What has population genomics told us about the dynamics of selection and plant adaptation?

Abstract

Population genomics has transformed our understanding of how natural selection shapes plant genomes. The explosion of whole-genome resequencing data has enabled estimates of nucleotide diversity across genomes, clarified how selection interacts with recombination and demography, and broadened inference beyond crops and model species to virtually any plant system. Here, we synthesize key lessons and outstanding challenges concerning the action of natural selection on plant genomes. One message is that purifying selection is a pervasive selective force acting against deleterious mutations and structural variants, but its efficacy varies predictably and is often relaxed after demographic contractions, including domestication bottlenecks, range expansions, shifts in mating systems, and polyploid formation. Adaptation also routinely shapes genetic diversity, as evidenced by the detection of selective sweeps. Swept genes have functional biases, but many outstanding questions about adaptive variation remain, including the selection coefficients of adaptive alleles, whether selection typically acts on segregating variants or de novo mutations, and how often selection is polygenic. As with polygenic selection, our knowledge about balancing selection is limited, due in part to challenges in its detection. It is known, however, to act on disease-resistance genes and genes that govern breeding systems. Finally, an emerging theme in plant population genomics is the interplay between introgression and adaptation, but it remains challenging to link introgression confidently to fitness, and further study requires an expanded, multispecies scale. This review reveals the power, and some of the limits, of population genomics to infer the dynamics of selection on plant genomes.

Introduction

Many major questions in evolutionary biology focus on understanding the prevalence, strength, dynamics, and genetic targets of natural selection. To bolster this understanding, substantial effort has been dedicated to detecting and quantifying selection at the DNA level. These efforts began with Kreitman's 1983 report of nucleotide polymorphism in the Drosophila melanogaster Adh gene, in which he documented the types and frequencies of sequence-level variants from a sample of 11 flies (Kreitman 1983). Although modest by today's standards, Kreitman's research spurred a wave of theoretical and statistical methods to detect signals of selection in nucleotide polymorphism data (Hudson et al. 1987; Tajima 1989).

The first molecular population-genetic analyses of nuclear DNA sequences from plants were not published until a full decade later (Gaut and Clegg 1993a, 1993b). Early efforts to measure genetic diversity in plants tended to focus on crop species, both because they had genetic information available and because it was reasonable to assume that they had experienced widespread and recent selection during and after domestication (Eyre-Walker et al. 1998). Even so, by the early 2000s—fully two decades after Kreitman's work—only a handful of studies had investigated nucleotide polymorphism in plant nuclear genes (Fig. 1). At that time, two of us wrote a comprehensive review about detecting adaptive evolution from plant nucleotide data (Wright and Gaut 2005). It was easy to be comprehensive because DNA sequence variation had been sampled from only a few plant species for a few dozen genes. One conclusion from the review was that, due to sampling biases and limitations in controlling for demographic factors, it was premature to assess the general extent and modes of selection acting across plant genomes.

Figure 1.

A histogram of the number of papers per year (y-axis, left) and the number of citations per year (y-axis, right) to papers that met Web of Science search parameters, as accessed on October 10, 2025. The search parameters are provided in the Supplementary Methods. The arrows point to illustrative papers either for conceptual advances (e.g. association mapping, demographic inference, or temporal sampling) or for marked increases in the scale and scope of data. The footnotes refer to 1. (Gaut and Clegg 1993a, 1993b), 2. (Eyre-Walker et al. 1998), 3. (Thornsberry et al. 2001), 4. (Nordborg et al. 2005; Wright et al. 2005), 5. (Coop et al. 2010; Eckert et al. 2010), 6. (Huang et al. 2012; Hufford et al. 2012), 7. (1001 Genomes Consortium 2016), 8. (Kreiner et al. 2022a), and 9. (Landis et al. 2024).

Twenty years later, it is interesting to reflect on how much the field has grown. Over the last several years, >500 plant studies have been published annually that contain the phrases “nucleotide polymorphism” or “selective sweeps,” with 5,000 such papers since 1993 garnering over 20,000 citations per year (Fig. 1). The field has grown, in part, because plants have distinctive features that make them particularly useful for addressing fundamental questions in evolutionary genetics. One feature is variation in mating systems. Flowering plants span a continuum from obligate outcrossing to complete self-fertilization, often with varying degrees of asexual reproduction and mixed strategies. In theory, mating systems affect effective population sizes, population differentiation, the efficacy of selection, and the extent of linkage disequilibrium (Charlesworth and Wright 2001); plants provide a comparative framework for testing these expected effects. Plants are also sessile, meaning they must tolerate or adapt to their local environment, producing genomic signatures that can be associated with geographic and environmental gradients. Their breadth of ecological and life-history traits—from ephemeral annuals to ancient perennials, from generalists to narrow endemics—provides additional useful contrasts. Of course, plants also exhibit a high propensity for polyploidization and associated genome rearrangement, providing natural experiments in genome duplication and its consequences (Wendel 2000). Finally, plant population genetics has been central for studying the evolutionary processes associated with domestication.

In many recent papers, short-read whole-genome resequencing (WGR) data have been generated from scores of, and often hundreds, of individuals per species. Moreover, research is no longer limited to crops and model species, due to the relative ease of generating long-read, high-quality reference genomes. With sufficient funding, it is now technically straightforward to generate large population genomic datasets for virtually any plant species or group; often, the more difficult challenge is collecting samples and verifying their provenance. As a consequence, the field of plant population genomics is replete with WGR data. These data have facilitated estimation of nucleotide sequence diversity across numerous plant taxa, revealed how diversity is distributed across genomes, and aided in the association of genotypes with phenotypes.

Here, we summarize lessons from the current literature about what we have learned about the action of selection on plant genomes. The empirical literature is so extensive (Fig. 1) that we cannot hope to perform a comprehensive survey as we did 20 years ago. For that reason, we avoid well-trodden topics that have been reviewed recently, such as the genomic effects of polyploidy (Otto and Whitton 2000; Mason and Wendel 2020), and we also tend to focus on flowering plants, reflecting existing biases in the literature. Furthermore, the literature is so extensive that we must rely on specific examples, rather than a comprehensive literature summary. However, to assess general trends in the literature, we have also amassed a dataset of 267 papers (Table S1). As detailed in the Supplementary Text, these papers were based on a focused search for WGR-based plant population genomics, while also requiring that the primary text is in the public domain. This dataset is not comprehensive and may reflect unintended biases, but we use it to extract information to help evaluate general trends in the literature (Table 1).

Table 1.

Search term(s)	%Papers with termsa
Introgression	46.4
Gene ontology; GO enrichment	46.1
GWAS	45.3
Selective sweep	39.0
Positive selection	30.3
Purifying selection	24.7
Genetic load	9.7
Balancing selection	8.6
Deleterious variants	7.9
Linked selection	7.9
Adaptive introgression	7.1
Polygenic adaptation	4.1
Selection coefficient	3.0
Diversifying selection	2.6
Ancestral recombination graph	1.9
Hard sweep	1.5
Soft sweep	1.5
Environmental association	1.1
Tree sequence	1.1

^aPercentages reflect the number of papers, out of the 267 paper exemplar set, that contained the specific search term or one of the set of search terms, based on exact grep searches of Markdown files with references removed (see Supplementary Text).

So, what has population genetics told us about selection and adaptation in plant genomes? To address this question, we have organized this review into separate sections that focus on distinct modes of selection. The sections cover purifying selection, adaptive selection and selective sweeps, polygenic adaptation, balancing selection, and, finally, adaptive introgression. In each section, we briefly touch on theoretical and methodological considerations, illustrate major empirical findings, and conclude by listing open questions for current and future research. After surveying these distinct modes of selection, we conclude with a call to action that includes renewed community standards, methodological development, especially related to multispecies datasets, and a sense that while much has been learned in the last two decades, there are still many open and compelling questions to be addressed.

Purifying selection

Purifying selection is the most pervasive and dominant mode of selection. By filtering deleterious mutations from populations, purifying selection reduces genetic diversity at sites under selection and also at linked neutral sites (Charlesworth et al. 1993). One way to measure purifying selection is to compare nucleotide diversity per site (π) at synonymous (π_s) and nonsynonymous (π_n) sites. When purifying selection acts on amino-acid-changing variants and when selection on synonymous sites is minimal or absent, π_n/π_s is expected to be less than 1.0. This is, in fact, precisely what Kreitman's work showed (Kreitman 1983). It is also worth remembering that, in the absence of selection, π provides insights into historical population sizes, because it estimates the product of effective population size (N_e) and mutation rate (µ) (that is π = 4N_eµ) (Nei and Li 1979). This is important because the efficacy of selection scales inversely with N_e; thus, purifying selection is expected to be stronger in species with higher effective population size.

But how diverse are plants, and is empirical data consistent with purifying selection scaling with N_e? Chen et al. (2017) addressed these questions by measuring π_s and π_n across thousands of coding regions in plant and animal species (Chen et al. 2017). They found that π_s varied ∼11-fold among 28 plant species, with significantly lower diversity in self-fertilizing plants than in outcrossing species. As a group, plants had higher median π_s and a much narrower range than the 34 animals in their sample. They also measured the strength of purifying selection in each of their species using π_n/π_s. As predicted, there was a negative relationship between π_n/π_s and π_s, indicating stronger purifying selection in species with higher N_e. Interestingly, the slope of the negative relationship differed between plants and animals. The difference in slope disappeared when accounting for potentially different µ across species, but it nonetheless suggests, as others have concluded (Hough et al. 2013), that the strength and dynamics of purifying selection differ between plants and animals. For example, fruit flies have pervasive purifying selection across all regions of the genome, including noncoding regions (Sella et al. 2009). In contrast, purifying selection in plants is more often focused on genes, with a smaller fraction of noncoding sites subjected to purifying selection (Hough et al. 2013).

The metric π_n/π_s is commonly used to quantify purifying selection, but other methods utilize the site frequency spectrum (SFS)—i.e. the frequency count of variants observed in a sample. Nonsynonymous SNPs tend to be rarer than synonymous variants and thus their SFS is more left-skewed; similar results apply to structural variants (SVs) (Zhou et al. 2019; Kou et al. 2020; Hämälä et al. 2021) (Fig. 2a). This left-skewed pattern is consistent with purifying selection restricting nonsynonymous SNPs and SVs to low population frequencies. The SFS can also be employed to estimate the distribution of fitness effects (DFEs), which provides a population-scale estimate of N_es, the product of N_e and the selection coefficient s, for purifying selection on new mutations (Fig. 2b). Once the DFE is estimated, N_es values can be compared across taxa with different mating systems (Arunkumar et al. 2015; Muyle et al. 2021), contrasting demographic histories (Jaramillo-Correa et al. 2020) or both (Laenen et al. 2018). The DFE can also be used to investigate the factors that affect purifying selection. For example, Chen et al. (2020) estimated the DFE for 11 plant species. For species with low N_e, polymorphism patterns were dominated by genetic drift. However, linked selection, which affects the π_n/π_s ratio in part through disproportionate effects on π_s, was required to explain the observed DFEs for higher N_e species.

Figure 2.

a) An unfolded SFS based on population genomic data from the wild relative of domesticated grapevine. The SFS illustrates differences among synonymous SNPs (sSNPs), nonsynonymous SNPs (nSNPS), and SVs. The latter two have pronounced left-leaning distributions that are consistent with strong purifying selection against deleterious alleles. b) Inferred DFEs based on a demographic model that used the SFS for sSNPs as the neutral control. The bins to the far left, where Nes < −100, represent the proportion of new mutations that are strongly deleterious and illustrate that most new nSNP and SV mutations are subject to strong purifying selection. The DFE was estimated using fitDadi (Kim et al. 2017). The data came from Figure 2 of Zhou et al. (2019) with the various types of SVs combined into a single category.

When is purifying selection relaxed?

At least four commonly studied situations reduce the population- or species-wide magnitude of purifying selection. The common thread among these situations is a historical decrease in N_e, which increases genetic drift, reduces the efficacy of selection, and can lead to the accumulation of slightly deleterious variants (Lohmueller 2014; Bertorelle et al. 2022). The accumulation of deleterious variants, which increases the mutation load, has been measured empirically by metrics related to π_n/π_s (e.g. Willi et al. 2018; Exposito-Alonso et al. 2018b; Fiscus et al. 2025; Jiang et al. 2025), by estimating the DFE, and also by counting the number of derived deleterious mutations within a genome, a measure that may be less sensitive to the confounding effects of population history (Simons et al. 2014) (reviewed in Moyers et al. 2018).

One situation that can weaken purifying selection is domestication. Domesticated species tend to have deleterious mutations at higher population frequencies than their wild ancestors, likely owing to domestication bottlenecks, linkage between deleterious and adaptive variants and perhaps relaxed selection due to cultivation (Gaut et al. 2018). Increased mutational loads have been documented in domesticated animals (Marsden et al. 2016; Xie et al. 2018), rice (Lu et al. 2006; Liu et al. 2017), sunflower (Renaut and Rieseberg 2015), barley (Kono et al. 2019), maize (Lozano et al. 2021), and grapes (Zhou et al. 2017), but there are exceptions like sorghum (Lozano et al. 2021). This phenomenon, which has been labeled the “cost of domestication” (Lu et al. 2006; Moyers et al. 2018), has led to the idea that one way to improve crops is to breed or engineer for reduced mutational load (Morrell et al. 2011). A second situation involves populations at the niche edge of a wild species' geographic range. Edge populations experience environmental marginality, which can, but not always (Moeller et al. 2011), lead to lower population densities, reduced N_e, and the accumulation of deleterious variants. For species undergoing range shifts, the effect may differ between the expanding and lagging edge of the range (Angert et al. 2020), because populations at expanding edges are especially prone to experience low N_e associated with serial founder events, assortative mating, and strong environmental selection pressure (Hampe and Petit 2005). Once established in expanding-edge populations, deleterious mutations can further propagate via allele surfing (Travis et al. 2007; Excoffier et al. 2009). Consistent with these predictions, leading-edge populations have increased numbers and frequencies of deleterious variants and elevated π_n/π_s ratios (González-Martínez et al. 2017; Willi et al. 2018; Takou et al. 2021; Fiscus et al. 2025; Jiang et al. 2025; Ning et al. 2026), with measurable impacts on plant fitness (Perrier et al. 2020; Leventhal et al. 2025).

Mating-system transitions, specifically a shift from outcrossing to selfing, also prompt a reduction in N_e (Peng et al. 2025). Substantial empirical evidence documents accompanying shifts in purifying selection—e.g. in selfing Eichhornia species (Ness et al. 2012; Arunkumar et al. 2015), Capsella rubella (Slotte et al. 2013), and Arabidopsis thaliana (Payne and Alvarez-Ponce 2018), among others. It is tempting to conclude that weak purifying selection is a universal feature of a shift to selfing, but the effects of the shift depend on evolutionary parameters like Ne, dominance, the DFE, and recombination rate (Nordborg 2000; Sianta et al. 2023). With moderate to high recombination, for example, greater selfing may increase the efficacy of purifying selection against strongly deleterious, recessive alleles, thus enabling genetic purging (Schemske and Lande 1985). One interesting feature of mating-system shifts, particularly elevated selfing and inbreeding, is that they may be more common at range edges, further exacerbating N_e effects associated with environmental marginality (Zeitler et al. 2023).

A polyploid event can also alter the strength of purifying selection but in both directions. On one hand, the increased number of gene copies raises N_e, which can increase the efficacy of purifying selection (Otto and Whitton 2000; Hollister et al. 2012). On the other hand, polyploid formation is often accompanied by a genetic bottleneck, which reduces N_e and so can weaken purifying selection (Soltis et al. 2014; Monnahan et al. 2019). Purifying selection can be further weakened by genetic redundancy: with duplicate gene copies present, one or even both copies may be released from selective constraint. Additionally, if the two parental species of a polyploid carried different deleterious burdens, that legacy can persist in the resulting polyploid (Kryvokhyzha et al. 2019). We do not focus on polyploids in this review, as they introduce exceptions beyond our current scope; nevertheless, whole-genome duplication clearly affects the genome-wide strength and efficacy of purifying selection (Conover and Wendel 2022).

What is left to learn about purifying selection?

We have learned a great deal about purifying selection in plants. We know it is strong against nonsynonymous variants and SVs (Fig. 2). We know it is relaxed across a range of situations, varies across genes (Slotte et al. 2011; Yang and Gaut 2011; Williamson et al. 2014), and correlates strongly with gene expression (Josephs et al. 2015). Yet, there is much more to learn. We are still discovering nuanced effects of purifying selection on, for example, synonymous variants that alter codon usage bias (Parvathy et al. 2022) or that affect mRNA secondary structures (Martin et al. 2025 ). These observations, along with the presence of linked selection, imply that synonymous SNPs are rarely strictly neutral, which affects our inference of metrics like the DFE that assume their neutrality (Martinez i Zurita et al. 2025).

Linked selection generally reduces diversity at neutral sites, but linkage can also generate unexpected consequences. For example, heterozygous, putatively recessive deleterious SNPs in maize are preferentially purged during selfing (Roessler et al. 2019), but low-recombination regions retain higher-than-expected heterozygosity, likely due to associative overdominance caused by recessive deleterious alleles on different haplotypes (McMullen et al. 2009; Roessler et al. 2019). There is growing recognition that associative overdominance can have genome-wide effects on patterns of neutral diversity, further complicating inferences of selection (Zhao and Charlesworth 2016; Gilbert et al. 2020).

Metrics related to π_n/π_s have been—and will continue to be—useful tools for studying purifying selection, but this line of inquiry needs to expand. One emerging approach is to update π_n/π_s metrics by incorporating increasingly powerful information about global protein structures based on deep learning analyses (Cheng et al. 2023; Norn et al. 2024). DFE estimation merits additional emphasis; only a handful of publications have inferred DFEs among plant populations and species. One found that DFEs are relatively stable within species but vary significantly between species (James et al. 2023). This pattern implies that DFEs are shaped more strongly by deep, species-level events rather than recent evolutionary events like local adaptation or population-level demographic shifts. But is this a general pattern? More generally, how much does selection on new mutations—and for that matter, dominance coefficients (Di and Lohmueller 2024)—vary among species?

Another major question concerns the extent and proportion of noncoding genomic regions subject to purifying selection. Studies suggest that these regions can be substantial; in A. thaliana, ∼4% of the genome is noncoding DNA subjected to purifying selection (Haudry et al. 2013). Further examination of the fraction of classes of noncoding regions in Capsella grandiflora revealed that over 50% of untranslated regions and over 70% of identified conserved noncoding sequences are subject to purifying selection (Williamson et al. 2014). Similarly, studies of 3D genome structure have shown that boundary regions of topological domains have low nucleotide diversity in both plants (Liao et al. 2022) and animals (Liao et al. 2021), suggesting these regions may also be under purifying selection. The synthesis of 3D structure and chromatin markers may yield additional insights into the dynamics of selection. However, documenting purifying selection in noncoding regions remains challenging because it requires sequence conservation across species or individuals. Identifying these conserved regions is challenging between difficult-to-align genomes, but new methods using “DNA language models” may prove helpful (Benegas et al. 2023; Morrell and Pakhomov 2025; Zhai et al. 2025a, 2025b). Yet, challenges remain for large language models (LLMs), including reference biases, polarizing ancestral and derived variants (Keightley and Jackson 2018), and the potential for transposable elements (TEs) to overwhelm the models. Current approaches have avoided some issues by focusing on sequences immediately adjacent to genes (Benegas et al. 2023; Zhai et al. 2025b), but eventually noncoding elements will need to be discerned more widely, perhaps by masking or downweighting the effect of heavily represented TEs.

Adaptive evolution: selective sweeps

Hard and soft sweeps

Decades of experimental work have demonstrated that plants adapt readily to their environments, as illustrated by reciprocal transplant experiments (Clausen et al. 1940). But what should a history of adaptation look like in population genomic data? A foundational expectation is that adaptation produces selective sweeps, but sweeps do not conform to a single model. The classical “hard” sweep model assumes that adaptation proceeds from a strongly advantageous de novo mutation that rises in frequency under directional selection and ultimately approaches fixation (Smith and Haigh 2007) (Fig. 3a). In contrast, “soft” sweeps involve directional selection acting on multiple haplotypes simultaneously (Fig. 3b). Multiple beneficial haplotypes can arise when variants segregate at appreciable frequencies before the onset of positive selection or when a recurrent mutation is introduced onto distinct genetic backgrounds (Hermisson and Pennings 2005; Pennings and Hermisson 2006; Harris et al. 2018). The magnitude of diversity reduction for soft sweeps is related to numerous evolutionary parameters, but diversity reductions are expected to be less pronounced than under hard selective sweeps (Fig. 3b).

Figure 3.

Patterns of polymorphism under various sweep models: a) The hard sweep model, in which a de novo mutation, denoted by the bold T, arises on a single haplotype and then rises to fixation under directional selection. The orange color reflects the selected haplotype across a linked region. Over time, recombination with other haplotypes shortens this haplotype in some individuals, as denoted by black letters on the orange background. Nonetheless, the outcome is notably reduced nucleotide diversity in the linked region around the beneficial variant. b) The soft sweep model, based on an adaptive variation (bolded T) arising on two haplotypes (orange and blue). Directional selection increases the frequency of both haplotypes, so that the reduction in diversity around the selected site is less severe than a hard sweep. c) An example illustrating how hard sweeps in two distinct populations can mimic a soft sweep when data from the two populations are analyzed together. This figure was modeled after Figure 1 in Novembre and Han (2012).

Why is it helpful to conceptualize soft versus hard sweeps? One reason is to better understand the dynamics of selection. Does selection typically lead to a complete dearth of genetic variation, as expected for hard sweeps? A related question is: Does selection typically rely on de novo mutations or act on segregating variants? This matters because it informs whether adaptation is mutation-limited or whether the strength of selection is the main determinant of adaptive success. There have been extensive debates about the prevalence of hard versus soft sweeps in Drosophila. Some groups argue that soft sweeps are more common than hard sweeps (Garud et al. 2015; Garud and Petrov 2016), while others find little evidence for soft sweeps (Jensen 2014; Harris et al. 2018 but see Feder et al. 2021). The debate hinges primarily on whether segregating variants are common prior to the onset of selection, which in turn depends on whether variants are conditionally neutral or antagonistically pleiotropic. A second reason relates to the power to detect selective sweeps. Numerous studies have used genome scans to identify putatively selected genes in plants. Yet published studies have almost exclusively used methods designed to identify hard sweeps, leaving soft sweeps mostly ignored. One example illustrates the point. In an important 2012 study (Fig. 1), Huang et al. identified selected genes in rice by ranking genes by the difference in π between wild and cultivated samples (Huang et al. 2012). This ratio is expected to be elevated for genes that have experienced selective sweeps associated with crop domestication or improvement, but it is likely insensitive to soft sweeps. If soft sweeps are common, their genome scan probably missed numerous genes of interest.

Methods to detect soft sweeps based on haplotype diversity have been used surprisingly rarely in the plant population genomics literature; fewer than 2% of our 267 curated papers even mention the term “soft sweep” (Table 1). Fewer still have utilized any of the slew of methods that may help to detect them (Sabeti et al. 2002, 2007; Voight et al. 2006; Garud et al. 2015). Machine learning (ML) approaches have recently been developed to detect sweeps and classify them as hard versus soft (e.g. Garud et al. 2015; Schrider and Kern 2016), but they have yet to be applied widely to plants. Perhaps the most compelling evidence for soft sweeps is identifying independent causal mutations (Pennings et al. 2014; Feder et al. 2016; Kreiner et al. 2022b), thereby directly demonstrating the presence of multiple adaptive alleles.

How prevalent are sweeps?

Given numerous studies to detect selection from plant population genomic data (Fig. 1, Table 1, and Table S1), what have we learned about the prevalence of sweeps in plant genomes? Do many genes bear the signatures of sweeps or just a few? One might assume that these questions have been answered thoroughly, but we still have at best incomplete insights, for two reasons. One is that the parameters for recognizing even hard sweeps are relatively narrow, making them difficult to detect (Braverman et al. 1995; Przeworski 2002, 2003); soft sweeps are even more difficult. There are also methodological limitations. For example, many (and perhaps most) papers use an a priori empirical cutoff for test statistics, in that scans identify candidate selected genes (or windows) as having the most extreme 5%, 1%, or 0.01% of the test statistic of interest (e.g. the composite likelihood ratio statistic; Nielsen et al. 2005). This approach predefines the number of genes that contain a signature of selection, which is reasonable for producing a list of candidate selected genes (Teshima et al. 2006) but not for estimating the proportion of genes that bear a signature of selection.

Surprisingly few studies have been designed to explicitly estimate the proportion of genes under selection. Notable exceptions include the first large-scale studies of nucleotide diversity in maize (Zea mays ssp. mays), which estimated that ∼2% of genes bore the signature of selection (Wright et al. 2005; Hufford et al. 2012). Taken at face value, this number suggests that ∼800 genes in maize exhibit the signature of adaptive selection, but we do not know whether this number is accurate or typical. [In contrast, an early study of A. thaliana in northern Sweden populations detected 22 swept regions (Long et al. 2013)]. We also do not know whether domesticates, such as maize, have more genes under selection than wild plants, given recent conscious and unconscious selection on domestication traits (Ross-Ibarra et al. 2007). And of course, recombination will affect the proportion of the genome affected by selective sweeps due to linkage.

How might we achieve better estimates of the number and genomic extent of hard sweeps? One solution is to simulate data under the neutral demographic model of the species or population in question, analyze the simulated data using the desired sweep metric, and thereby set a simulation-based cutoff for empirical data. This approach moves away from an a priori empirical cutoff and will provide a better estimate of the number of regions that are more extreme than expected under a given demographic model. Another approach is to use ML methods to estimate the proportion of the genome attributable to hard sweeps. One study of grapevine used ML methods to estimate that <1% of genes exhibited diversity patterns consistent with a hard sweep (Xiao et al. 2023). Because ML models are often trained on simulated data, both approaches rely on simulations to mimic population-genetic history. If the simulations are incorrect, they are likely to lead to Type I and/or Type II errors. Sampling poses a further challenge: most plant datasets span broad geographic ranges, whereas local sweeps require in-depth within-population sampling.

Another genome-wide approach that provides insight into the prevalence of sweeps comes from Sella et al. (2009), who demonstrated a genome-wide reduction in nucleotide diversity surrounding nonsynonymous fixations in Drosophila. This pattern suggests that a significant fraction of nonsynonymous fixations have undergone selective sweeps. This approach has not yet been applied broadly to plant population genomic datasets, but outcrossing C. grandiflora exhibited a pattern similar to Drosophila (Williamson et al. 2014).

Are hard or soft sweeps more common?

Just as it is difficult to estimate the proportion of genes that have experienced a sweep, it is as difficult to compare the prevalence of soft versus hard sweeps. This is partly because, as we have already mentioned, methods for detecting soft sweeps have been largely ignored in empirical plant population genetics. Evidence is accumulating, however, that soft sweeps could be an important feature of plant adaptation. For example, crop improvement in Eucalyptus grandis has relied predominantly on selection on segregating variants, leading to the signature of soft sweeps (Mostert-O’Neill et al. 2022). Another paper—one of the few to classify soft versus hard sweeps using ML methods—found that ∼12% of the genome contained signals of soft sweeps, with another 21% linked to soft sweeps and only 1% categorized as hard sweeps (Xiao et al. 2023). The high estimated proportion of soft sweeps does lead one to pause: Why do soft sweeps appear to be so much more prevalent than hard sweeps? One reason could be that classic hard sweeps are uncommon, as has been argued for humans (Hernandez et al. 2011). Another potential explanation is unrecognized population structure in a global sample. If a sample comprises multiple structured populations, independent hard sweeps within populations appear as a soft sweep across the global sample (Messer and Petrov 2013) (Fig. 3c). Incomplete (or partial) hard sweeps also mimic some features of soft sweeps, which can lead to further misclassification.

One way to assess the potential prevalence of soft sweeps is by identifying whether adaptive variants segregate prior to selection. A small but growing body of literature has used domestication as a model to determine whether selection has acted on de novo mutations or on standing variation. One compelling example is the teosinte glume architecture 1 (tga1) locus of maize (Fairbanks and Ross-Ibarra 2025). This domestication locus contains a mutation that alters the fruitcase surrounding the kernel, enhancing edibility. The tga1 mutation was thought to have arisen de novo during domestication, given its apparent absence in wild populations. Leveraging a larger dataset from wild populations than previously available, Fairbanks and Ross-Ibarra (2025) identified the mutation at a low but detectable ∼1% frequency in Z. mays subsp. mexicana, a wild taxon that contributed to modern maize (Yang et al. 2023). They then used ancestral recombination graphs (ARGs) to estimate the mutation's origin at ∼45,000 years before present, a time frame that vastly exceeds the onset of domestication ∼10,000 years ago. Hence, the tga1 mutation preexisted in wild populations long before artificial selection. Notably, this appears to be true for every domestication gene identified in maize thus far (Fairbanks and Ross-Ibarra 2025). Rice is similar: Alam et al. (2025) used ARGs to estimate the age of putatively causative mutations that contribute to domestication phenotypes. They estimated that 73% of mutations had allele frequencies >5% in wild populations, with the majority originating before domestication. Their ARG-estimates complement previous work showing that domestication-related haplotypes segregate in wild rice populations across a wide geographic range (Jing et al. 2023). The evidence for nontrivial allele frequencies in wild rice and maize suggests that soft sweeps are a common feature of domestication. Additional evidence for the prevalence of soft selective sweeps comes from the multiple independent herbicide resistance alleles in Amaranthus tuberculatus (Kreiner et al. 2018). ARG reconstruction also revealed multiple origins of the same resistance mutation across A. tuberculatus populations (Kreiner et al. 2022b), highlighting the potential for soft sweeps under intense herbicide pressure in a plant with a large N_e.

What types of genes are prone to positive selection?

A primary motivation for detecting selective sweeps is to identify the genetic mechanisms and functions that contribute to adaptation. Gene function is usually inferred from gene ontology (GO) enrichment analyses, which are surprisingly common given that GO enrichment was mentioned in ∼50% of the papers in our dataset, eclipsing even “selective sweeps” (Table 1). Any discussion of GO enrichment results requires caveats, however. One major caveat is reporting biases, because authors typically discuss only a subset of results that fit a narrative—e.g. enrichment for agronomic trait genes during domestication. GO analyses also typically include all genes within a sweep, even though many may be linked to, rather than a direct target of, selection. Moreover, the type of genes under selection can vary between species-wide sweeps and local adaptation, and so the scale of analysis matters. Finally, many of the inferred functions are biased by functional discovery in A. thaliana. In short, GO analyses are, at best, a blunt tool.

What have they revealed about functional trends in plant adaptation? The literature is, frankly, overwhelming and well beyond our current scope. As a proxy, we examined GO terms extracted from the curated dataset of 267 papers, representing species from >70 genera (Table S1). Although it is difficult to generalize, some functional themes emerged. For example, numerous empirical studies suggest that natural selection frequently targets genes involved in responses to biotic stresses. One recent comparative analysis identified selective sweeps across 17 diverse plant species, showing that genes associated with disease-resistance were disproportionately represented within sweeps (Nocchi et al. 2024). The authors proposed that these genes may be especially responsive to selection because they often exhibit large fitness effects, relatively low pleiotropic constraint, and membership in gene families (Nocchi et al. 2024). Disease-defense genes are also enriched within putative selective sweep regions for wild Vitis species (Morales-Cruz et al. 2021), Populus cathayana (Xiang et al. 2024), Medicago trunculata (Bonhomme et al. 2015), and rice (Alam et al. 2025), to name just a few examples. A particularly interesting example comes from a genome scan of invasive ragweed (Ambrosia artemisiifolia), where selection on disease-resistance and defense-response genes suggests pathogen-mediated adaptation during range expansion (Bieker et al. 2022). The repeated identification of defense-related functions suggests host–pathogen interactions commonly shape the dynamics of plant genome evolution, as they have in humans (Enard et al. 2016; Marini et al. 2026).

Genes contributing to flowering time, photoperiodic responses, and circadian regulation are also frequently identified in selection scans of wild and domesticated species. These genes mediate adaptation to clinal variations in day length and to other environmental factors such as temperature. Examples include the FLOWERING LOCUS C (FLC), a major repressor of flowering in A. thaliana (Deng et al. 2011); FRIGIDA (FRI), which contributes to the vernalization response and acts epistatically with FLC to shape flowering phenotypes across environments (Caicedo et al. 2004); FLOWERING LOCUS T (FT), which promotes flowering in response to long photoperiods (Kardailsky et al. 1999; Kobayashi et al. 1999); TERMINAL FLOWER 1 (TFL1), which is functionally antagonistic to FT (Bradley et al. 1997); and ELF3 (EARLY FLOWERING 3), which enables flowering in short growing seasons (Faure et al. 2012). Most of these genes bear evidence of positive selection across distinct species, indicating they are frequent selection targets. For example, loci homologous to FT have evidence for positive selection in A. thaliana (Flowers et al. 2009), rice (Han et al. 2023), maize (Guo et al. 2018), cucumber (Wang et al. 2020), and soybean (Dong et al. 2022). The genes mentioned here represent a subset of larger developmental networks that are targeted by selection and contribute to photoperiod sensing, vernalization pathways, and circadian clock regulation (reviewed in Blackman 2017).

A third functional category emerged from our survey (Table S1): genes associated with fruit traits. GO categories that included keywords such as sugar metabolism, fruit development, ripening, flavor, and aroma were especially common in domesticated horticultural crops, including peach, pear, apricot, grapevine, and strawberry. These observations are reasonable, but like all GO-based analyses it can be difficult to separate narrative from biological relevance. We are not immune to this narrative bias; the most commonly mentioned set of GO enrichment terms across our 267 papers were related to “specific enzyme activity” (Table S1). Because this is a diffuse and amorphous category, it is difficult to construct a useful narrative.

In addition to GO biases, genome scans are always limited by concerns about confounding factors, imperfect model specification and false positives (Jensen 2014). They do, however, generate testable hypotheses that are increasingly amenable to functional analysis and verification. One striking example is the transition to polyploidy in Arabidopsis arenosa. Initial population genomic scans revealed a strong enrichment of meiosis genes in genome-wide scans of differentiation between diploid and polyploid populations (Hollister et al. 2012; Yant et al. 2013; Wright et al. 2015). These early results helped fuel additional functional characterization of the evolution of recombination rates and pathways in polyploid populations (Morgan et al. 2020, 2021), which have largely validated initial inferences about the importance of meiotic adaptation in nascent polyploids.

What is left to learn about adaptive selection and beneficial mutations?

We have already outlined some topics that need further attention—e.g. the consideration of soft sweeps, further improvement of model specification, and efforts toward functional hypothesis testing—but there are more. One surprising omission is estimates of the strength of selection, s. Only ∼3% of the papers in our exemplar set used the term “selection coefficient,” far fewer have actually estimated s for individual adaptive alleles (Table 1 and Table S1). A notable exception is the work of Kreiner et al. (2022b), who used temporal allele-frequency data from the outcrossing weed A. tuberculatus (Kreiner et al. 2022b) to estimate that s was 0.19 for seven herbicide resistance alleles, implying a 19% advantage for individuals with these variants in herbicide-laden environments. As this example illustrates, knowledge of s is critical, because it (i) dictates the timing and speed of allele-frequency change, (ii) shapes the effect of linked selection, (iii) affects the power to detect selective sweeps, and (iv) determines whether selection is likely to overcome genetic drift (i.e. |N_es| >> 1).

The question is how to fill this obvious gap in the field. One way to estimate s is to extend DFE methods fit to SFS data by incorporating beneficial and deleterious mutations (Schneider et al. 2011; Tataru et al. 2017). However, this approach underestimates the genome-wide strength of positive selection, because strongly beneficial mutations spread rapidly and contribute little to the SFS (Booker 2020). Methods that also integrate genome-wide patterns of selective sweeps may permit less biased inference of both the rate and the distribution of beneficial selection coefficients (Uricchio et al. 2019). The use of ARGs to estimate s is growing and likely to be the basis for studying selection coefficients moving forward (e.g. Brandt et al. 2024; Nielsen et al. 2025). It is also possible to estimate the strength of selection by following temporal allele-frequency change, but this approach is limited either to taxa with historical samples, such as herbarium specimens (e.g. Exposito-Alonso et al. 2018a; Bieker et al. 2022; Lang et al. 2024; Eckert et al. 2026), seed banks (Gómez et al. 2018), or long time-course experiments (Landis et al. 2024).

To gain further insights into the strength and prevalence of adaptive evolution, one can turn to alternatives that do not focus on selective sweeps per se. One option is the metric α, which estimates the proportion of nonsynonymous substitutions fixed by natural selection (Smith and Eyre-Walker 2002; Galtier 2016). Estimates of α are appreciable; they exceed 30% in outcrossing C. grandiflora (Slotte et al. 2010; Williamson et al. 2014), between Populus species (Lin et al. 2018), and between Helianthus species (Strasburg et al. 2011). A useful property of α is that it can be compared across species (Fay 2011), generally yielding lower estimates for selfers compared to outcrossers (Moutinho et al. 2020; Zeng et al. 2024). It can also be applied to different categories of sites, not just nonsynonymous variants. For example, α values for intronic and noncoding sites exceed 15% in Capsella, Populus, and Norway spruce (Williamson et al. 2014; Lin et al. 2018; Wang and Ingvarsson 2023).

Two emerging challenges include (i) the assessment of the contribution of SVs to adaptation and (ii) functional validation of candidate adaptive variants. The contribution of SVs can be substantive; one study of an invasive species inferred that 26% of parallel (and presumably adaptive) events involved large chromosomal inversions (Battlay et al. 2023). SVs also explain more genetic variation than SNPs for some phenotypic traits (e.g. Liu et al. 2024), implying a potential role in adaptation. The study of SVs will certainly improve as more population genomic datasets are generated from samples of complete, phased genome assemblies (Alonge et al. 2020; Igolkina et al. 2024), but it will remain difficult to disentangle causative SVs from linkage effects. For the second challenge (functional validation), CRISPR-engineering has expanded the possibilities to test effects of candidate selection genes. However, CRISPR and other engineering approaches can be difficult to pursue in nonmodel systems, especially in perennials that may take years to develop a focal phenotype (e.g. flowering) and are often inefficient to transform. Our impression is that some journals are beginning to require phenotypic validation of evolutionary inferences as a prerequisite for publication and peer review. This is a ridiculous standard for some systems, such as hard-to-transform perennials, and it thus biases studies toward species with specific life-history traits. This standard contributes to another growing falsehood—i.e. that evolutionary conclusions require functional validation. This falsehood fails to recognize that evolutionary inferences can be correct even if they cannot be confirmed experimentally. Subtle phenotypic and genetic differences may be undetectable under the limited conditions of an experiment, but they may nonetheless be crucial to a selective response over evolutionary time (Muyle et al. 2022).

Adaptive evolution: polygenic adaptation

Polygenic adaptation is evolutionary change driven by simultaneous selection on variants at many loci, each contributing a small effect to a quantitative trait (Pritchard et al. 2010). The concept derives from Fisher's infinitesimal model (Fisher 1919), which posits that gradual phenotypic evolution can be explained by Mendelian inheritance at an infinite number of loci of infinitesimal effect (Barghi et al. 2020). When a polygenic trait is under selection, the response increases the frequency of variants of small effect across many loci, potentially reaching a phenotypic optimum without fixing individual variants. However, many shifts at individual loci may be too subtle to detect as soft or hard sweeps, raising the question: To what extent does polygenic adaptation affect the ability to detect selection from population genomic data?

Polygenic selection can lead to selective sweeps

To address this question, it is first important to note that polygenic adaptation, hard sweeps, and soft sweeps are not mutually exclusive (Barghi et al. 2020). The effect sizes of variants contributing to polygenic adaptation are likely skewed toward many loci of small effect, a smaller number of loci of intermediate effect size, and a handful of larger-effect loci (Barghi et al. 2020). The outcomes of polygenic selection depend on numerous factors, including the number of loci contributing to the trait, their effect sizes, the dominance of individual variants, epistatic interactions, recombination, and the nature and timing of environmental factors that affect the trait (Barton et al. 2017). Nonetheless, large and medium-effect loci under selection can produce polymorphism patterns consistent with a selective sweep (Fig. 4), as shown via simulation. Stetter et al. (2018)—building on earlier work by Thornton and Long (Thornton et al. 2013; Baldwin-Brown et al. 2014)—used forward simulations to study the temporal dynamics of genetic diversity associated with directional selection on a quantitative trait. Their quantitative trait was relatively simple, with 20 underlying QTL. After simulating directional selection under varying selection strength and a range of demographic models, they observed both soft and hard selective sweep signatures at individual loci. The overarching point, again, is that polygenic adaptation can lead to detectable sweep patterns (Barghi et al. 2020). Hard sweeps, soft sweeps, and polygenic adaptation differ in some, but not necessarily all, of their characteristic signatures in population genomic data.

Figure 4.

A cartoon of the effect sizes of loci contribution to a polygenic trait (inset), along with the trajectory of allele-frequency shifts when the trait is under selection. The colored circles represent four loci with different effect sizes. While it may be possible to detect large shifts in allele-frequency for a locus with large effect (green), shifts in allele frequencies are expected to be less distinct for loci with smaller effects.

Environmental associations as case studies for local adaptation and polygenic selection

Population genomic analyses tend to focus on identifying sweeps and rarely consider the genetic architecture of traits or the dynamics of polygenic adaptation. In fact, “polygenic adaptation” is mentioned in only ∼4% of the papers in our dataset (Table 1). There is, however, an important empirical exception: associations between genotypes and environmental data. These associations go by several names (Bragg et al. 2015; Rellstab et al. 2015)—e.g. environmental associations, environmental genome-wide associations, landscape genomics, and genotype-environment associations (GEAs)—but they have become central for detecting genetic variants that contribute to local adaptation. GEA analyses implicitly study polygenic adaptation, because they explore a multifaceted trait (e.g. climate tolerance) with a polygenic basis. They also serve as a first step toward projecting the genetic fate of populations under predicted climate shifts (Fitzpatrick and Keller 2015; Aguirre-Liguori et al. 2021).

An early GEA application examined potentially adaptive differences among loblolly pine populations by identifying loci where allele frequencies aligned more closely with environmental variables than geographic structure (Eckert et al. 2010). At roughly the same time, Coop et al. (2010) defined approaches to detect putatively adaptive variants using a covariance matrix across samples from multiple populations, leading to the development of BayEnv (Günther and Coop 2013). Today, GEA is more often explored using latent-factor mixed models (Caye et al. 2019) or multivariate ordination (Forester et al. 2018). It is also worth noting that GEA is likely to be most successful with samples of multiple individuals from multiple populations, so it may not be powerful with some common sampling designs—e.g. gene bank repositories with single accessions representing each sampling locality (Lei et al. 2019).

GEA has been particularly effective when combined with other experiments and data sources, such as common garden studies. For example, Fournier-Level et al. (2011) found that many variants in A. thaliana were adaptive in a localized environment. Another pattern that has emerged is that chromosomal inversions can generate signals that exceed those in other parts of the genome (Fang et al. 2012; Pyhäjärvi et al. 2013). Because inversions create large blocks of associated markers, they may need to be excluded from comparisons to detect other potentially adaptive loci. GEA presents several other challenges, such as the need to properly account for population structure (Günther and Coop 2013). There can also be difficulties in handling multiple environmental variables that are often highly correlated (Hoban et al. 2016), but formal statistical approaches can address this issue explicitly (e.g. Fiscus et al. 2025). Additionally, multiple environmental factors can interact to create complex selective pressures, shaping allele frequencies in patterns that may not align with those expected from single environmental variables alone (Lotterhos 2023).

Despite these challenges, GEA has provided remarkable insights into some features of selection on plant genomes. For example, Exposito-Alonso et al. (2018b) used WGR data from hundreds of A. thaliana accessions to identify alleles correlated with climatic variables such as drought and temperature seasonality. They showed that local adaptation is highly polygenic, without evidence for classical hard sweeps. In a related analysis, they demonstrated a link between polygenic adaptation and evolutionary responses to local climates (Expósito-Alonso et al. 2019). GEA has also been central for interpreting the genetic basis of repeatability, or parallel adaptation, by assessing whether homologous/orthologous genes exhibit associations across different lineages. Yeaman et al. (2016) first applied this approach to two conifer species that diverged ∼140 million years ago. They identified 47 genes associated with climatic variables (especially temperature and cold hardiness) in both species, indicating potential parallel adaptation. Their results imply that climate adaptation can be genetically constrained despite the highly polygenic nature of the trait. More recently, they applied this approach to 25 plant species with divergence times ranging from 2.5 Mya to >300 Mya (Whiting et al. 2024). Genes with specific functions were overrepresented in climate associations, particularly in core physiological pathways. In short, these analyses suggest that selection can be identified at least for some major effect loci, even for a highly polygenic trait such as climate adaptation. These results may also partially reflect that local adaptation across heterogeneous environments can favor genetic architectures composed of fewer, larger-effect loci (Yeaman 2022).

Future themes in polygenic selection

Historically, polygenic selection has been all but ignored in the plant population genomics literature, because the major emphasis has been on finding genes of large effect through sweep mapping. However, identifying the number and effects of contributing mutations will be a major emphasis moving forward, in part by detecting correlated shifts in allele frequencies across many trait-associated loci. Under polygenic adaptation, each locus may have only a weak population-genetic signal, but alleles that collectively increase a trait of interest should shift systematically in the same direction across populations. Such a signal can be detected, for example, by weighting allele frequencies at Genome-wide association studies (GWAS)-identified loci by their estimated effect sizes (Turchin et al. 2012; Berg and Coop 2014; Barghi et al. 2020). Two complications are that loci are often pleiotropic, so an allele-frequency shift may instead reflect selection acting on a correlated trait, and that epistatic interactions add complexity, since an allele's effect size can depend on the genetic background.

GWAS-associated loci are not always under selection, but GWAS signals can be interpreted to infer the evolutionary forces—including purifying selection and balancing selection—that maintain genetic variation for quantitative traits (Josephs et al. 2017). For example, loci associated with cis-regulatory variation in C. grandiflora have lower minor allele frequencies than expected compared to controls, suggesting purifying selection against associated variants (Josephs et al. 2015). In contrast, cis-regulatory variants show an excess of common alleles in Mimulus guttatus, but trans-acting regulatory variants show evidence of purifying selection (Brown and Kelly 2022). More studies examining the allele-frequency spectra of QTL loci relative to null distributions will aid inference about the nature of polygenic selection, although fully accounting for ascertainment biases and identifying a true “null” distribution are challenging.

Genomic prediction methods are also useful. As an example, Metheringham et al. (2025) recently investigated polygenic adaptation in a population of European ash trees after invasion by a fungal pathogen. Using genome-wide markers and genomic prediction models trained on phenotyped trees, the authors identified thousands of loci contributing small effects to disease-resistance. They then estimated genomic breeding values for individual trees. By comparing breeding values from pre-epidemic adult trees with postinfection juveniles, they detected a significant shift toward higher predicted resistance in the younger cohort. They inferred that strong viability selection acted within a single generation to remove 31% of juvenile trees, an evolutionary force sufficient to drive detectable changes in allele frequencies across many loci.

In addition to the broader use of genomic prediction tools in an evolutionary context, long-term “evolve and resequence” experiments (Long et al. 2015) will continue to yield insights into the dynamics and architecture of polygenic selection. There are, however, few plant experiments that mirror successful long-term experiments in E. coli and Drosophila; the most prominent are probably the maize long-term oil selection experiment (Laurie et al. 2004; Beissinger et al. 2014) and the barley composite crosses (Landis et al. 2024). There have been several shorter-term efforts in A. thaliana, although its selfing nature is likely to limit many types of evolutionary insights (Long et al. 2015). Even in outcrossing species, the understandably small sample sizes in experiments can lead to long-range linkage effects that make it challenging to discern the number of loci under selection.

Balancing selection

Balancing selection maintains advantageous genetic diversity within a population by preventing the fixation of any single allele. Several mechanisms can drive this outcome, including heterozygote advantage (overdominance), negative frequency-dependent selection, antagonistic selection, and spatially or temporally varying selection that favors different alleles across environments or time (Bitarello et al. 2023). Regardless of the mechanism, balancing selection typically results in linked variants held at intermediate frequencies (Charlesworth 2006). However, identifying this signature is notoriously difficult because recombination gradually erodes the genomic signal surrounding a causative variant. Consequently, haplotypes under long-term balancing selection can be quite small (Charlesworth 2003), perhaps even a single SNP.

Somewhat to our surprise, only 8.6% of the studies in our dataset explicitly mentioned the term “balancing selection” (Table 1). We suspect the infrequent focus on balancing selection reflects challenges in its detection rather than its overall prevalence in plant genomes. Evidence from human genomics suggests that balancing selection may affect over a thousand alleles (Munby and Przeworski 2025)—including 1% to 8% of human genes (Bitarello et al. 2018)—raising the possibility that it affects a similar number in plants.

How is balancing selection detected?

Early approaches to detect balancing selection, such as Tajima's D (Tajima 1989), focused on identifying an excess of intermediate-frequency variants. The Hudson–Kreitman–Aguadé framework (Hudson et al. 1987) also has the potential to identify balancing selection by identifying loci that harbor unusually high diversity relative to divergence. However, contemporary studies emphasize that no single summary statistic is sufficient to infer balancing selection; instead, inference typically emerges from combining multiple signals (DeGiorgio et al. 2014; Cheng and DeGiorgio 2020), including allele-frequency patterns (DeGiorgio et al. 2014; Siewert and Voight 2017), divergence contrasts (Innan 2006), and haplotype structure (Siewert and Voight 2017).

The hallmarks of a genomic region under balancing selection include elevated local polymorphism, clusters of variants that segregate at similar frequencies, and reduced fixation relative to an outgroup. Trans-specific polymorphisms provide particularly compelling evidence, because shared haplotypes across species imply maintenance of alleles over long evolutionary intervals (Leffler et al. 2013) (Fig. 5). Importantly, the genomic footprint of balancing selection is spatially restricted because recombination rapidly erodes linked variation, making the analytical scale—i.e. window size—a critical determinant of detection power (Bitarello et al. 2023). Recent analytical advances increasingly integrate population-genetic theory with ML and genealogical inference. For example, deep learning approaches can jointly model demography and selection, distinguishing genome-wide demographic effects from locus-specific balancing selection without reliance on predefined summary statistics (Sheehan and Song 2016). Complementary frameworks, such as ARGs, leverage local genealogies to identify regions with unusually deep times to the most recent common ancestor, which can be a consequence of long-term allele maintenance. These developments reflect a broader conceptual shift: balancing selection is best understood in terms of its effects on genealogical structure rather than on allele frequencies alone.

Figure 5.

An illustration of a trans-specific polymorphism, which can be indicative of a balancing selection. The neutral variant is polymorphic only in one species, whereas the transpecific G–T polymorphism is polymorphic in two species. Outgroup rooting helps differentiate the derived T mutation from the ancestral G variant.

An emerging view is that balancing selection may often act recurrently or transiently rather than continuously over deep evolutionary time (Bitarello et al. 2023), due to temporally fluctuating selection. Fluctuating selection occurs when the fitness effects of alleles vary across time, with different alleles favored in different seasons, years, or environmental conditions. A few studies have provided evidence for temporally alternating allele frequencies in plants—e.g. thousands of SNPs in M. guttatus exhibit temporally alternating allele-frequency shifts due to conflicting selection regimes (Kelly 2022), and oak species show evidence for fluctuating allele-frequency changes associated with climate shifts (Saleh et al. 2022; Caignard et al. 2024). In practice, however, fluctuating selection is difficult to confirm empirically, because it requires temporal sampling of allele frequencies across ecologically variable time periods (Bergland et al. 2014; Buffalo and Coop 2020). Alleles maintained by fluctuating selection can also produce genomic signatures that resemble those of directional selection, complicating inference and perhaps contributing to an underestimation of the prevalence of balancing selection (Wittmann et al. 2023; Soni and Jensen 2024).

Empirical examples

The clearest examples of balancing selection in plants occur at loci associated with mating systems (Delph and Kelly 2014), particularly self-incompatibility (SI) loci. At the S-locus of Solanaceous and other SI species, negative frequency-dependent selection favors rare alleles, allowing numerous allelic lineages to persist for extended evolutionary periods (Richman et al. 1996). SI loci are canonical examples because they frequently exhibit extreme allelic diversity with allelic lineages that are much older than the species in which they are found. The signature of balancing selection is easier to detect at SI loci than some other genomic regions because SI alleles require multiple allelic variants with pleiotropic interactions, driving the evolution of a cluster of tightly linked genes for functional SI interactions (Charlesworth et al. 2005). In theory, the maintenance of high diversity can also shield linked deleterious mutations from purifying selection; evidence from A. thaliana and A. lyrata suggests that this “sheltered load” varies as a function of the dominance of SI alleles (Le Veve et al. 2024).

SI loci are not the only loci related to mating systems that are subject to balancing selection. A recent example comes from heterodichogamy, a system in which individuals consistently differ in the temporal order of male and female flowering. Rare flowering morphs gain a reproductive advantage because their flowering schedules overlap with those of more compatible partners. In walnuts, the flowering morphs differ in SVs near a single gene, TPPD1, that affects flower development (Groh et al. 2025b). The SVs have remained polymorphic in the walnut genus Juglans for ∼40 million years. A strikingly similar phenomenon has been documented in avocados, even though they exhibit a different (and independently-derived) form of heterodichogamy (Groh et al. 2026).

Although balancing selection is prominent in mating-system genes, the classic example comes from the Major Histocompatibility Complex (MHC), a cluster of immune genes, in humans. High MHC diversity is maintained through pathogen-mediated selection, including heterozygote advantage and negative frequency-dependent selection, leading to transpecies polymorphisms (Hughes and Nei 1988; Klein et al. 1993). The MHC has direct parallels to plant disease-resistance genes, which are also prone to balancing selection (Charlesworth 2006). The best-characterized example is the A. thaliana rpm1 gene (Stahl et al. 1999; Tian et al. 2002). rpm1 encodes a receptor that recognizes pathogen effector proteins. Functional and nonfunctional rpm1 alleles segregate across populations, a pattern of polymorphism that has been maintained for extended evolutionary periods despite the apparent advantage of resistance. Experimental analyses suggest that the maintenance of variation reflects fitness trade-offs: resistance alleles confer protection against pathogens, but they impose measurable fitness costs in pathogen-free environments (Stahl et al. 1999; Tian et al. 2002). As a result, fluctuating pathogen pressure and ecological heterogeneity maintain both resistant and susceptible alleles. Göktay et al. (2021) extended observations about disease-resistance genes by cataloging SVs across the genomes of 1,300 A. thaliana accessions. SVs encompassing disease-defense genes occurred at intermediate frequencies, perhaps signaling widespread balanced polymorphisms.

Capsella rubella has been a useful species for identifying balancing selection, because this selfer generally lacks genetic diversity across its genome (Brandvain et al. 2013). This low diversity makes it comparatively easier to identify regions of aberrantly high diversity, especially compared to its outcrossing sister species, C. grandiflora, which has higher diversity throughout its genome. In C. rubella, and even more so in the selfing species C. orientalis, the regions of high diversity generally map to regions harboring disease-resistance genes (Gos et al. 2012; Koenig et al. 2019). One interesting feature of selfing species is that they are not commonly expected to experience balancing selection due to overdominance, but frequency-dependent selection may act since it is not strongly affected by mating system (Glémin 2021).

Chromosomal inversions provide additional examples of balanced polymorphisms. An inverted region captures variants already present in the region, and it also accumulates segregating variants that differ between inversion arrangements. When recombination is limited or eliminated, and when rates of homologous gene conversion among arrangements are low, high rates of allele-frequency differentiation can result (Guerrero et al. 2012), and long haplotypes associated with the inversion can occur (Kirkpatrick and Barton 2006; Fang et al. 2012). Because a potentially large number of variants may contribute to a specific signal of selection, inversion variants can outweigh other portions of the genome in outlier analyses, such as environmental association studies (Fang et al. 2012). In M. guttatus, for example, a large inversion differentiates coastal perennial and inland annual ecotypes, capturing alleles affecting flowering time, growth, and drought escape, with reciprocal fitness trade-offs maintaining both arrangements across heterogeneous environments. Similar dynamics occur in sunflowers, where multiple inversions distinguish dune and nondune ecotypes of Helianthus petiolaris, preserving locally adaptive allele combinations despite ongoing gene flow (Huang et al. 2025a). In teosinte and maize, inversions that differentiate populations across highland–lowland gradients and are maintained by environmentally dependent selection elevation differences (Pyhäjärvi et al. 2013; Calfee et al. 2021). Structural rearrangements also underlie mating-system polymorphisms. In Primula, the heterostyly supergene involves a hemizygose region that is present in thrum flowers and absent in pin flowers. Hemizygosity can suppress recombination, similar to the repression of recombination in inversions. The Primula floral morphs are maintained through negative frequency-dependent selection (Huu et al. 2016; Li et al. 2016).

Balancing selection: prospects

Balancing selection is known to preserve diversity in genes involved in breeding systems and defense, but it is not yet clear whether these examples reflect conspicuous cases of a relatively rare phenomenon or instead indicate a process that acts more broadly (Charlesworth 2006; Fijarczyk and Babik 2015). Statistical power is likely to remain an impediment to detection, because a substantial fraction of loci under balancing selection likely have signatures eroded below detectable thresholds (Brandt et al. 2026). No single statistic is likely to distinguish balancing selection from the confounding effects of population structure, bottlenecks, or background selection (DeGiorgio et al. 2014; Cheng and DeGiorgio 2020), but approaches that simultaneously integrate allele-frequency spectra, divergence patterns, and haplotype structure outperform single-statistic tests (Cheng and DeGiorgio 2020). ML frameworks that jointly model demography and selection show promise (Sheehan and Song 2016), and ARGs may be particularly powerful for identifying regions with unusually deep genealogies (Deng et al. 2025). ARGs generally require phased sequence data, however, although long-read sequencing is reducing this impediment. Even with improved methods, recent or transient episodes (e.g. fluctuating selection) can be indistinguishable from positive selection or neutral drift (Gao et al. 2015). These observations imply that even large-scale genomic surveys will substantially underestimate the true prevalence of balancing selection, even with the very best methods.

When a candidate locus is identified, determining which form of balancing selection is operating—heterozygote advantage, negative frequency-dependent selection, or fluctuating selection—typically requires experimental evidence beyond what genomic inference can provide (Llaurens et al. 2017). Knowing the mechanism can help answer a longstanding question—i.e. whether balanced variants constitute a reservoir of standing genetic variation that facilitates rapid adaptation to novel or changing conditions or whether they more often represent evolutionary stasis maintained by antagonistic trade-offs (Hermisson and Pennings 2005).

Adaptive introgression

When we evaluated the use of specific terms across our exemplar dataset, we were surprised that “introgression” was the single-most mentioned term in our queries (Table 1). Its prevalence reflects that researchers have long recognized the potential adaptive power of introgression in plant evolution (Rieseberg and Wendel 1993; Rieseberg et al. 2003), an effect far more prominent in plants than animals (Mallet 2005). Although the term “adaptive introgression” was mentioned less frequently than “introgression” alone (Table 1), there is a growing awareness that introgression profoundly affects the adaptive landscape. Adaptive introgression follows an initial hybridization even, which can transiently elevate genetic diversity; further backcrossing; and an eventual sweep of the introgressed fragment (Fig. 6).

Figure 6.

A representation of the process of adaptive introgression. The top panel illustrates the process of adaptive introgression, in which an adaptive variant (star) is introduced from a donor species into a recipient species. Initially, the variant is at low frequency in the recipient species but ultimately rises in allele-frequency, potentially becoming fixed. The bottom graphs illustrate nucleotide diversity over time in the region of adaptive introgression. Diversity in the region is shown for the two parental species (left), for the recipient species when nucleotide diversity can transiently increase due to the presence of segregating variants at the region (middle), and then the sweep reduces diversity in linked regions (right).

Documenting adaptive introgression

The inference of introgression has been bolstered by numerous methodological advances (Huang et al. 2025b), especially the ABBA–BABA framework (Durand et al. 2011; Sousa and Hey 2013). ABBA–BABA approaches scan for excess allele sharing and can localize candidate introgressed tracts using related, window-based methods like fd and fdM (Malinsky et al. 2015; Martin et al. 2015). They can, however, have high false-positive rates in some cases, such as unequal mutation rates among taxa (Frankel and Ané 2023; Pang et al. 2025). Other methods rely on genealogical inference. For example, Twisst2 infers genealogies across chromosomal regions for multiple taxa and then flags regions where aberrant topological relationships are not readily explained by incomplete lineage sorting (Martin 2025). Probabilistic models provide another powerful framework for detecting introgression. Hidden Markov models have been used widely to infer local ancestry (Svedberg et al. 2021; Sun et al. 2025). These models treat ancestry along a chromosome as a hidden state and use observed patterns of genetic variation to probabilistically infer the boundaries of introgressed segments.

While these and other methods can identify introgressed regions, they do not necessarily demonstrate whether the regions confer an adaptive benefit. One straightforward solution is to infer whether introgressed regions overlap with putative selective sweeps (e.g. Numaguchi et al. 2020; Gong and Han 2022). Another approach is to infer whether such regions are enriched for putative functions like stress response or GEAs (e.g. Marburger et al. 2019; Leroy et al. 2020; Morales-Cruz et al. 2021). For example, an early study generated transcriptomic data from 13 wild tomato species and performed phylogenomic analyses to detect recent and historical introgression at specific loci. Some of the potentially introgressed regions associated with phenotypes related to variables like altitude, temperature, seasonal climate variability, water, pH, and soil characteristics (Pease et al. 2016), hinting at adaptive effects.

A few methods have been developed to jointly infer adaptation and introgression. Racimo et al. (2016) introduced metrics to detect adaptive introgression based on the number and frequency of alleles shared uniquely between archaic and modern populations. Another tool is VolcanoFinder, which detects characteristic patterns of genetic diversity around an introgressed, beneficial variant (Setter et al. 2020). ML methods for joint inference are also emerging (Gower et al. 2021; Zhang et al. 2023), as is the use of ARGs (e.g. Hubisz et al. 2020; Brandt et al. 2024). Thus far, these methods have primarily been applied to human data and may require some adjustment for plant datasets. They nonetheless hold great promise; our understanding of the human species has been transformed by these methods, including inferring that archaic ancestry from Neanderthals and Denisovans has contributed to adaptations like immune response and high-altitude physiology (Enard and Petrov 2018; Zhang et al. 2023; Villanea et al. 2025).

Empirical trends

One emerging theme is the prevalence of introgression between crops and their wild relatives (Janzen et al. 2019; Blanco-Pastor 2022). Gene flow from wild relatives to crops can restore genetic diversity lost during domestication and also introduce genetic variants that enable global expansion to new locations and environments. For example, the introgression of flowering time alleles has aided adaptation to altered day lengths in new growing locales for rice and maize (Fujino et al. 2010; Guo et al. 2018; Bellucci et al. 2023). Adaptive introgression between crops and their wild relatives has been detected for sunflower (Whitney et al. 2010; Hübner et al. 2019), rice (Jing et al. 2023), maize (Calfee et al. 2021; Yang et al. 2023), common bean (Bellucci et al. 2023), date palms (Flowers et al. 2019), grapevine (Xiao et al. 2023), and cotton (Yuan et al. 2021), to name just a few examples. The extent of wild-to-crop introgression suggests it is a necessary precursor for the geographic expansion of some crops (Janzen et al. 2019), but it can have unintended consequences. For example, wild maize (teosinte) acquired herbicide resistance from cultivated maize, helping establish it as a weed in Europe (Corre et al. 2020), just as the introgression of herbivore resistance has led to weediness in cultivated sunflower (Whitney et al. 2006).

WGR studies have revealed genomic biases to introgressed regions, particularly that they tend to be more common in high-recombination regions of the genome (Calfee et al. 2021; Morales-Cruz et al. 2021; Feng et al. 2023; Wang et al. 2023). This relationship is expected because positive selection will be more efficient when high-recombination rates decouple beneficial variants from deleterious load (Edelman and Mallet 2021). However, in some species, such as wild oak trees (Fu et al. 2022), introgressed alleles tend to be found in low-recombination regions. In these regions, recombination is less likely to break up favorable combinations of beneficial alleles across a large, introgressed haplotype. Interestingly, introgression can be adaptive not because it introduces a beneficial variant but instead because it reduces mutation load (Gaut et al. 2018). In such cases, the retention of an introgressed region might be more favorable in low-recombination regions, but we do not know how often this occurs. It is worth noting that introgressed regions in poplar do not decrease mutation load (Liu et al. 2022) but they do in wheat (He et al. 2019).

Introgressed regions can also be functionally biased, as usually inferred by GO analyses. Although introgressed regions harbor genes with a wide range of functions, there are some themes. One is that introgressed alleles commonly contribute to environmental adaptation. In Populus, gene flow from cold-adapted relatives has helped North American species adapt to higher latitudes (Suarez-Gonzalez et al. 2018), while in Europe, introgression aided local adaptation to the short growing seasons of Northern Scandinavia (Rendón-Anaya et al. 2022). Similar dynamics apply to sympatric Asian oak species (Fu et al. 2022), sunflowers (Owens et al. 2025), and the evolution of invasiveness (Bieker et al. 2022; Hodgins et al. 2025). Not surprisingly, the functions identified as enriched within sweeps, such as flowering variants and disease-defense genes, also appear to be enriched within regions of adaptive introgression (Bechsgaard et al. 2017; Suarez-Gonzalez et al. 2018; Morales-Cruz et al. 2021).

A fascinating example of a functional bias comes from introgression among polyploids. A series of studies has investigated introgression between polyploids of two Arabidopsis species, A. arenosa and A. lyrata. The work detected introgression from older A. arenosa polyploids to more recently formed A. lyrata polyploids (Marburger et al. 2019). The introgressed regions were enriched for genes involved in meiotic functions that fine-tune crossover control. The implication is that pre-adapted alleles from polyploid A. arenosa helped stabilize meiosis in nascent A. lyrata polyploids (Marburger et al. 2019). A more recent study of polyploid A. lyrata populations has also detected introgression of meiosis-related genes, suggesting it could be a common phenomenon during polyploid establishment (Scott et al. 2025). Adaptive introgression need not be limited to taxa with the same ploidy level; for example, frequent interploidy introgression from diploid Brassica species into tetraploid B. napus fueled phenotypic diversification during domestication (Wang et al. 2023).

What is left to learn about adaptive introgression?

To date, most population genomics studies have focused on introgression between a pair—or at most a few—populations or species. It has been proposed, however, that adaptive introgression plays a central role in species' divergence and specialization during adaptive radiations (Combrink et al. 2025), implying that a full understanding of its effects and role requires an expansive, multispecies scale. It is becoming clear that many—and maybe all?—plant genera are typified by distinct species with extensive histories of recent and ancient introgression (Cannon and Petit 2020; Leroy et al. 2020; Rendon-Anaya et al. 2021; Buck and Flores-Rentería 2022; Wang et al. 2026). In these cases, the evolutionary unit of interest is not an isolated species or pair of species but a shared pool of adaptive variants across many species. Yet, population genetics across multiple species, or an entire genus, have been explored in only a few plant groups thus far, such as (Helianthus) (Todesco et al. 2020; Soudi et al. 2023), Solanum (Pease et al. 2016; Hamlin et al. 2020), Zea (Chen et al. 2022; Tittes et al. 2025), and Vitis (Morales-Cruz et al. 2021; Wang et al. 2026).

Given the prevalence of adaptive introgression across multiple species, plant population genetics needs a paradigm shift to focus on genus- or family-scale adaptive processes. However, a multispecies framework has concomitant challenges. One is that the identification of adaptive introgression currently relies primarily on pairwise comparisons. Thus, one challenge is methodological: detecting the regions and timing of introgression events across many species or populations simultaneously. The development of ARGs, Local Ancestry Inference, and ML may help address some aspects of complex reticulate histories, but method development is an urgent need within this expanded, multispecies paradigm. Another challenge is to understand the underlying biology—i.e. the selection pressures, adaptive benefits, functional biases, and phenotypic traits—that shape the retention of introgressed regions. In wild-to-crop comparisons, there is compelling evidence that introgression contributes to flowering time, disease-resistance, and agronomic traits in crops, but the picture is much less clear in wild-to-wild adaptive introgression. Finally, abundant evidence for introgression among species raises a complementary question: which loci are resistant to introgression and thus may be responsible for maintaining the phenotypic and evolutionary distinctiveness of individual species (Wang et al. 2026)?

Conclusions

As we wrote this review, three things became obvious. The first is that the field has come a long way since our 2005 review (Wright and Gaut 2005). Population-genetic approaches have yielded much greater insight into the prevalence, dynamics and determinants of different selection modes. The second is that significant knowledge gaps remain. We have mentioned some of those gaps, but we emphasize that greater effort is needed to study DFEs, selection coefficients, and other basic evolutionary parameters that have largely been ignored, with a commensurate expansion of a multispecies scope. Finally, there are substantial limits to our ability to infer natural selection from WGR data. Take balancing selection as one example: promising new methods are likely to help us identify balanced variants, but many (if not most) of these variants will remain invisible. Similar limits apply to identifying genetic variants that are minor contributors to polygenic selection, detecting soft sweeps, and categorizing noncoding regions that are subject to purifying selection.

That is not to imply that further investigation is a waste of effort. We envision four themes that need further emphasis. The first, as we have mentioned throughout, is the application and development of newer methods. ARGs may be especially instrumental for detecting old variants maintained by balancing selection and for interpreting the history of introgressed regions. ML (and AI) approaches are poised to be equally useful, especially as these methods continue to explode in breadth and utility (Whitehouse et al. 2024). Their applications range from LLMs to identify sequence conservation across genomes, to neural networks for categorizing sweeps, to methods that perform joint inference of processes such as adaptation and introgression. Yet none of these approaches is likely to fully circumvent issues of low statistical power, and they will continue to be limited by the accuracy of simulated training data. Since plants tend to have strong population structure and rampant introgression, it is challenging to accurately capture true demographic history using simulations. If methods are trained to detect subtle signatures of selection, they may be as apt to identify mismatches between true and simulated histories as it is to detect bona fide cases of selection.

A second theme is the maintenance of community standards. The plant literature is uneven with respect to the competence of population-genetic inference. Population-genetic analyses of plant WGR data often appear as an “add-on” to papers that are focused on cataloging the origin, genotypic variability, or phenotypic variability of a crop. These “add-on” analyses often use off-the-shelf methods, whether appropriate or not, that contribute to a false sense of inferential limits and perhaps even misleading results. No journal is immune, but we have seen especially disturbing examples in the highest profile journals. So, what should our community standards be? This question deserves conversation beyond this article, but Brault et al (2026) have published a pertinent and thought-provoking commentary. Building on their insights, we believe that community standards should include at the minimum: (i) clear presentation of methods, which are often obfuscated in short-format journals, (ii) public availability of variant calls (e.g. VCFs) after publication, with the inclusion of monomorphic sites in genomic VCFs when possible, (iii) systematic reporting of a suite of standardized statistics like π, and (iv) the cessation of a priori empirical cutoffs for sweep mapping. Instead, forward or coalescent simulations, based on demographic inference, should be a minimum expectation for defining significance, even given potential inaccuracies of simulations. To this end, efforts to create standardized means of specifying demographic models (Gower et al. 2022) will permit easier comparisons across studies (Adrion et al. 2020).

A third theme is larger datasets and the integration of different data types with WGR-based inferences. For the former, human datasets have progressed from thousands to hundreds of thousands of individuals (e.g. Gudmundsson et al. 2022; UK Biobank Whole-Genome Sequencing Consortium 2025), making it possible to detect a broader swathe of alleles associated with phenotypes or that vary meaningfully in frequencies among groups due to selection. Similar sample sizes may be achievable for some plant species like rice—e.g. the International Rice Research Institute (IRRI) in Los Baños, Philippines contains ∼130,000 accessions—but it will prove to be difficult for species without large germplasm collections. For crops with sizable collections, we advocate WGR sequencing of complete collections as soon as possible, even if they have already been genotyped by reduced-representation methods. It is of course equally important to sample the wild relatives of crops and ecologically important noncrop species to better understand evolutionary processes and for ecological and conservation applications. However, as sample sizes increase, it brings challenges for efficiently storing, parsing, and analyzing variant calls, particularly as the analysis of SVs becomes more common. At scale, VCF files are no longer efficient; the use and development of tree sequences and new file formats, such as zarr files, are crucial (Kelleher et al. 2019; Czech et al. 2025). New software tools are facilitating analysis of these larger datasets by improving speed and accuracy (Mirchandani et al. 2025; Jeffery et al. 2026; Pope et al. 2026).

Finally, new(er) types of data will provide further clues about selection on plant genomes. One obvious example is long-read sequencing and the emergence of accurate, phased assemblies (Cochetel and Cantu 2026). These assemblies facilitate SV characterization and the analysis of difficult-to-study features such as centromeric and telomeric repeats. The phased nature of these assemblies is useful for many population-genetics applications. The integration of epigenetic markers (e.g. methylation data and ATAC-seq data) and 3D chromosome conformation with population-genetic data is another important and growing emphasis because the dynamics of selection are often linked to chromatin status (Drinnenberg et al. 2019). Since chromatin can vary by cell type, increased exploration of single-cell methods and inferences may yield further insights into relationships among function, chromatin and selection dynamics.

While there is much work to do, our overarching conclusion is one of triumph: we have learned a great deal since the first population-genetic analysis of plant nuclear DNA sequences. The plant evolutionary genomics community has established a firm foundation for understanding the signatures and impacts of selection on plant genome structure and function. The rapid pace of methodological and technical innovation promises an even more exciting future, with novel insights just beyond the horizon.

Author contributions

All authors participated in the conception and writing of this review. B.S.G., P.L.M., and T.W. performed analyses.

Supplementary material

Supplementary material is available at Molecular Biology and Evolution online.

Funding

This work was funded in part by NSF DEB-2414478 to B.S.G., from USDA BRAG 2023-33522-41008 to P.L.M, and the Minnesota Agricultural Experiment Station fund MIN-13-122 to P.L.M.

Data availability

No new data were generated for this review. The code for accessing and manipulating the 267 publications is available at https://github.com/pmorrell/Plant_Selection, with additional information in the Supplementary Text.

Artificial intelligence use

Artificial intelligence played diverse roles in the construction of this manuscript; it was used for a basic plot, given the data (Fig. 1); for extracting information from the dataset (Table S1, see the Supplementary Text); as a search engine to find papers and to complement other search protocols (like Google Scholar); and for occasional smoothing rough prose. It was not, however, used to generate original prose content.