Promises and challenges of crop translational genomics

Abstract

Crop translational genomics develops breeding techniques based on genomic datasets to improve crops. Technological breakthroughs in the last ten years have made it possible to sequence the genomes of ever more crop varieties and assisted in the genetic dissection of crop performance in numerous ways. Still, translating research findings into breeding applications, “to move from lab to field”, is challenging. Here, we review recent progress and future prospects of discipline. Genetic mapping and genomic selection employ rapid genotyping of large populations, as does the sequence-assisted characterization and deployment of plant genetic resources. All three approaches have had an impact on breeding for qualitative traits where single genes with large phenotypic effects exert their influence. The complex genetic architectures underlying quantitative traits such as yield and flowering time, especially in newly domesticated crops, will require further basic research, also into the regulation and interaction of genes, and the integration of genomic approaches and high-throughput phenotyping before targeted interventions can be designed. Future translational priorities are the support of genomics-assisted breeding in low-income countries and the adaptation of crops to changing environments.

The concept of translational genomics first arose in medical genetics in the wake of the human genome project¹. Scientists might help improve human health by distilling discoveries made through genetic research^2,3 into new treatments for diseases. The goal of translational genomics in crops is to build on such discoveries to increase yield, nutrient use efficiency and other parameters of plant performance. Methods in humans and crops are partially overlapping: most tools of molecular biology, especially genome sequencing, are species-agnostic. Basic principles of genetics such as mutation, recombination and selection exert their powers in mammals as they do in plants. Initially, some thought the best course of action would be to transfer the knowledge about gene actions in model plants like Arabidopsis thaliana or Medicago truncatula^4,5 to crops. In recent years, the technical constraints on sequencing crop genomes themselves were lifted⁶, and the limits of comparative gene mapping have come to the fore⁷. As a consequence, crop translational genomics has turned into a search for ways and means to enlist genomics datasets and breeding techniques premised on them for the purposes of crop improvement, among which are higher yields, improved nutrient use efficiency, resistance to pests and pathogens and, with ever-increasing importance, getting plants ready for climate change. The remit of translational research overlaps with that of applied research. We believe the key difference to be one of outlook: while the applied researcher enlists scientific tools to further practical aims such as those of commercial breeders, the translational scientist is looking for problems to whose solutions his or her scientific findings might be applied. Despite different points of departure, both groups may arrive at the same destination.

The idea of breeding better crops with the help of molecular biology predates the rise of genomics. Mutation breeding took off after Herrmann Muller had discovered mutagenic effects of radiation in the 1920s⁸. Many translational applications have been proposed over the years (Table 1). Some, like molecular markers, have been taken up widely. Others have yet to come by success. This mixed record has been framed in narratives that range from Panglossian tale to scathing satire. An unabashedly pessimistic vignette of crop translational research was drawn in an engagingly written, albeit judgmental, review⁹ of the field spanning the decades prior to 1990. In its telling, the opportunity costs far outweigh the touted benefits of what is mockingly termed “molecularology”. At the other end of the spectrum are many a genome paper’s introduction, in which the newly assembled genome sequences are presented as the key that will reveal insights into agronomic traits and unlock hidden potentials, be it of staple or underutilized crops.

Table 1. Crop translational genomics: different approaches, their limitations and how to overcome them.

Developments	Approach	Limitations	Way forward
Model plants	Genetic and genomic resources for Arabidopsis thaliana	Representativeness of a single model	Fostering basic research in crop plants
Genetic mapping	Correlating genotype and phenotype in populations	Genotype to phenotype gap Complex genetic architectures GxE interaction	Improving our understanding of gene regulatory networks and the epigenome
Genome sequences	Crop reference genome sequences	Structural variation	Pangenomics
Deployment of plant genetic diversity in breeding	Genebank genomics	Phenotyping bottleneck Lack of recombination Long breeding cycles	High-throughput phenotyping Genomic prediction Altering recombination landscape Gene editing Speed breeding
Gene editing	Targeted induction of mutations	Recalcitrance of non-model organisms Restrictive regulatory environments	Development of species-specific protocols Engagement with stakeholders, legislators, and the general public

Consecutive technological breakthroughs in DNA sequencing have made possible what was unthinkable a mere decade ago: comparing plant genomes in their thousands. We share the excitement begot by these advances and express the optimistic belief that catalogues of genes and alleles in crops and their wild relatives might somehow help us nudge plant performance in a direction conducive to human sustenance. To achieve this goal, scientists have devised numerous applications of genomic research in plant breeding. Here, we discuss the interrelated topics of molecular markers and trait discovery, plant genetic resources, deleterious variants in crop genomes, de novo domestication and interactions between genes.

Markers for genetic mapping and genomic selection

The scientific discipline of genomics is concerned with the structure and function of genomes. It arose to cater to the needs of genetic mapping. That activity refers to the search for causal links between genotype and phenotype. The human geneticists that first attempted it struggled, with the technologies of the day, to discover and apply molecular markers, genetic polymorphisms linked to phenotypes of interest, for even genetically simple traits, namely Mendelian diseases. As DNA sequencing methods improved, the aim was not merely to find as many markers as possible from ever more contiguous bits and pieces of the genome, but also to harness the technology for genotyping, i.e. determining allelic states of sequence variants in many samples simultaneously¹⁰. Statistical methods were developed to associate marker patterns with traits of interest either in experimental¹¹ or natural populations¹².

The most complete instantiation of a genetic map, and these days the most commonly reported on, is a genome sequence that contains all genes along with the intervening non-coding elements¹³. First accidentally, then in systematic surveys, genomicists have discovered structural variants, i.e. differences between individuals of the same species in the presence and arrangement of genetic elements^14,15. This gave rise to the concept of the pangenome¹⁶: a representative collection of complete genome sequences is now deemed necessary to understand evolution within and between species. Beyond their original remit of genetic mapping, genome sequences have been put to many other uses, among them inquiries into the history of human beings and their crops with the help of ancient DNA¹⁷.

This, in a nutshell, is the outline of genomics’ development from markers to pangenomes over the past 40 years. All along the way it has been hoped that knowledge first about the sequences of genes, then about their actions would make it possible to harness beneficial alleles for crop improvement^18,19 (Figure 1). This turned out to be more complicated than initially thought. Two genetic mapping techniques targeting single genes, genome-wide association studies (GWAS) and quantitative trait locus (QTL) mapping, have been found wanting by Bernardo²⁰ in regard to their translational potential (Table 1). He notes the difficulty of converting knowledge about the positions and actions of single genes into better crops and attributes that lack of success to complex genetic architectures^21,22, i.e. those that are characterized by the presence of many interacting genes that influence a trait in question. If all discrete genetic factors of statistical significance exert but small phenotypic effects, the minuscule benefits of introducing any one of them into an elite variety are easily offset by the costs of doing so. We wish to offer an optimistic counterpoint to this cautionary note. Scientists have cloned genes such as the flowering time genes of the cereals²³ that have played outsized roles in crop evolution²⁴ and whose allelic variation is exploited by breeders to the present day²⁵. Practical proposals have been put forward to move from resistance gene cloning to durable resistance in the field²⁶. Genetic mapping is a mainstay of basic research in crop plants because it is an effective means of linking genotype to phenotype. Genome sequence assemblies, cheap genotyping and gene editing for functional dissection have made that goal easier to attain²⁷. Even if any single gene is unlikely to be a silver bullet, knowledge about the molecular basis of yield formation, nutrient use efficiency, disease resistance and other biological processes is accumulating – gene by gene – and one day we may be able to devise discrete targeted intervention to improve crop plants (Figure 2).

Figure 1. Genomics-assisted breeding.

(a) Reference genome sequences are now available for most cultivated plant species. Cost-effective high-throughput sequencing informs about the genetic makeup of breeding material and plant genetic resources. This knowledge, in conjunction with agronomic phenotypes, helps researchers establish links between genotype and phenotype and resolve these to causal sequence variants. Beneficial genes and alleles can be introduced into cultivars by genetic transformation and gene editing. (b) Genomic-assisted breeding has the potential to speed up crop improvement by shortening the breeding cycle, i.e. the time it takes to develop new varieties from existing ones.

Figure 2. From genome sequences to function.

(a) Genetic mapping, the establishment of statistically robust links between genotype and phenotypes, require access to genetic diversity. Crop geneticists can take advantage of curated and freely accessible germplasm repository whose holdings are representative of global diversity in cultivated plants and their wild relatives. Genebank accessions are increasingly being subjected to sequence-based genotyping. A common pattern that emerges is a correspondence between geographic and genetic structure as represented by the matching colors in the world map and the ancestry components. The map was created with mapchart.net. (b) Cost reductions and algorithmic improvements, respectively, in DNA sequencing and sequence assembly have enabled the construction of pangenomes, collection of genome sequences for many individuals of a species. (c) Studies in model organism and crop plants have shown that knowledge of DNA sequence is often insufficient to explain all phenotypic variation. Epigenomic studies identify gene regulatory elements and investigate the impact of epigenetic modifications such as DNA methylation, chromatin openness and histone marks on gene expression. (d) Gene actions can rarely be understood in isolation. The continuous effort of molecular geneticists is required to unravel the molecular functions of genes and dissect complex gene regulatory networks.

Another application of genome-wide genotyping, genomic prediction, gets a better mark from Bernardo²⁰. That method does not put much stock in any single genetic factor, but predicts phenotypes through statistical model predicated on genome-wide marker profiles²⁸. The requisite data are a genetic variant matrix and phenotypes of a training set. Statistical models encapsulating the relationship between genotype and phenotype in the training data are used to predict phenotypes in wider set of breeding lines solely from genetic marker data. As genotypic data is often cheaper and easier to collect than agronomic phenotypes, genomic selection can widen the set of germplasm under breeders’ scrutiny or speed up the development of new varieties (Figure 1). This neat concept has largely been borne out in practice. Genomic selection has been taken up in many crops and is often as accurate as phenotypic selection^29–31. Current practices may be further improved by including structural variants from pangenomes^32,33. Genomic selection might still be more widely adopted than it is at the moment: crops with small breeding communities or those with particularly challenging genomes, such as autopolyploids, are only now acquiring genome sequence assemblies and the attendant infrastructure of genotyping platforms, computational analysis pipelines and logistics that underpin genomic selection^34,35.

Predictive resistance gene atlases

The success of genomic selection notwithstanding, there is a place for single genetic loci with large effects and the molecular tools that find and manipulate them. A case in point are resistance (R) genes, which act in gene-for-gene manner: a plant R gene recognizes the product of a single gene of a given pathogen (the avirulence gene)³⁶. The most common class of R genes involved in that process are nucleotide-binding site leucine-rich repeat (NLR) genes^37,38. Crop varieties bearing suitable R genes can be made impervious to a particular variety (or “isolate”) of a pathogen, but the fast evolution of avirulence genes must be answered by ever new R genes. Rarer, hence widely sought after, are R genes that are effective against multiple isolates or even multiple species and provide longer-lasting resistance. Examples are Lr34 and Lr67 of wheat^39,40 and Xa21 of rice⁴¹.

As NLR genes are conserved in structure and sequence within and between species, they can be targeted by hybridization-based capture sequencing⁴². Thanks to ever-falling sequencing costs, whole genome sequencing might also do the job: pangenomes provide ready access to the sequences of all classes of genes, including those of the resistance sort. Whether compiled through enrichment sequencing or whole genome sequencing, “resistance gene atlases⁴³” of major crops are underway and might well be ready by the end of this decade. Such atlases, possibly complemented by their counterpart catalogues of avirulence factors in pathogens^44,45, could underpin the predictions of protein structures in host and pathogen and how these interact. Förderer⁴⁶ et al. used cryo-electron microscopy to resolve the structure of the complex formed by the wheat resistance protein Sr35 and its avirulence counterpart AvrSr35 in stem rust. This and similar^47,48 research opens up the possibility of “structure-guided NLR engineering”: in the future, variants of R genes might be conceived in the lab to respond to changes in avirulence factors that would otherwise escape recognition by the host. Computationally predicted protein structures are likely to take center stage in that effort. As of now, Förderer et al.⁴⁶ noted discrepancies between their experimentally determined structures and those predicted computationally with AlphaFold2 (ref. ⁴⁹). If successfully implemented by means of better models possibly premised on more and more detailed crystal structures, designer NLRs might be the closest plant analogue to personalized medicine in humans.

Activating plant genetic resources

“Genetic resources” have been recognized as the raw material of plant breeding for decades⁵⁰. The most common instantiation of plant genetic resources (PGR) is, as the International Treaty on Plant Genetic Resources for Food and Agriculture puts it, “material of plant origin, including reproductive and vegetative propagating material, containing functional units of heredity” stored in germplasm repositories called “genebanks”. PGR might also be maintained in situ: traditional varieties in farmers’ fields⁵¹ and wild relatives in their natural habitats⁵². Taking them from there is more difficult than working with the holdings of genebanks⁵³, which can be thought of as living libraries: seeds are stored in shelves in cold rooms; care is taken to keep them viable; seed lots are referred to as “accessions”; each accession is accompanied by “passport data” that record, among other things, its taxonomic status and provenance. We have argued that textual passport records should, with the cheap genome sequencing now at hand, be complemented with “molecular passport data,”⁵⁴ dense genome-wide marker profiles, possibly even whole genome sequencing (Figure 2). Such “genebank genomics” has been conducted in the major cereal crops wheat^55,56, barley⁵⁷, maize⁵⁸ and rice⁵⁹. At present, obtaining full sequences for all accessions in a genebank is a costly undertaking. If technological progress continues apace, “digital genebanks” might be ready in the 2030s.

Marker data for many genebank accessions will be a great boon for collection management: they help spot duplicates, determine heterozygosity and heterogeneity of seed lots and correct decades-old passport records⁵⁴. Complementing genebanks with sequence data does not, by and in it itself, lead to improved varieties. But it could help devise and implement strategies to enhance the value of germplasm repositories. Consider the example of disease resistance. In their perennial search of R genes against new types of pathogens, breeders might turn to genebanks. They face three interconnected challenges. The first can be figuratively expressed as “finding the needle in the haystack⁶⁰”: it is impractical to screen thousands of accessions, each possibly composed of several genotypes. Second, assuming we have come across resistance in a genebank accession, the possibility remains it is not novel, i.e. has already been employed in a released variety. Finally, new and effective alleles need to be crossed into an elite background without compromising crop performance (Table 1).

The first challenge might be tackled with genomic prediction. The same statistical tools that allow breeders to select the most promising recombinants in early stages of the breeding cycle should enable pre-breeders to impute desirable phenotype from a tractably small training set to entire genebank collections. The germplasm in the training set may be chosen so as best to represent the global diversity of the crop or individual germplasm groups (e.g. Japonica rice, temperate maize, winter wheat)⁶¹. If phenotypic data is available for a sufficient number of accessions, trait-customized collections may be selected that are both genetically diverse and segregate for a trait of interest⁵⁶. These approaches have been trialed for biomass and grain yield, respectively, in sorghum⁶² and bread wheat⁵⁶. The “novelty” criterion can be checked for with genome sequences. Once we have sequence databases both of all genebank accessions and all currently grown varieties, it will be straightforward to find out if a putatively beneficial haplotype is already present in elite varieties. Pangenomics also works the other way around: sequence inventories of genebanks can be scoured for novel variants of known genes. As long as the designer NLR mentioned above remain a distant prospect, allele mining of resistance genes in genebank collections is the best way forward to tap sequence diversity in known resistance genes⁶³. Sequence look-up requires at least a rough positional mapping of favorable loci by genome-wide association scans or QTL mapping in PGR x elite crosses⁵⁶.

A decomposition of beneficial phenotypes into discrete genetic factors is anyway required for the final step of introducing allelic variants of PGR into cultivars: introgression. In the case of single loci with large effect such as resistance genes, this can be achieved in a single step via transgenes and gene edits. It might work even across species borders in taxa like the Triticeae that harbor several closely related species, as was demonstrated by the transfer of wheat resistance genes into barley^64,65. Polygenic “gene cassettes” with several resistance genes effective against multiple isolates were proposed as a means of preventing any one of them from being overcome by rapid evolution of single avirulence factors^66,67. Putting together gene stacks requires cloning their constituents parts first. Genome sequences have done away with a dearth of markers or the need for arduous physical mapping and thus expedited resistance gene mapping in crops⁶⁸ and their wild relatives^69,70 (Table 1).

Gene editing may not always be possible for technical reasons. An obstacle to the wider deployment of Cas9 gene editing technology is the recalcitrance of plant genetic resources owing to challenges in tissue culture, regeneration and mutant detection⁷¹. An alternative to transgenic approaches are crosses: a donor is crossed with an elite variety and the offspring is repeatedly backcrossed with the elite parent so as to lower the proportion of donor chromatin by meiotic recombination. But crossovers are rarer than we might wish for. On average, there are one to two per gamete per chromosome per generation. More breeding cycles might be crammed into one year by speed breeding, the drastic shortening of maturation times effected through changes in temperature and light intensity, duration and spectrum^72,73. But even rapid cycling cannot compensate for absent recombination. “Recombination deserts” occupy the pericentric regions of many crops and extend for hundreds of megabases in some^74,75. The effects of structural variants are more local. Inversion polymorphisms prevent cross-overs between their breakpoints between carriers and non-carriers. While much rarer than SNPs or other types of structural variation, inversions are frequent and large enough for breeders to worry about^76,77.

On top of slowing down breeding, long linkage blocks can be freighted with undesirable alleles. Such “linkage drag” is common in “alien” introgressions, which introduce chromatin from crop-wild relatives into an elite background. While alien introgressions have been widely deployed without inordinate penalties on agronomic performance in some crops such as wheat⁷⁸ and tomato⁷⁹, others like barley remain recalcitrant⁸⁰. Some of these obstacles (Table 1) could be removed if we were able to make crossovers more frequent overall or induce them in places where they would otherwise not happen. Promising steps in that directions have been made recently. Mieulet et al.⁸¹ reported that knock-out of a single component of the recombination machinery tripled crossover frequency in rice, tomato and pea. Pericentromeres of the mutants, however, remained recombinogenically inert. “Chromosome engineering⁸²” has been proposed as a remedy against inversions. Specifically, CRISPR/Cas9 gene editing was enlisted both in a model plant⁸³ and maize⁸⁴ to induce inversion to either reverse natural or create new variants. Increased recombination might also be needed to bring about the whole-sale change of the mutational landscape proposed by those concerned with deleterious variants.

De novo domestication

De novo domestication refers to the process of turning wild species into crops, as did the first farmers millennia ago, but in a faster, and presumably more targeted, manner with the help of molecular biology⁸⁵. Frothy claims about “feed[ing] the world through de novo domestication”⁸⁶ are, for the time being, just that. Yet still, new cultigens might be uniquely suited to agricultural environment that past farmers and breeders did not select for and which might better be served by new domesticates rather re-purposed old ones. Examples are perennial grain crops^87,88 and cultivation on saline soil⁸⁹. If domestication in the new crops proceeds along the same lines as in the old, initially large genetic gains might be expected from a few “domestication genes” controlling plant architecture, the dimensions of seeds of fruits and growth habit⁹⁰. Genome sequences and genetic mapping will help finding and deploying them⁹¹.

Evolutionary biologists are fascinated by domestication as a case of rapid adaptation to new environments. They ask questions⁹² like: How long does it take to domesticate a plant? How many genes need to be tinkered with? When does domestication start and crop evolution⁹³ end? These questions also have practical implications. Agriculture has created a new environment, the arable field, in which selection pressures active in pristine habitats are relaxed or even reversed. Alleles that were loss-of-function in wild progenitors might be beneficial and even indispensable in incipient crops. Knock out a couple of genes and you get large, non-dormant seeds that are retained on the inflorescence upon maturity⁹⁰. Domestication of the Neolithic founder crops yielded in the evolutionary short time frame of a few thousand years cultivated plants that are recognizable as such in the archaeological record⁹⁴. It is reasonable to expect that with knowledge about gene action from related species and modern molecular tools at hand, de novo domestication might proceed even faster^95,96. Roughly speaking that will take us to where domestication ended and crop evolution began⁹³. Were we, hypothetically, to re-domesticate wild barley or chickpea, we would end up with plants that look like domesticated forms. But they would have the limited geographic range and yield potential of those species thousands of years ago. These semi-adapted lines would be useful for trait mapping and introgression breeding, but genetic improvement must go beyond domestication genes if de novo domesticates are to become profitable crops in their right. In contrast to initial domestication, subsequent crop evolution operates on quantitative traits such as flowering time, seed dimensions, nutrient composition and disease resistance, which are likely to be under the control of many genes, each with small effects. At the same time, limited standing genetic diversity and low recombination rates are barriers to the rapid combination of beneficial alleles (Table 1). For the new domesticates to outyield their traditional counterparts, we must hope, as expressed by Fernie et al.⁹⁷, that greatly improved knowledge and powerful gene editing will enable us to fast-forward crop evolution.

Deleterious variants

A lack of genetic diversity is often bemoaned in current elite varieties^98,99, but doubts as to the economic relevance of such “genetic erosion” is as old as the diagnosis¹⁰⁰. One reason why large regions of the genome have lower diversity in crops than their wild progenitors is genetic linkage, the co-inheritance of loci that reside on the same chromosome. Selection is at the core of breeders’ business. Leaving aside the diversifying sort, selection at one locus favors one of its alleles to the detriment of all others. But owing to linkage such losses of diversity are not limited to the genes under selection but affect their wider surroundings as well. These “selective sweeps” lead to “haplotype blocks”, where genetic diversity is reduced and the extent of which is determined by local recombination rates. A favorable interpretation of linked selection is that haplotype blocks help breeders maintain combinations of beneficial alleles¹⁰¹. But the “genetic hitchhiking¹⁰²” can also cause deleterious variants to be swept along as long as they do not offset the gains conferred by the target of selection. Since farmers and breeders have imposed selective pressures on crop genomes for a long time, unfavorable alleles may have accumulated in various places.

These theoretical considerations have been borne out by empirical studies in maize^103,104, sorghum¹⁰⁵ and rice¹⁰⁶. Which sequence variation is harmful can, in the simplest case, be inferred by overlaying variants with gene annotations¹⁰⁷. More sophisticated approaches¹⁰⁸ parse multiple sequence alignment to infer evolutionary constraints and check how well amino acid changes in crops abide by them. As reference genomes of more crop wild relatives, possibly of entire genera and families¹⁰⁹, become available, such approaches should become more powerful. Human genetics offers a foretaste of what is to come: machine learning models trained on 233 primate genome sequences predicted the pathogenicity of 4.3 million protein-altering variants in the human genome¹¹⁰. The overwhelming majority (94 %) were not benign. An important caveat in domesticated plants is that variants that compromise fitness in the wild might have the opposite effect in a farmer’s field – and those would arguably be most interesting to crop evolutionists.

But it is still likely that there are many truly deleterious variants in crop genomes. What to do about them? Means are needed to purge them from crop genomes. One proposal is to include evolutionary constraints in genomic selection models to penalize deleterious alleles¹¹¹. A farther-reaching scheme was put forward by Wallace et al.¹¹²: recombination sites or gene sequences themselves might be engineered by gene editing so as to, respectively, unlink harmful from favorable alleles or rid the genome of its mutational load. Population genetic theory predicts that clonally propagated plants accumulate deleterious variants as a lack of recombination makes selection against them less efficient^113,114. Exciting progress towards overcoming this predicament has been made in potato. Modern commercial varieties of the crop are autotetraploid clones that trace back to sexually reproducing diploid progenitors¹¹⁵. A move away from vegetative propagation is only part of an ambitious research program that aims at reinventing potato as hybrid seed crop^116,117 to make breeding faster and do away with the costly distribution of seed tubers to farmers. Genomics has helped remove some stumbling blocks. Long reads lend themselves to haplotype-resolved genome sequence assembly of diploid¹¹⁸ and tetraploid¹¹⁹ potatoes and their wild relatives¹²⁰. With such resources in hand, the deleterious mutation load can be quantified^121,122, rare recombinants between beneficial and harmful loci in tight linkage be found, and self-incompatibility be overcome¹²³. Similar approaches can be envisioned in other vegetatively propagated crops like other tubers or fruit trees¹²⁴. But genomics does not only help remake existing crops, it also lends a hand in domesticating new ones.

Regulatory and interactional variation

“Breeding 4.0” is the moniker given to a proposed breeding strategy in which editing “tens to hundreds of sites per generation” bestows upon breeders the “ability to combine any known alleles into optimal combinations”¹¹² (Table 1). Truly optimal combinations might not be needed to breed better crops, simply those that are good enough. But some combinations of alleles might not even meet that lower bar. The actions of alleles at different loci might interfere with each other in pernicious ways. “Gene pyramiding” is a breeding technique in which alleles are “stacked”, i.e. introgressed into the same genetic background, so as to confer beneficial traits such as durable resistance to one or more pathogens¹²⁵. Gene interference might prevent that process from working smoothly. Hurni et al. (ref ¹²⁶) describe an example in wheat. They combined transgenically two genes, Pm3 and Pm8, that confer resistance to isolates of powdery mildew. The transgenic plants became susceptible to varieties of the pathogen that Pm8 is effective against. The reason at the molecular level might be that the two genes sequester themselves away into ineffectual heterodimers. As more genes or alleles are stacked on top of each other, the risk of mutual interference between any two of them grows. Tools are needed to gauge the probability of that happening before embarking on the piling up of elaborate gene pyramids.

Gene-by-gene action is referred to by geneticists, molecular and otherwise, as epistasis, although different camps of scientists mean somewhat different things by the term¹²⁷. In a plant breeding context, negative epistasis describes the phenomenon that alleles at two different loci, each beneficial on its own, interfere through their interaction with crop improvement. A systematic inquiry into the phenomenon and a search for means to overcome it has begun in tomato. Soyk et al.¹²⁸ studied mutations in two of genes of that crop that control inflorescence architecture. Each on its own exerts benign effects on yield components, but a combination of the two leads to excessive branching¹²⁸. By combining suitable natural or induced alleles of both genes in a hybrid, it was possible to so modulate transcript abundance of both genes as to bypass negative epistasis¹²⁸. Tomato geneticists have uncovered more such epistatic interactions affecting the color, shape and size of fruits, some in a qualitative, some in a dosage-dependent manner¹²⁹. Tomato is but one example: gene-by-gene interactions dominated over single-loci effects in a quantitative genetic dissection of grain yield in hybrid wheat¹³⁰. Regulatory networks of flowering time genes in the model plant Arabidopsis thaliana are well studied^131,132. Genetic and transcriptomics studies have confirmed epistatic effects in the crops rice¹³³, barley¹³⁴ and maize¹³⁵. Domestication may have affected the wiring of gene regulatory networks: epistatic and environmental effects are weaker in foxtail millet than in its wild progenitor¹³⁶.

Agronomic traits are affected by differences in gene regulatory sequences as much as they are by variation in genes themselves and such regulatory variation can be harnessed for crop improvement. In maize, induced variation in the promoter of a gene that regulates meristem sizes altered ear morphology¹³⁷. Soyk et al.¹²⁹ proposed to use induced mutations in both genic and regulatory regions to dissect “interactional” variation in crops. Specifically, an allelic series of regulatory variants in one gene is crossed with a loss-of-function mutant in a putative interactor and the phenotypic effects quantified. This approach has already borne fruit and revealed dosage-dependency in epistatic interactions in tomato¹³⁸. Medical geneticists have introduced the omnigenic model, which posits that many, if not all, sequence variants in the genome have some, albeit weak, effect on any given trait¹³⁹. Large-scale genomics will play a role in testing that prediction in plants. Pangenomes, epigenome maps^140,141 and gene expression atlases^142,143 will help delineate regulatory sequences, catalogue variation in it, and understand its impact on gene expression (Figure 2). These resources and the research built on them add to the mechanistic understanding of crop performance that may empower future genetic engineers.

From genomes and phenotypes to crop models

One goal of crop translational genomics is to bridge the gap between advances in the science of genomics and the breeding of improved crop varieties. Phenotyping is an indispensable part of this endeavour: phenotypes have first to be observed experimentally in order to conduct genetic mapping. Any promising breeding line will have to undergo multi-environment field trials before it is released as a commercial variety. Genome-wide predictions and gene edits require phenotyping validation. Only through the synergistic development of genomic approaches and phenotyping methods can the full promise of crop translational genomics be realized. Methods for high-throughput phenotyping (HTP) or “phenomics” have been developed in the last two decades^144,145. Arguably most relevant to crop science is field-based HTP, which is mostly a derivative of remote sensing and targets the radiation reflected or emitted by the canopy. Field-based phenotyping can be implemented, for example, by unmanned aerial vehicles (“drones”)¹⁴⁶ and tractor-mounted cameras and sensors¹⁴⁷. Lab-based systems often combine automated image acquisition with growth in a controlled environment. There is a trade-off between the control over environmental parameters and the correlation between trial environments and real-life field conditions¹⁴⁸. The sophisticated simulation of field conditions¹⁴⁹ is one proposed solution to achieve biological and agricultural realism. HTP also encompasses molecular phenotyping such as metabolite profiling¹⁵⁰. Gene expression and epigenetic profiles can also be conceptualized as molecular phenotypes, especially when they are used in expression and DNA methylation QTL mapping^151,152. Yet, gene expression and epigenetic marks can feed back to the genome sequence in ways macro-phenotypes do not¹⁵³.

High-throughput sequencing, after twenty fledgling years, seems on the verge of settling down. HTP has yet to attain this level of technological maturity. Phenotyping methods differ in their levels of readiness¹⁵⁴. A wider deployment in practical breeding is hindered by high costs and technical challenges¹⁵⁵. Reynolds et al.¹⁵⁶ argued for taking “breeder friendliness” into account, a quaint resonance of which is found in Awada et al.’s inclusion in their agro-economical decision making model of “breeders’ aversion” towards the adoption of new technologies ¹⁵⁷. We should note that similar concerns pertain also to genome sequencing.

Community standards that harmonize data formats across phenotyping platforms remove avoidable obstacles. The acronym FAIR (Findable, Accessible, Interoperable, Reusable)¹⁵⁸ summarizes the principles that are guiding this effort. Subsumed under FAIR are MIAPPE (Minimum Information About a Plant Phenotyping Experiment)¹⁵⁹, a metadata standard for plant phenotyping and Breeding API (BrAPI)¹⁶⁰, an emerging set of protocols that make datasets and databases interoperable.

These commendable initiatives solve practical questions, but they do not reduce the size and complexity of the data, and possibly more importantly, of the scientific questions those data are destined to help answer. Some speak of a “data deluge”¹⁶¹. Crop science is not in a state of watery chaos, but a fresh look at how to make sense of more and more complex datasets is in order. One way forward is a more widespread adoption of mathematical modeling, which has a long tradition in plant science and agronomy^162,163. Mechanistic or process-based models, such as those of the primary metabolism, are informed by physical and biochemical principles. They underpin efforts to improve photosynthesis in a targeted manner¹⁶⁴. Statistical models, by contrast, fit equations to observed data to extrapolate trends. The abovementioned genomic prediction models, which integrate dense marker data and phenotypic data to help speed up the breeding cycle, are one example. A concern with this sort of models is that they are “black boxes” because they condense complex polygenic relationships into “breeding values” without answering the question why the model prioritizes certain genotypes¹⁶⁵. A balance has to be struck between parsimony – having as few predictor variables as possible – and biological realism, which includes the accurate representation of ecophysiological parameters¹⁶⁶. Modelling the complex dynamic processes that govern plant growth is one of 15 open research questions in plant cell biology¹⁶⁷.

Outlook

Translational genomics is likely to assist in the improvement of any crop in which it is attempted. Some priorities can be set by considering which applications might contribute most to safeguarding food supply and preventing further environmental damage. A report¹⁶⁸ published by the United Nations Food and Agriculture Organization (FAO) concedes that its analysis might give the impression “that all is well from the standpoint of potential for further production growth based on the use of existing varieties and technologies to increase yields.” The authors temper their optimism in two regards: (i) the “further spread of high external input technologies […] might aggravate related environmental problems” and (ii) yield growth will have to be engineered in the “often unfavourable agro-ecological and often also unfavourable socioeconomic environments of the countries where the additional demand will be.”

Let’s start with the second assertion. Economists agree that for the past two hundred years the world as a whole has avoided a Malthusian trap¹⁶⁹, a hypothetical scenario in which population growth outpaces the gains in agricultural productivity. This aggregate view hides the woefully inequitable state of global food security¹⁷⁰. Genomic research in tropical crops¹⁷¹ such as tef¹⁷², sesame¹⁷³, cassava^174,175 and others are welcome as is bioinformatics capacity building¹⁷⁶. Methods of proven efficacy, such as marker-assisted selection for resistance genes or genomic selection for yield and nutritional quality, should be adopted¹⁷⁷. Geneticists should double down on programmes to build up bioinformatics capacity¹⁷⁸ in low-income countries, especially those focused on crop plant research¹⁷⁹. The FAO report’s other caveat relates to the external costs of agriculture. The general public and regulatory bodies have become aware of the environmental impacts of intensive farming¹⁸⁰. In response to such concerns, or simply driven by climate change, changes in fertilization regimes, pest control or even the choice of crop might upend agricultural practices and alter fitness landscapes in cultivated plants. A precedent for rapid change is the rise of semi-dwarf cereals in the 20^th century after millennia-old constraints on nitrogen supply had been lifted. Hope for a second Green Revolution may be excessively optimistic, but at least with cheap sequencing, the opportunity costs of crop genomics are less likely to outweigh its benefits. Pangenomes and “pan-epigenomes” – atlases of gene expression, gene regulation and chromosome structure – will be among the foundational resources that will help geneticists, molecular biologist, physiologist and breeders to adapt our crops to a changing climate.