Meta-analysis of genome-wide association studies provides insights into genetic control of tomato flavor

Tomato flavor has changed over the course of long-term domestication and intensive breeding. To understand the genetic control of flavor, we report the meta-analysis of genome-wide association studies (GWAS) using 775 tomato accessions and 2,316,117 SNPs from three GWAS panels. We discover 305 significant associations for the contents of sugars, acids, amino acids, and flavor-related volatiles. We demonstrate that fruit citrate and malate contents have been impacted by selection during domestication and improvement, while sugar content has undergone less stringent selection. We suggest that it may be possible to significantly increase volatiles that positively contribute to consumer preferences while reducing unpleasant volatiles, by selection of the relevant allele combinations. Our results provide genetic insights into the influence of human selection on tomato flavor and demonstrate the benefits obtained from meta-analysis.

The manuscript could be improved for a broader audience I believe without too much additional effort from the authors. In present form, the manuscript is mainly of value to a narrow audience interested in tomato and flavour. They suggest the value of meta analysis in identifying additional QTLs but I would argue this is obvious in cases where sufficient additional variation and/or recombination is present and monitored in the additional genotypes as compared to each prior analysis. While a number of additional QTLs are identified, the reader learns little about why or how the combined data aided QTL identification. How many of the new loci are strong versus weak QTLs? Could they have been identified in individual populations by lowering thresholds? Can the authors make inferences about LD, numbers of accessions, diversity or other general metrics that could be used to guide meta analysis or primary GWAS assessment in other systems? A brief review of prior publications by these authors indicates strong genetic and genomics backgrounds and I believe they could present their data in a way where it had more general relevance to the broader genetics and breeding communities.
Line 47 -RF: Rothan in press……this needs to be updated. There is nothing in the reference list.
Lines 182 -184: I am confused by the use of the term "suffered" as this implies a negative connotation. I believe the authors are simply referring to possible causes of genetic diversity for citrate and their possibly underlying genes. I do not believe they intend to associate this diversity with content "suffering" in terms of low or high levels. Perhaps "was influenced" is better terminology?
Lines 258-260. This sentence is confusing. Are the authors saying cherry (small), large fruited and wild species are the "most promising candidates for high-flavor breeding purposes"? They seem to be suggesting as much, though what is left? I suspect the intended meaning here is different than conveyed. Please rewrite.
Overall the manuscript could benefit from editing for grammar, word usage, clarity and assurance of intended meaning.
Reviewer #2 wrote in the Remarks to Editor section.
(S)he thinks the using IMPUTE and filter after imputation by MAF, HWE, and INFO are appropriate. Meta-analysis strategy, i.e., using METAL as default and METASOFT under the presence of at least modearate heterogeneity, is also appropriate.
Reviewer #3 (Remarks to the Author): The present manuscript "Meta-analysis of genome-wide association studies provide new insights into genetic control of tomato flavor" definitely presents a new chapter in the quest to understand and apply genetic control of complex crop traits such as flavour. It combines different association studies in order to produce more information about genetic loci that possibly control tomato fruits flavour and will serve as a useful resource for researchers and breeders working in this area.
I believe it would help understanding the results of this study if a few issues listed below are addressed.
Meta-analyses in human studies commonly detect genetic variation that may explain very little proportion of phenotypic variation. In this study about 2/3 of the associations detected were new compared to those had been found previously. This is a considerable amount of new information, but what quantitative value these new loci would have in addition to already known loci? Are we not looking here at e.g. environmentally vulnerable small effects which now popped up just because the sample size and therefore the statistical power increased? Can this be estimated and discussed? The genetic structure of the combined data was described well and was dealt appropriately in the GWAS analysis. However, very little was said about phenotypic structure -the data came from different studies with different plant growth conditions, different harvesting and sampling procedures, different metabolic analysis protocols etc., which are all non-genetic sources of variation which could lead to significant phenotypic batch effects. Could the authors show the phenotypic variation just as they did for the genetic variation, e.g. using PCA, and to explain how they dealt with the batch effects?
It has been found by the authors that the flavour related chemical composition tends to improve with an increase of alternative alleles at the associated genetic loci, e.g. sugar content positively correlates with the number of alternative alleles of the associated genes. However, most of the alternative alleles in this case come from the berry-sized wild accessions, whose overall genetic background and therefore physiological background is quite different. It is, therefore, quite possible that all these alternative alleles and their combinations may lead to the chemical phenotype only when the overall background is "suitable". As it looks in Fig.2d, the sugar content mainly depends on fruits size. And indeed most of the loci mentioned are in close proximity to known loci for fruit size. Thus, is this then possible that we are simply looking at the fruit size effect?
Do Beta values in Supplemental Table S9 indicate size and directionality of the effects of alternative alleles? If so then counting the effects up of the alleles of the selected sugar genes for the two most extreme allele combinations in Fig.2e RAARA allele combination should give a higher sugar content than ARAAA combination, but in Fig. 2E they are the other way around. Also in the Discussion (p.12,l.247) it is stated that that the sugar content can be significantly improved almost all alternative alleles are selected, also combining the effects of these alleles does not really support this. How can these contradiction be explained? Or the data should be interpreted in a different way? Can the authors then provide more extensive guidance how to interpret different parameters in terms of their biological meaning? P9L174 The authors refer to cherry tomato on Fig 3c,d and also use this later in the text (Fig.4h) in comparison with "modern" tomato. Cherry tomato varieties has been introduced to the market since 1990th and therefore also should be considered as modern tomato. Since many wild cerasiforme accessions were used in this study, it would be better if the authors come up with an alternative term for what they call cherry tomato in order to avoid any confusion.

Reviewers' comments:
Reviewer #1 (Remarks to the Author): The manuscript could be improved for a broader audience I believe without too much additional effort from the authors. In present form, the manuscript is mainly of value to a narrow audience interested in tomato and flavour. They suggest the value of meta analysis in identifying additional QTLs but I would argue this is obvious in cases where sufficient additional variation and/or recombination is present and monitored in the additional genotypes as compared to each prior analysis. While a number of additional QTLs are identified, the reader learns little about why or how the combined data aided QTL identification. How many of the new loci are strong versus weak QTLs? Could they have been identified in individual populations by lowering thresholds? Can the authors make inferences about LD, numbers of accessions, diversity or other general metrics that could be used to guide meta analysis or primary GWAS assessment in other systems? A brief review of prior publications by these authors indicates strong genetic and genomics backgrounds and I believe they could present their data in a way where it had more general relevance to the broader genetics and breeding communities.
Reply: Many thanks for this comment. We would like to provide a point-to-point reply.
Q1: While a number of additional QTLs are identified, the reader learns little about why or how the combined data aided QTL identification. How many of the new loci are strong versus weak QTLs? Could they have been identified in individual populations by lowering thresholds?
Reply 1: We added a new paragraph in the discussion section to show readers the main advantages of meta-analysis of GWASs (Line 229-245). Briefly, we demonstrated the great potentials of meta-analysis of GWASs in other major crops. We discussed the main differences between mega-analysis (re-analysis of pooled data) and meta-analysis of GWASs. We then discussed the main benefits of meta-analysis. As for strong versus weak QTLs, some parameters derived from the meta-analysis should be detailed. For example, if we use P = 1.00×10 -09 as the genome-wide threshold of strong QTLs, then a total of 87 strong meta-QTLs were identified (Supplementary Table 9; Line 90-91). Among all the newly identified loci, most of them passed the suggestive thresholds (P = 10 -04 ) in at least one panel (Figure S13-75; Line 91-92).
Q2: Can the authors make inferences about LD, numbers of accessions, diversity or other general metrics that could be used to guide meta analysis or primary GWAS assessment in other systems?
Reply 2: We cited two publications which detailed protocols to perform meta-analysis of GWASs (Line 230-232). Most of the steps can be directly applied for other systems. METAL 1 (fixed-effect model) and METASOFT 2 (random-effect model) are the two most commonly used meta-analysis software 3 (Line 570-571). We explained in details about these two methods in the online methods (Line 571-577). We also provided a detailed script to perform meta-analysis along with all the raw meta-analysis data (See data availability Line 623-625; also see link at: https://data.inra.fr/privateurl.xhtml?token=a1b23652-4d2f-4a49-9b83-88a6236f9151). Q3: Line 47 -RF: Rothan in press……this needs to be updated. There is nothing in the reference list. Reply 3: The paper of Rothan was recently online and we have updated this (Line 48). Q4: Lines 182 -184: I am confused by the use of the term "suffered" as this implies a negative connotation. I believe the authors are simply referring to possible causes of genetic diversity for citrate and their possibly underlying genes. I do not believe they intend to associate this diversity with content "suffering" in terms of low or high levels. Perhaps "was influenced" is better terminology? Reply 4: Thanks for this concern. Following your suggestions, we replaced the term "suffered" to "was influenced" (See Line 189-190; 270; 281-282).
Q5: Lines 258-260. This sentence is confusing. Are the authors saying cherry (small), large fruited and wild species are the "most promising candidates for high-flavor breeding purposes"? They seem to be suggesting as much, though what is left? I suspect the intended meaning here is different than conveyed. Please rewrite.
Reply 5: Sorry for this misleading. In fact, the cherry tomato genomes are mosaic of big-fruited tomatoes and wild species genomes. We have modified this sentence in the revised manuscript (See Line 287-291). Q6: Overall the manuscript could benefit from editing for grammar, word usage, clarity and assurance of intended meaning. Reply 6: Many thanks. The two co-authors Harry Klee and Denise Tieman, from University of Florida, USA (both of them are native English speakers), have checked for these minor issues and a new careful reading was performed.
Reviewer #2 wrote in the Remarks to Editor section. (S)he thinks the using IMPUTE and filter after imputation by MAF, HWE, and INFO are appropriate. Meta-analysis strategy, i.e., using METAL as default and METASOFT under the presence of at least modearate heterogeneity, is also appropriate. Reply: Many thanks for your positive feedbacks.

Reviewer #3 (Remarks to the Author):
The present manuscript "Meta-analysis of genome-wide association studies provide new insights into genetic control of tomato flavor" definitely presents a new chapter in the quest to understand and apply genetic control of complex crop traits such as flavour. It combines different association studies in order to produce more information about genetic loci that possibly control tomato fruits flavour and will serve as a useful resource for researchers and breeders working in this area. I believe it would help understanding the results of this study if a few issues listed below are addressed.
Q1: Meta-analyses in human studies commonly detect genetic variation that may explain very little proportion of phenotypic variation. In this study about 2/3 of the associations detected were new compared to those had been found previously. This is a considerable amount of new information, but what quantitative value these new loci would have in addition to already known loci? Are we not looking here at e.g. environmentally vulnerable small effects which now popped up just because the sample size and therefore the statistical power increased? Can this be estimated and discussed?
Reply 1: Many thanks for these comments. In fact, we have tried to analyze how much new heritability or percentage of phenotypic variation can be explained based on summary statistics of these new identified loci. We have tried both the LD score regression approach (LDSC) 4 and Heritability Estimation from Summary Statistics (HESS) 5 . However, there are at least two crucial problems: substantially large population size (in humans, around or above 50,000 samples) and large high sequencing-depth reference panels. Based on the available data we have, the heritability estimated was not that trust-worthy. We agree that a large part of the new associations may result from the power of increased sample size and SNP imputation as stated earlier that among all the newly identified loci, most of them passed the suggestive thresholds (P = 10 -04 ) in at least one panel (Figure S13-75; Line 91-92).
For your concern about environmentally vulnerable small effects, in meta-analysis of GWASs, this has already been handled properly. In fact, environmental effect is just one cause of cross-study heterogeneity. Together with other causes, such as population structure, LD, phenotyping platforms, G×E interactions, were managed by using the heterogeneity I 2 . If I 2 is lower than 25%, it means the heterogeneity is quite small. If I 2 is higher than 25%, it means that is strong heterogeneity, which can be caused by many unknown factors, including environmentally vulnerable small effects. We properly handled this using the random-effect model as stated at Line 574-577.

Q2:
The genetic structure of the combined data was described well and was dealt appropriately in the GWAS analysis. However, very little was said about phenotypic structurethe data came from different studies with different plant growth conditions, different harvesting and sampling procedures, different metabolic analysis protocols etc., which are all non-genetic sources of variation which could lead to significant phenotypic batch effects. Could the authors show the phenotypic variation just as they did for the genetic variation, e.g. using PCA, and to explain how they dealt with the batch effects?
Reply 2: Thanks for this comment. For the phenotype data of each individual panel, they have been properly managed and normalized before GWASs 6-8 (see Materials and Methods of the 3 GWAS articles). All your concerns are crucial when you want to combine the phenotypes and genotypes of all panels and run a single GWAS using the combined data, which is technically referred as mega-analysis of GWASs. For mega-analysis, genotypes and phenotypes from all panels should be first combined and then analyzed, which requires proper management of the phenotypic structure. In contrast, meta-analysis has great benefits when pooling raw data of individual panels (mega-analysis) are impossible. In this study, we only had the summary data from each GWAS. This is discussed at Lines 229-245 in the discussion.
Q3: It has been found by the authors that the flavour related chemical composition tends to improve with an increase of alternative alleles at the associated genetic loci, e.g. sugar content positively correlates with the number of alternative alleles of the associated genes. However, most of the alternative alleles in this case come from the berry-sized wild accessions, whose overall genetic background and therefore physiological background is quite different. It is, therefore, quite possible that all these alternative alleles and their combinations may lead to the chemical phenotype only when the overall background is "suitable". As it looks in Fig.2d, the sugar content mainly depends on fruits size. And indeed most of the loci mentioned are in close proximity to known loci for fruit size. Thus, is this then possible that we are simply looking at the fruit size effect?
Reply 3: Fruit weight was influenced by strong domestication and improvement sweeps, which have been detailed in Lin et al. (2014) 9 . In their study, they provided a summary table of fruit weight loci selected during domestication and improvement (Supplementary Table 11 of Lin's paper; can also be downloaded at: https://www.nature.com/articles/ng.3117#supplementaryinformation). However, for all the 6 loci we analyzed in Figure 2, only the locus on chromosome 11 was influenced by a domestication sweep (DW149). In Lin's study, all the three main QTLs (fw11.1, fw11.2, fw11.3) were influenced by improvement sweeps, instead of domestication sweeps. For the remaining 5 loci, they were not influenced by any domestication and improvement sweeps. So, we conclude that though sugar content has a strong negative correlation with fruit weight, it is more likely for the 6 loci associated with both fructose and glucose, were not strongly selected or were loosely linked to fruit weight QTLs (Line 269-272). By the way, we identified more loci for these two sugars, some of which being close to fruit weight QTLs but we did not analyze them. Table S9 indicate size and directionality of the effects of alternative alleles? If so then counting the effects up of the alleles of the selected sugar genes for the two most extreme allele combinations in Fig.2e RAARA allele combination should give a higher sugar content than ARAAA combination, but in Fig. 2E they are the other way around. Also in the Discussion (p.12,l.247) it is stated that the sugar content can be significantly improved almost all alternative alleles are selected, also combining the effects of these alleles does not really support this. How can these contradiction be explained? Or the data should be interpreted in a different way? Can the authors then provide more extensive guidance how to interpret different parameters in terms of their biological meaning? Reply 4: If the genotypes of each individual panel were aligned with the reference genome and the first allele in the genotype files were always the alternative allele, then the beta values indeed would indicate the size and directionality of the effects of alternative alleles. However, we did not do this and the directions in Supplementary Table S9 also indicated that the first allele in the genotype files of different panels was not always the reference genome. As for your main concerns about the allele combinations, we would like to provide more detailed explanations. 1) The allele combination analysis was only based on panel T as explained in Online Materials Lines 604-621). We did not combine three panels because the sugar content had strong crossstudy heterogeneity and we could not simply combine the panels. Panel T had the largest population size among the three panels and we therefore only focused on panel T. 2) As panel T was genotyped by resequencing, we did not perform genotype imputation for this panel. Before we analyze the relationship between sugar contents and allele combinations, we had to remove accessions with too much missing genotype calls. Thus for some allele combinations, the number of individuals was quite small, 4) For some allele combinations, including RAARA, RRRAR, ARRRR and AARRR, for which there was only one observation and we could not perform statistical tests as explained in Online Materials Line 620-621. Q5: P9L174 The authors refer to cherry tomato on Fig 3c,d and also use this later in the text (Fig.4h) in comparison with "modern" tomato. Cherry tomato varieties has been introduced to the market since 1990th and therefore also should be considered as modern tomato. Since many wild cerasiforme accessions were used in this study, it would be better if the authors come up with an alternative term for what they call cherry tomato in order to avoid any confusion.

Q4: Do Beta values in Supplemental
Reply 5: Though cherry tomato varieties have been introduced to the market since 1990s, they are admixture of modern big-fruited tomatoes and wild species 9,10 . So, cherry tomatoes are not really modern tomatoes. We have clarified this at Lines 287-291.