above excluded one a priori candidate gene within Inv4m: the microRNA MIR172c, which is located inside Inv4m at coordinates: 4:174154928-174155050. MIR172 has known roles in regulating developmental transitions as well the development of pigmentation and macrohairs in maize [47] and therefore was a strong candidate given that increased pigmentation and macrohair density is common in highland landraces and Inv4m overlaps a QTL for pigmentation and macrohairs in teosinte [35]. The expression of the pre-miRNA 172c was too low for us to measure differential expression in all but the leaf sheath tissue, and was not associated with Inv4m genotype in that tissue. However, of 31 genes that are predicted targets of miR172, 6 were Inv4m-regulated in both populations in at least one tissue. Of these, four were up-regulated by the High haplotype of Inv4m, one was down-regulated, and one was upregulated in one tissue and downregulated in another tissue (S4 Table).

Population genetics of Inv4m

We downloaded unimputed genotype-by-sequencing (GBS) data from 94,726 loci on chromosome 4 for 4,845 maize plants from the SeeD-maize GWAS panel [33, 67] and ran a principal components analysis on all positions within the Inv4m locus (between 168832447 and 182596678 in AGPv2 coordinates, [31]) with < 25% missing data and minor allele frequencies > 0.05. PC1 explained 22% of the genetic variation among these plants in this interval. Scores on PC1 neatly divided plants into three groups, representing the two homozygous classes at Inv4m and their heterozygotes. We cross-referenced plants with landrace passport data from Germinate 3, the CIMMYT Maize Germinate Database germinate.cimmyt.org/maize/ and extracted country of origin, latitude, longitude and elevation records. All but 7 of the plants containing the minor haplotype at Inv4m were from Mexico, so to study associations with elevation and to calculate other diversity statistics, we subsetted to only those plants collected in Mexico.

To calculate the association of Inv4m with elevation, we divided landraces into 100m bins, calculated the haplotype frequency of the minor haplotype in each bin, and fit a loess curve to the log-transformed haplotype frequencies, weighted by the number of landraces in each bin. Based on this analysis, we labeled the minor haplotype in Mexico “High”, and the major haplotype “Low”. We used the R function HWExact from the HardyWeinberg R package to test genotype counts against the Hardy-Weinberg expectation. We calculated diversity statistics separately for plants homozygous for the “High” or “Low” haplotypes at Inv4m across all of chromosome 4 in 500 marker windows using TASSEL 5 [68]. Because many more plants were homozygous for the “Low” haplotype, we randomly sampled 371 landraces to calculate the diversity statistics to make the sample sizes equal.

To calculate nucleotide diversity statistics, we downloaded published whole genome sequencing reads from 9 Mexican landraces mapped to the AGPv4 reference genome [36] from /iplant/home/lilepisorus/landrace_AGPv4. As described by those authors, one of the six sequenced landraces from Mexican Highlands (RIMMA0677) carries the lowland haplotype at Inv4m and so was excluded. We calculated π within each population and F_ST between the two population using the Analysis of Next Generation Sequencing Data (ANGSD) v0.931-2 [69–72] across an ≈ 130Mb region including Inv4m. We lifted over coordinates of the inversion breakpoints using the EnsemblPlants Assembly Converter Tool https://plants.ensembl.org/Zea_mays/Tools/AssemblyConverter to: (171771502,185951149). We ran ANGSD with the following options:

1. -GL1 -uniqueOnly 1 -remove_bads 1 -only_proper_pairs 1
   -trim 0 -C 50 -minMapQ 20 -minQ 20
   -doMaf 2 -doMajorMinor 4 -doSaf 1
   following the tutorial at http://www.popgen.dk/angsd/index.php/Thetas,Tajima,Neutrality_tests. We calculated average between-population pairwise nucleotide diversity following the R scripts available here: https://github.com/tvkent/Pi-Between assuming non-reference alleles matched between populations. All analyses used a 1Mb sliding window with a step-size of 100Kb.

Association of Inv4m with agronomic traits

We re-analyzed phenotypic data from the F1 Association Mapping (FOAM) panel of Romero-Navarro et al [33] and Gates et al [37] to more fully characterize associations signatures of Inv4m. Full descriptions of this experiment and data access are described in those references. We downloaded BLUPs for each trait and line from Germinate 3, and kept lines with GBS genotype data from Mexico. We fit a similar model to the GWAS model used by Gates et al [37] to estimate the effect of Inv4m genotype on the trait’s intercept and slope on trial elevation, accounting for effects of tester ID in each field and genetic background and family effects on the trait intercept and slope using four independent random effects. We implemented this model in the R package GridLMM [73]. We extracted effect sizes and covariances conditional on the REML variance component estimates and used these to calculate standard errors for the total Inv4m effect as a function of elevation. To test whether the phenotypic effects of Inv4m on yield components could be explained as indirect effects via flowering time, we additionally re-fit each model using Days-To-Anthesis as a covariate with an independent effect in each trial.

Experimental material for isolating Inv4m

To directly assess phenotypic effects of the Inv4m locus, we selected two highland landrace accessions which both carry the High haplotype of Inv4m, Palomero Toluqueno (PT) and an accession from the Cónico landrace. These Inv4m landrace accessions were obtained through the International Maize and Wheat Improvement Center (CIMMYT); PT came from accession mexi5 (referred to henceforth as PT) and Cónico from accession Michoacán 21 (referred to henceforth as Mi21). B73 is a modern inbred from the United States that carries the non-inverted Inv4m haplotype. Both landraces were crossed with B73 and one resulting F1 individual from each cross was backcrossed to B73 for 5 generations, selecting on a diagnostic SNP for Inv4m each cycle with a cleaved amplified polymorphic sequence (CAPS) assay. The diagnostic SNP is at position 4:179617762 (Matt Hufford, Personal communication). DNA was extracted from leaf tissue using a Urea lysis buffer extraction protocol (https://github.com/RILAB/lab-docs/wiki/Wetlab-Protocols). Primers were designed to amplify the fragment of DNA carrying the diagnostic SNP (Forward: CTGAGCAGGAGATGATGGCCACTC; Reverse: GGAAAGGACATAAAAGAAAGGTGCA). Amplification consisted of 5 minute denaturation at 95°C, 35 cycles of 95-60-72°C for 30 seconds each, 7 minutes of final extension step at 72°C, followed by a 4°C hold. Amplified DNA was then digested with the Hinf 1 enzyme for 1 hour at 37°C, and the resulting product was run out on a 1% agarose gel for genotyping. Two of the resulting BC₅ individuals per population identified as heterozygous for Inv4m were self-pollinated to produce BC₅S₁ families segregating for Inv4m.

Growth chamber experiment to identify phenotypic effects of Inv4m

We planted seeds from the four segregating families (two parents per Inv4m donor) in the UC Davis controlled environment facility growth chambers. Chambers were programmed to mimic temperatures in Mexican lowlands (22°C night, 32°C day, 12 hr light) and highlands (11°C night, 22°C day, 12 hr light). Kernels were soaked in distilled H₂0 for 12 hours and planted in 10.2cm x 34.3cm nursery pots (Steuwe & Sons: CP413CH) in a substrate mixture composed of a 3:1 ratio of Sungro Sunshine Mix #1 to sand. Pots were organized in racks with 9 pots per rack (Steuwe & Sons: tray10). Plants were watered every other day with a 1x Hoagland nutrient solution, and emergence was recorded daily. The experiment was replicated and growth chambers were switched to account for variation between instruments between replicates (See S1 Fig for a graphical workflow). The first replicate of the experiment began March, 2017 and the second replicate began April, 2017.

Two seeds were planted in each pot, one in the center and one near the corner, and a total of nine tissues were sampled from the two plants when they reached specific developmental stages. These nine tissues were selected to maximize the diversity of gene expression profiles based on the transcription atlas of [39]. Corner plants were removed from the pot and sampled when they reached the V1 stage, while center plants were sampled when they reached the V3 stage (S2 Fig). Two tissue types were sampled from the V1 stage, and 7 tissue types were sampled from the V3 stage (Table 1). Sampling occurred between 2 and 4 hours after simulated sunrise. Plant tissue was placed in 2 ml centrifuge tubes, immediately flash frozen in liquid nitrogen, and stored at -70°C.

Inv4m Genotyping and RNA sequencing

We used the same DNA extraction and CAPS genotyping methods as previously described to genotype the seedlings for the Inv4m haplotype. We harvested and successfully genotyped 364 individual plants from families segregating for Inv4m. The Low, Heterozygote, and High Inv4m haplotypes were segregating in the PT families at a 42:78:45 ratio, within Hardy-Weinberg equilibrium (HWE; D = -2.24, p-value = 0.53). The inversion was segregating in the Mi21 families at a 59:97:43 ratio, also within HWE (D = -0.928392, p-value = 0.7768964).

We randomly sampled 3 biological replicates per experimental replicated from each tissue and temperature treatment for the two homozygous Inv4m genotypes per donor, for a total of 432 samples from 96 plants. Approximately 20 mg of tissue from each sample was placed in a 2ml centrifuge tube, flash-frozen in liquid nitrogen and ground using stainless steel beads in a SPEX Geno/Grinder (Metuchen, NJ, USA). mRNA was extracted using oligo (dT)₂₅ beads (DYNABEADS direct) to isolate polyadenylated mRNA using the double-elution protocol. We prepared randomly primed, strand specific, mRNA-seq libraries using the BRaD-seq [74] protocol with 14 PCR cycles. Samples underwent a single carboxyl bead clean-up, were quantified using the Quant-iT PicoGreen dsDNA kit, and then normalized. We took 2ng per library and multiplexed 96 samples for sequencing. Each multiplexed library was sequenced on 1 lane of a Illumina HiSeq X platform, generating a mean of 4,241,500 reads per sample. Raw reads were quality checked using FastQC v.0.11.5 [75]. Adapter sequences, low quality reads (q<20), and sequences less than 25 bp were removed using Trimmomatic v.0.36 [76].

Effects of genotype at Inv4m on seedling emergence

The effect of Inv4m on seedling emergence were analyzed using the following random slope and intercept model for each donor and temperature treatment separately: (Mod 1)

y_ijk is the emergence time for individual plant k in experimental replicate j in Inv4m genotype i. μ is the model intercept, β₁G_i is the effect of Inv4m genotype, u = u_ijk is a random effects term for the experimental replicate, and e_ijkl is residual error. Variance components, coefficients and standard errors were estimated by REML using the R function lmer [77], and p-values were calculated using conditional F-tests [78].

Population characterization

BC₅S₁ plants are expected to contain ∼ 3% residual DNA inherited from the donor parent across the remaining 9 chromosomes. We used the RNAseq reads from each plant to genotype all residual regions across the genome by calling variants in the expressed regions. Paired reads that passed filtering were aligned to the B73 reference genome version 4 [79] using hisat2 [80], and variant loci were called using GATKv3 [81, 82]. We ran MarkDuplicates, SplitNCigarReads and HaplotypeCaller on every sample, including all 435 samples from our segregating families and an additional 46 B73, 48 PT and 47 B73-PT F1 samples from plants run in parallel with the families but were not otherwise used in this experiment, and then ran GenotypeGVCFs on all samples jointly. We next used SelectVariants to extract SNPs, and VariantFiltration to remove SNPs with FS score < 30 and QD > 2.0. We further filtered for SNPs called homozygous-reference in all B73 samples, and which exhibited allele frequencies > 1/8 and < 7/8 in either the PT or Mi21 families (expected frequencies of each variant should be 0.25 or 0.5 depending on recombination between the two BC₅ parents of each population, but we allowed for some sampling error). We used this set of highly filtered SNPs for each population to genotype each of the individual plants. For each plant, at each locus we first combined all genotype likelihoods across all RNA samples from the same sample. We then identified the approximate breakpoints of the introgressed regions by inspecting the density of variant sites. We identified 3 regions (on chromosomes 2, 4 and 5) in the PT family and 2 regions in the Mi21 family (on chromosomes 3 and 4). Within these introgressed regions, we used R/QTL [83] to assign genotype probabilities across the Inv4m locus for each plant, allowing error.prob = 0.2. Finally, we observed that several genes outside these 5 introgressed regions each of which exhibited >= 2 SNPs relative to the reference. We hypothesized that these genes may have a different chromosome location in the landraces relative to B73, and actually reside inside one of the 5 introgressed regions. We therefore assigned their genotype to the most common genotype among these variant loci.

RNA quantification

To quantify gene expression, we ran kallisto v.0.42.3 [84] separately on each sample using the B73 AGPv4.36 transcript models downloaded from the maize genome database [79]. We limited to only one bootstrap replicate, and then summed the transcript counts for each gene. Genes were retained for analysis when at least a third of the samples had 10 or more reads. Gene counts were normalized using the weighted trimmed mean of M-values (TMM) with the calcNormFactors function in edgeR [85]. Normalization using TMM reduces bias of very highly and lowly expressed genes. The voom function [86] in the limma package [87] was used to convert normalized reads to log2-counts per million (log2CPM), estimate a mean-variance relationship, and assign each observation a weight based on its predicted variance. Observation-weights were used in downstream analyses to account for heteroscedasticity. We estimated batch effects using the removeBatchEffects function in limma using the experimental replicate as batch, which corrected the log2CPM expression values. Global patterns of gene expression across the experiment were visualized with the plotMDS function from edgeR.

Analysis of Inv4m effects on gene expression

We divided genes into three groups to estimate the effects of Inv4m or other introgressed landrace alleles, based on whether each gene resided in a “clean” genomic region with only B73’s allele present in the families, inside the Inv4m locus itself, or if it resided within one of the genomic blocks containing any of the residual landrace genome outside of Inv4m. Each group of genes served a different purpose in the analysis of Inv4m. Genes in the “clean” region were used to assess the effects of Inv4m on global gene expression and indirectly assess the effects on development and physiology more broadly. Genes inside the Inv4m locus were scanned for candidate alleles underlying Inv4m’s effects. Genes in the residual introgression blocks were used as controls to assess the similarity of PT and Mi21 alleles in random genomic loci, as well as compare effect sizes and expression correlations with Inv4m.

For genes that resided in “clean” genomic regions with only B73’s allele present in the segregating families (approximately 89.8% of genes expressed in both donors), we estimated the effect of the Inv4m locus separately in each Inv4m donor, temperature treatment, and tissue using the linear model: (Mod 2) where y_ij is a normalized, batch-corrected log2CPM value for a single gene in a single sample, μ is the intercept for that gene in the particular population, environment and tissue, β is the corresponding effect of the landrace Inv4m haplotype, and e_ij is the model residual, which is assumed to be independent of all other residuals and have variance proportion to ϕ_ij, the empirical weight factor calculated by voom. We fit this model to the whole set of “clean” genes using the lmFit function from limma, and extracted β and its standard error (|β|/t). This provided 18 independent Inv4m effect-size estimates and standard errors for each gene per family. Since our goal was to maximize our power to identify any effect of Inv4m, we combined results across conditions using the mash method [42] implemented in the mashr R package. mash leverages correlations across conditions to more accurately estimate effect size and variance statistics. mashr requires complete data (i.e. each gene requires an effect size estimate and standard error in every condition). Since some genes in particular tissue:temperature conditions were excluded above due to low expression, we imputed their effect sizes as 0 with standard error set to 1000 to convey our lack of knowledge about the true effect size of Inv4m in these conditions. We used the output of mashr to identify a union set of genes regulated by Inv4m by identifying genes with lfsr < 0.05 in any condition.

We also fit a separate model to test for interactions between Inv4m and the temperature environment, separately for each Inv4m donor and tissue: (Mod 3)

This model adds β₂, the main effect of temperature on expression, and β₃, the interaction between Inv4m and temperature. However it is less flexible than the first model because the residual variance σ² is constrained to be equal for the two temperature environments. This model was also fit to each gene using lmFit, and the estimate of β₃ and its standard error were extracted. We again used mashr to identify the union set of genes affected by this interaction.

For genes residing inside the Inv4m locus, we fit the same two statistical models with the lmFit function. However, we did not include these genes in the multiple adaptive shrinkage analysis.

For genes residing outside Inv4m, but within one of the genomic blocks containing residual landrace DNA in both donors, we fit a slightly different statistical model: (Mod 4) where cis is the local genotype of the gene, and β is the associated effect. For genes in residual genomic blocks on chromosome 4, the cis genotypes were highly correlated with the Inv4m, so some of the cis effect may have been caused by Inv4m, but these effects were difficult to separate statistically. However for genes on other chromosomes, the two genotypes were largely uncorrelated.

Sequence divergence and expression correlation between donors within the High Inv4m haplotype and the residual genomic regions

We estimated the sequence divergence between the two landrace donors and B73 for each gene within the chromosome 4 introgression containing Inv4m present in both families. For each gene, we calculated the genetic similarity of the PT and Mi21 alleles relative to B73 by counting the number of shared SNPs divided by total number of observed SNPs within each gene window, using only the highly filtered SNP set described above. We calculated the correlation in expression effect sizes due to Inv4m between PT and Mi21 for each tissue and temperature combination. T-tests were used to determine whether genetic similarity and expression correlation were higher within Inv4m relative to the shared introgressed region.

Gene ontology enrichment

Genes were assigned to gene ontology (GO) categories for functional annotation using an updated ontology annotation [88] which we expanded to include all ancestral terms for each gene with the buildGOmap function of the R package clusterProfiler [89]. Genes in “clean” genomic regions which responded to the High Inv4m haplotype in the same direction in both donor populations were classified as Inv4m-regulated and tested as foreground genes in a GO enrichment analysis. Genes that were expressed in each tissue and temperature combination in both donor populations, but were not Inv4m-regulated were included in the set of background genes. We calculated the enrichment of each GO term using the enricher function in the clusterProfiler R package. We selected GO terms with a false discovery rate less than 1% after a Benjamini-Hochberg multiple test correction. We then calculated the percent of genes in each GO terms that were Inv4m-regulated, and ranked all GO terms by their maximum enrichment across conditions. We then selected the highest enriched GO term among each set of GO terms that had a semantic-similarity >0.5 as the representative term set.

Candidate gene pathway assessment among Inv4m-regulated genes

We inspected two additional candidate gene sets: genes known to regulate flowering time in maize from two separate sources [45, 46], and genes regulated by the microRNA miR172. Inv4m has previously been associated with flowering [33]. miR172 is a highly conserved micro-RNA across the plant kingdom that regulates development and flowering time and one miR172 gene, miR172c is located inside Inv4m. For miR172 targets, we found the mature sequence for zma-miR172c: “AGAAUCUUGAUGAUGCUGCA” from miRBase http://www.mirbase.org, and used this as a query of the Plant Small RNA Target Analysis Server (psRNATarget, http://plantgrn.noble.org/psRNATarget), and collected all predicted target genes. We also used TAPIR’s pre-computed target genes for zma-miR172a-b-c-d. These two categories of genes were inspected by hand for evidence of regulation by Inv4m. From each of the candidate lists, we reported genes which were in the clean genetic background, and differentially expressed in a majority of tissues in the same direction.

Candidate genes within Inv4m

We used the intersection of two separate methods to identify candidate adaptive genes within Inv4m. First, we quantified the proportion of conditions (tissue:temperature combinations) that each gene within Inv4m was differentially expressed according to Inv4m genotype. To be considered differentially expressed, the gene needed to be differentially expressed in both Inv4m donor populations and in the same direction. Genes where at least one donor was not expressed were removed from this analysis per condition. As a complimentary approach, we calculated the correlation in expression between each Inv4m gene and all the genes in “clean” genomic regions that we determined to be Inv4m-regulated above. We then quantified the proportion of these correlations that were significant and in the same direction in both donor populations for each gene. For this analysis, we used the lm function in R to implement a linear model with the Inv4m-regulated gene expression as the response variable, and the Inv4m gene’s expression and Inv4m genotype as predictors.

De-novo assembly of novel genes

We collected all un-mapped reads from the samples homozygous for the high Inv4m haplotype, and used Trinity v2.4 [90] to assemble un-annotated transcripts using default settings. We then used Kallisto to quantify the expression of each of these novel transcripts using the un-mapped reads from each RNAseq sample. To search for candidate “novel” genes in the highland haplotype, we filtered for Trinity genes that had zero estimated counts in any of the samples that were homozygous for the B73 haplotype of Inv4m, but had non-zero estimated counts in at least 2 samples homozygous for the PT or Mi21 haplotype in each segregating donor family (to exclude genes that may reside in either PT or Mi21).