Chemical and genetic basis of orange flavor

Abstract

Sweet orange (Citrus sinensis) exhibits limited genetic diversity and high susceptibility to Huanglongbing (HLB). Breeding HLB-tolerant orange-like hybrids is in dire need. However, our understanding of the key compounds responsible for orange flavor and their genetic regulation remains elusive. Evaluating 179 juice samples, including oranges, mandarins, Poncirus trifoliata, and hybrids, distinct volatile compositions were found. A random forest model predicted untrained samples with 78% accuracy and identified 26 compounds crucial for orange flavor. Notably, seven esters differentiated orange from mandarin flavor. Cluster analysis showed six esters with shared genetic control. Differential gene expression analysis identified C. sinensis alcohol acyltransferase 1 (CsAAT1) responsible for ester production in orange. Its activity was validated through overexpression assays. Phylogeny revealed the functional allele was inherited from pummelo. A SNP-based DNA marker in the coding region accurately predicted phenotypes. This study enhances our understanding of orange flavor compounds and their biosynthetic pathways and expands breeding options for orange-like cultivars.

Esters, produced by high-expressed CsAAT1 allele, were key compounds differentiating orange flavor from mandarin flavor.

INTRODUCTION

Citrus production in the United States has been devastated by Huanglongbing (HLB) or citrus greening disease, especially in Florida where historically citrus production was 90% sweet orange [Citrus sinensis (L.) Osbeck], a citrus type highly susceptible to Candidatus Liberibacter asiaticus, the bacterium considered responsible for the disease. In the 2022–2023 season, Florida citrus production had the lowest output in nine decades, taking acreage down to ~50% and production to ~10% of levels before the arrival of the disease (1). Thus far, there is no effective way to control HLB (2). By 2017, the cumulative economic loss due to HLB was estimated to be $6 billion (3). The most sustainable solution to maintain orange production is to plant cultivars that are resistant or tolerant to HLB.

Now, only a narrow range of citrus cultivars are classified as sweet orange, C. sinensis, which are all derived from cumulative somatic mutations derived from a presumed single complex interspecific introgression hybrid of mandarin (Citrus reticulata Blanco.) and pummelo [Citrus maxima (Burm.) Merr.] (4, 5). Therefore, the limited genetic diversity of sweet orange has rendered it difficult to find a resistant source within the species. Breeding new cultivars via outcrossing with HLB-tolerant materials provides the best breeding solution. Poncirus trifoliata (L.), conferring tolerance to HLB, has been used for citrus breeding. It is challenging to improve the fruit quality of hybrids harboring Poncirus introgressions, because fruits of the parent P. trifoliata have unacceptable flavor, and the hybrids with Citrus present various levels of off-flavor even after a few generations of backcrossing to elite citrus cultivars (6, 7). However, after generations of backcrosses to mandarin, new generations of orange-like hybrids are available, such as “US SunDragon” included in this study, which has P. trifoliata in its pedigree and low P. trifoliata off-taste (3). US SunDragon exhibits tolerance to HLB and a high resemblance to orange flavor and, thus, is a promising new variety to sustain the orange juice industry in Florida.

One key to accelerating the development of orange-like hybrids and enhancing the acceptance of using non–C. sinensis in orange juice is to define fruit components important for the distinctive flavor and aroma characteristics in orange. Studies before 2005 focused on the identification of abundant and odor-active compounds in orange fruit or juice samples alone. The fruity aroma esters ethyl 2-methylpropanoate, ethyl butanoate and ethyl 2-methylbutanoate, 3a,4,5,7a-tetrahydro-3,6-dimethyl-2(3H)-benzofuranone (“wine lactone”), and the grassy aroma (Z)-hex-3-enal were found to be the most potent volatiles in Valencia and Navel oranges, based on aroma extract dilution analysis and odor activity values (8, 9). Recently, efforts shifted to comparing aroma profiles between orange and other citrus species to find alternatives for HLB-sensitive C. sinensis. Using the same techniques but comparing “Valencia” orange to a mandarin cultivar, a recent study identified ethyl butanoate, ethyl 2-methylbutanoate, octanal, decanal, and acetaldehyde as the most odor-active volatiles in orange (10). Nevertheless, all abovementioned studies are of small scale, consisting of three cultivars at most. A new chemical and sensory evaluation, testing a diverse collection of accessions within and across citrus species, was desired to validate and expand on the previous results.

Woody perennial plants like citrus have a long juvenile phase and require a long and expensive evaluation process before releasing new hybrid cultivars. To date, no aroma genes have been directly discovered and validated via traditional forward-genetic approaches. On the other hand, using reverse-genetic approaches, various terpene synthases (TPSs) in citrus were identified and functionally validated (11, 12). TPSs with different enzymatic activities catalyze production of monoterpenes and sesquiterpenes found in citrus fruit. CsTPS1, discovered in C. sinensis, is responsible for valencene production in orange (11). A natural InDel in the promoter region of CsTPS1 was associated with deficiency of valencene production in some mandarin cultivars (13). Two transcription factors, CitERF71 and CitAP2.10, regulate E-geraniol and valencene production by interaction with CitTPS16 and CsTPS1, respectively (14, 15). Nevertheless, genetic studies on citrus fruit flavor are limited to the terpene biosynthesis pathway. No information about other aroma biosynthetic genes or molecular markers in citrus is available.

Here, we report a comprehensive chemical and flavor evaluation over 179 harvest/accession combinations, including sweet orange (C. sinensis), mandarin (C. reticulata), poncirus (P. trifoliata), and citrus hybrids with poncirus introgressions (hereafter, poncirus hybrids) with the objective to predict orange flavor using chemical data and to deduce the most important and key chemical compounds contributing to orange flavor. Integration of chemical clustering and transcriptome analyses facilitated the identification of a C. sinensis alcohol acyltransferase 1 (CsAAT1), catalyzing the biosynthesis of both straight- and branched-chain esters contributing to orange flavor. Its biological function was validated in citrus fruits using overexpression transient assays. Last, a DNA marker based on an exonic single-nucleotide polymorphism (SNP) was developed and tested.

RESULTS

Volatilome and flavor quality of mandarin (C. reticulata), sweet orange (C. sinensis), P. trifoliata, and poncirus introgressed hybrids

A total of 60 volatiles, consisting of five alcohols, eight aldehydes, 12 esters, two ketones, 14 monoterpenes, one terpene ester, one terpene ketone, four terpene alcohols, and 13 sesquiterpenes (table S5) were identified and quantified for 198 harvest/accession combinations (data S1). On the basis of principal components analysis (PCA) of the volatilome, clear separation was observed among poncirus (P. trifoliata), sweet orange (C. sinensis), and mandarin (C. reticulata) (Fig. 1A). However, poncirus hybrids exhibited a wider distribution, overlapping with all three species on the first two principal components (PCs) (Fig. 1B), reflecting their admixture genetics. Some poncirus hybrids showed similar volatile profiles to those of orange, revealed by their proximity to orange on the first two PCs (Fig. 1B) and clustering close to four orange cultivars in the hierarchical clustering analysis (Fig. 1C). In general, compared to orange, two poncirus accessions were higher in ketones, terpene ester, sesquiterpenes, and terpene alcohols but lower in alcohols, aldehydes, esters, and terpene ketone (Fig. 2A). After generations of modified backcrossing, poncirus hybrids recovered levels similar to orange, for a majority of volatiles, except for esters (Fig. 2A). Specifically, significantly higher concentrations of methyl butanoate (P = 1.6 × 10⁻¹³, Bonferroni corrected), ethyl butanoate (P = 2.5 × 10⁻¹⁰), ethyl hexanoate (P = 1.3 × 10⁻⁵), ethyl 3-hydroxyhexanoate (P = 3.9 × 10⁻¹⁸), and ethyl octanoate (P = 4.1 × 10⁻¹⁰) were detected in orange than in poncirus hybrids (table S6). In addition to chemical measurements, juice samples were also evaluated for orange and mandarin flavor attributes by a trained panel. Similar to volatile data, sensory scores of orange flavor for poncirus hybrids exhibited large variation, ranging from 1.1 to 8.1 (Fig. 2B and data S2). One hybrid, US SunDragon, scored consistently high for orange flavor (orange score > average orange score for orange cultivars, 6.4) across 11 of the 14 harvests (Fig. 2B), while its volatile profile was not the closest to orange (Fig. 1C). These results indicate that volatiles found in orange did not equally contribute to the orange flavor, with some having more importance than others.

Fig. 1. Clustering of volatile abundances among citrus accessions.

(A) Biplot of first two principal components (PCs) based on volatile data. Individual hybrids are colored according to their breeding class. Samples of hybrids, predominately mandarin (C. reticulata), sweet orange (C. sinensis), and P. trifoliata, are included in this principal components analysis (PCA). (B) Poncirus introgressed hybrids were added to PCA. (C) Hierarchical clustering of all accessions, constructed using absolute distance matrix of volatile data.

Fig. 2. Sensory and chemical differences across different classes of citrus accessions.

(A) Comparisons of total volume of volatile classes among different species. Y axis represents log₁₀ peak area. (B) Mandarin flavor scores regressed against orange flavor scores. Samples of US SunDragon are colored in light green.

Prediction of orange flavor using machine learning algorithms

Unlike mandarin flavor [coefficient of determination (R²) = 0.21; fig. S1], orange flavor (R² < 0.01; fig. S2) was poorly predicted by soluble solid concentration (SSC)/titratable acidity (TA). Therefore, predicting orange flavor entails an integration of sugars, acids, bitter compounds, and volatiles. Here, we tested six popular statistical models (16, 17) for predictions of orange, mandarin, and relative orange (orange score − mandarin score) flavors, based on measurements of SSC, TA, pH, limonin, nomilin, and 60 volatiles. To test the prediction accuracy of our models in real-world scenarios, all samples (n = 19) collected in the 2021–2022 season were not included in the training but only used for testing. The random forest (RF) model consistently performed the best across three flavor attributes and two testing schemes (Fig. 3), with R² of 0.67, 0.74, and 0.78 for respective orange, mandarin, and relative orange flavors for the 2021–2022 samples (Fig. 3B). The high predictive ability of the RF model gave us high confidence that RF model could be used to identify important chemical compounds and predict orange flavor in future seasons.

Fig. 3. Prediction of orange flavor using chemical data.

Six prediction models [Bayesian linear model (Bayesian), general linear mixed model (GLM), least absolute shrinkage and selection operator (LASSO) regression, partial least squares (PLS), random forest (RF), and extreme gradient boosting decision tree (XGBoost)] used to predict orange, mandarin, and relative orange flavors and tested with the 10-fold cross-validation (A); true validation dataset which includes 19 samples collected from season 2021–2022 (B). Y axis is the R² value.

Important compounds for predicting orange flavor

Among SSC, TA, pH, limonin, nomilin, and 60 volatiles, RF models identified 26 chemical compounds, important in prediction of relative orange and orange flavors. The important compounds included eight esters, two aliphatic alcohols, two aliphatic aldehydes, one aliphatic ketone, seven terpenes or terpene derivatives, both limonoids, SSC, TA, pH, and SSC/TA (Table 1). Both bitter limonoids, limonin and nomilin, exhibit negative effects on relative orange flavor, while both positive and negative effects were observed for terpenes and terpene derivatives. Two alcohols (1-pentanol and 1-hexanol) and two aldehydes [(E)-2-pentenal and (E)-2-hexenal] showed positive relationships with relative orange flavor. 1-Octen-3-one, which was reported to be the most odor-active aliphatic ketone in orange juice (18, 19), showed a high positive effect on relative orange flavor. A negative effect of SSC/TA was observed for relative orange flavor, consistent with the observation of lower SSC/TA in orange compared to that in mandarin in previous studies (20). Seven esters (methyl hexanoate, ethyl hexanoate, ethyl 3-hydroxyhexanoate, ethyl octanoate, methyl butanoate, ethyl butanoate, and ethyl 2-methylbutanoate), all with positive effect on orange flavor, were the only compounds shared between the prediction model for orange flavor scores and the classification model for orange (table S7). The total level of seven esters explained 21% of the total variation in relative orange flavor (fig. S3). A specific hybrid with P. trifoliata in its background, US SunDragon, had high orange flavor and high total esters (figs. S3 and S4). On the other hand, esters were negative predictors for mandarin flavor, yet only three contributors—SSC/TA, (E)-2-pentenal, and SSC—showed positive effects on mandarin flavor. In summary, orange flavor requires high amounts of esters, which are the key and essential compounds, as well as several alcohols and aldehydes, but low limonoids, lower SSC/TA than mandarin, and a balanced combination of terpenes.

Table 1. Important compounds [variable importance in projection (VIP) > 6] for prediction of orange and mandarin flavor in citrus fruit, and their relative intensity (relative orange = orange − mandarin).

P, positive; N, negative.

Compound	Class	Relative orange			Orange			Mandarin
		VIP	Effect	Rank	VIP	Effect	Rank	VIP	Effect	Rank
Methyl hexanoate*	Ester	63.87	P	1
Ethyl hexanoate*	Ester	34.01	P	2				10.33	N	5
pH		27.83	N	3	7.64	N	13
1-Pentanol	Alcohol	24.54	P	4
Limonin	Limonoid	24.24	N	5	8.89	N	9
SSC/TA		21.82	N	6	10.50	P	8	31.32	P	2
1-Octen-3-one	Ketone	17.88	P	7
α-Cubebene	Sesquiterpene	16.86	P	8	35.85	P	2
(E)-2-Pentenal	Aldehyde	13.18	P	9				8.07	P	6
(E)-2-Hexenal	Aldehyde	12.96	P	10	8.65	P	10
Ethyl 3-hydroxyhexanoate*	Ester	11.92	P	11				55.38	N	1
Ethyl octanoate*	Ester	10.57	P	12
TA		10.23	P	13				7.80	N	7
Nomilin	Limonoid	8.19	N	14	61.73	N	1
α-Terpinyl acetate	Terpene ester	7.62	P	15
Methyl butanoate*	Ester	7.30	P	16
Ethyl butanoate*	Ester	7.08	P	17
Ethyl 2-methylbutanoate*	Ester	6.79	P	18
SSC					10.99	P	5	14.86	P	3
β-Bourbonene	Sesquiterpene							13.90	N	4
Carvone	Terpene ketone				18.41	P	3	7.64	N	8
Caryophyllene	Sesquiterpene							7.36	N	9
β-Elemene	Sesquiterpene							6.11	N	10
Ethyl propanoate	Ester				11.75	P	4
Humulene	Sesquiterpene				10.84	N	6
Terpinen-4-ol	Terpene alcohol				10.78	N	7
1-Hexanol	Alcohol				8.47	P	11
(E)-Alloocimene	Monoterpene				7.67	N	12
δ-Cadinene	Sesquiterpene				6.67	P	14

*Overlap with important predictors for classification as orange based on pedigree data.

Flavor compounds and their related transcriptome networks in Citrus

Chemical clustering showed that 50 of the 66 analyzed compounds were clustered into nine groups (Fig. 4), mostly consistent with their chemical classes. There were two groups for sesquiterpenes (ST1 and ST2), two for monoterpenes (MT1 and MT2), one for alcohol (ALC1), two for aldehydes (ALD1 and ALD2), and two for esters (E1 and E2). Compounds within a cluster might share the same genetic control and indicate existing natural genetic variation across different accessions. To identify gene co-expression modules that jointly altered with chemical groups, the weighted correlation network analysis (WGCNA; fig. S5) was built using RNA sequencing (RNA-seq) data from juice samples of 11 accessions with different volatile profiles. Six gene co-expression modules (fig. S6A and data S3) were correlated to chemical clusters, including E2 to MEivory [correlation coefficient (r) = 0.77], E1 to MEdarkolivegreen (r = 0.82), TA to MEindianred4 (r = 0.67), SSC to MEbrown4 (r = 0.80), ALD1 to MEplum2 (r = 0.87), and ST2 to MEfirebrick4 (r = 0.82, only observed in dendrogram with individual compounds; fig. S6B). CsTPS1 (Cs4g_pb015710) was within the co-expression module MEfirebrick4, found positively correlated with concentrations of five sesquiterpenes (valencene, γ-muurolene, α-selinene, β-elemene, and humulene), consistent with the previous functional characterization (11). The identification of CsTPS1 validated the effectiveness of WGCNA to discover metabolite biosynthetic genes. Otherwise, the SSC-related module MEbrown4 contained 26 genes, enriched for galactose metabolic process (GO:0006012). The ALD1-related module MEplum2 contained a D-isomer specific 2-hydroxyacid dehydrogenase (Cs1g_pb014450), a fatty acid desaturase (Cs6g_pb001810), and an alpha/beta hydrolase (Cs9g_pb018430), all potentially involved in fatty acid metabolic process. Last, a GDSL lipase/esterase like gene (Cs3g_pb021240) in the module MEivory was related with E2 concentrations.

Fig. 4. Heatmap of correlations among chemical abundances.

Each cell is colored according to Pearson correlation coefficient (−1 to 1). Hierarchical clustering (HCA) is shown on the left side of the plot. Nine groups (right side of the plot) are assigned on the basis of chemical class and HCA. The letters in group names represent chemical classes: sesquiterpenes (ST), monoterpenes (MT), alcohols (ALC), aldehydes (ALD), and esters (E).

Discovery of CsAAT1 responsible for ester biosynthesis in orange

Using chemical clustering, five of the seven key esters for orange flavor were clustered into the E1 group, and methyl hexanoate showed a high correlation with the E1 group despite not being clustered in E1 (Fig. 4). To identify candidate genes responsible for ester production in orange, we compared gene expression between five ester producers (all esters in the E1 group) and six low producers (table S2). After filtering with false discovery rate adjusted P < 0.05 and fold change > 2, only 93 up-regulated genes (0.2% of total genes) were identified. Among them, a previously undescribed alcohol acyltransferase Cs6g_pb013840.1 (CsAAT1) was the only one annotated as acetyltransferase and showed fourfold increase of expression in ester producers than in low producers (Fig. 5A), with a correlation coefficient of 0.74 with the first eigenvector of E1 cluster (fig. S7). The increased expression of CsAAT1 in orange (Wilcoxon, P = 0.055; fig. S8) and the high production of esters (fig. S11) were confirmed using quantitative polymerase chain reaction (qPCR) and gas chromatography–mass spectrometry (GC-MS) with five additional orange cultivars, compared to low-ester-producer mandarins and hybrids. The expression of CsAAT1 was ripening-induced, as an increase of expression was observed along the ripening process, while no expression was observed in roots, leaves, or fruits 45 days after flowering (fig. S9) (21). A close examination of synteny regions from multiple genome assemblies revealed that the CsAAT1 and its flanking regions were incomplete in the C. sinensis genome V2 (Cs6g_pb013840) (22) and complete but not annotated in the genome V3 (4). After reannotation with RNA-seq data and homolog search, a tandem array of two AATs was identified, with an interspace of 8593 base pairs (bp) (Fig. 5D). However, only CsAAT1 was expressed across samples, while CsAAT1t (the tandem homolog of CsAAT1) had close to zero expression (Fig. 5D). Sanger sequencing of cDNA confirmed that CsAAT1 was 1347-bp long, with only one exon, encoding a polypeptide of 449–amino acid residues and a predicted molecular mass of 50.56 kDa. It contained characteristic HXXXDG (residues 155 to 160) and DFGWG (residues 378 to 382) motifs of the BAHD protein family (fig. S10) (23). In a maximum likelihood tree with 23 diverse acetyltransferases, CsAAT1 was clustered within the clade including three strawberry alcohol acetyltransferases (FvAAT, FcAAT1, and FvAAT) and RhAAT identified in Rosa hybrida (Fig. 5B), with sequence similarity between 39.3 and 40.9% to four AATs. The sequence similarity to strawberry AATs explained the resemblance of ester profile to cultivated strawberry (Fragaria × ananassa) as all esters in orange except for ethyl 3-hydroxyhexanoate could be detected in strawberry (16, 24, 25). Alignment among 10 CsAAT1 orthologs from the genus Citrus including both alleles of CsAAT1 revealed 96.78 to 99.78% sequence identity and a unique insertion of tyrosine at residue 305 only in C. sinensis and C. maxima (data S4). A gene tree indicated the high-expressed allele in sweet orange was inherited from C. maxima, whereas the low-expressed allele was from C. reticulata (Fig. 5C), consistent with the evolution path to C. sinensis (5). Between alleles from C. sinensis and C. maxima “Majiayou,” there were only three non-synonymous SNPs (fig. S10). Among the five C. maxima cultivars that we tested, one cultivar, named “Large Pink,” showed a high level of ester production comparable to that of orange cultivars, supporting our hypothesis of the origin of the orange CsAAT1 allele (fig. S11). However, within C. maxima, the phenotypic variation was large, which is consistent with the allelic diversity observed in the gene tree.

Fig. 5. Identification and validation of CsAAT1.

(A) Volcano plot of log₂ fold change against −log₁₀ P value comparing ester producers to low-producers. CsAAT1 is highlighted with a large dot. NS, not significant. (B) Maximum likelihood tree of 24 acyltransferases with 100 fast bootstrapping. CsAAT1 is underlined. Branches are annotated with bootstrapping values. The scale represents branch length. (C) A gene tree of CsAAT1 orthologs in the genus of Citrus. CsAAT1_A1 and CsAAT1_A2 are high-expressed and low-expressed alleles in Citrus sinensis. Branches are labeled with FastTree support values. AAT1 from P. trifoliata is used as an outgroup. The scale represents branch length. (D) Comparison of expression of CsAAT1 and its tandem replicate CsAAT1t. Y axis represents fragments per kilobase of transcript per million mapped reads (FPKM). (E) Comparisons of production for six esters between control fruits and fruits agroinfiltrated with CsAAT1 overexpression construct across four cultivars. M indicates mandarin cultivar. O indicates orange cultivar. Asterisks "*" indicate P values of <0.05 based on Student’s t tests. Y axis represents volatile abundance (peak area). (F) Concentrations of six esters with zero, one, and two dosages of the orange allele of CsAAT1 across 27 accessions. The number of samples in each category is provided. P values from Kruskal-Wallis tests are provided. Y axis represents volatile abundance (peak area).

Functional validation and DNA marker for CsAAT1

The function of CsAAT1 was validated via transient overexpression assays in three mandarin cultivars (ester non- or low-producer) at the fully ripe stage and one orange cultivar at the small green stage (1 to 2 cm in diameter). Across mandarin cultivars, fruits agroinfiltrated with the overexpression construct of CsAAT1 had significantly increased production of methyl butanoate, methyl hexanoate, ethyl butanoate, ethyl hexanoate, and ethyl octanoate (Fig. 5E and table S8). The highest concentrations were found in three ethyl esters, likely facilitated by a high volume of free substrate, ethanol, in the ripe citrus fruits (data S1). Although there was no increase of ethyl 3-hydroxyhexanoate in the overexpressed fruits, significant increases of the other four branched-chain esters were observed, especially ethyl trans-2-butenoate (fig. S12), suggesting that CsAAT1 was also capable of synthesizing branched-chain esters. On the other hand, transient overexpression of CsAAT1 in young green fruits of orange resulted in higher production of ethyl hexanoate, methyl hexanoate, and ethyl octanoate (Fig. 5E). The difference in ester profile comparing to overexpressed mandarin may be related to changing substrate availability during fruit development. A high-resolution melting (HRM) marker was developed to target 1122A->C. It showed 100% sensitivity and 85.1% accuracy to predict the production of six esters (Fig. 5F, fig. S13, and table S4). All accessions with zero dose of the CsAAT1 orange allele (n = 11) produced no esters in fruit. Four accessions with heterozygous calls produced low levels of esters (table S4), indicating potential additional genetic controls for ester production.

DISCUSSION

Unique odor of different citrus types is the result of a distinct composition of volatiles. Despite characteristic compounds found in some citrus such as grapefruit (26), no consensus of key compounds for orange has been reached. It was generally accepted that a combination of 20 to 30 volatiles is needed to imitate orange flavor (18). However, this hypothesis was challenged by other studies. Using aroma reconstitution experiments, a mix of 14 aroma-active compounds in their actual concentrations highly resembled orange juice reconstituted from concentrate (19). Feng et al. (10) cut the list down to five key compounds that were ethyl butanoate, ethyl 2-methylbutanoate, octanal, decanal, and acetaldehyde. Reactive aldehydes such as acetaldehyde, (Z)-hex-3-enal, neral, and geranial, imparting a pleasant green and citrus note to fresh squeezed orange juices, are present at supra threshold levels in fresh squeezed oranges yet diminish in processed orange juice (18). The two aldehydes, (E)-2-pentenal and (E)-2-hexenal, found to contribute to orange flavor in our work, fit into this category. However, concentrations of straight-chain aldehydes such as octanal, nonanal, and decanal were not found important for orange flavor in our study because both decanal and octanal are two major aldehydes in other citrus species such as mandarins (27–29). Our prediction models, incorporating the largest sensory and chemical dataset of citrus accessions to date, identified seven esters, including the most odor active compounds in orange juice, ethyl 2-methylbutanoate and ethyl butanoate (9, 10), as key compounds for orange flavor. Consistent with these results, orange juice made from HLB affected orange fruits, with a lower concentration of esters, was often described as lacking orange flavor (30).

Because citrus types produce a large diversity and quantity of terpenes, genetic research has focused on the discovery of genes for terpene biosynthesis up to now. Some characterized terpene biosynthetic genes include limonene synthase (31), linalool synthase (32), γ-terpinene synthase (33), sabinene synthase (34), and valencene synthase (11). Less is known about genetic controls for other aroma-active compounds in citrus such as esters. Alcohol acyltransferase (AAT) is a member of the BAHD family and is able to accept a range of alcohol and acyl–coenzyme A (CoA) substrates to form esters (23). AATs have been functionally characterized in a variety of fruit crops, such as strawberry (35), melon (36), and banana (37). Our phylogeny revealed that CsAAT1 belonged to Clade III according to previous classification (23), together with other AATs. More specifically, CsAAT1 is most phylogenetically related to three AATs in strawberry. Strawberry AATs can use short- and medium-chain, branched, and aromatic acyl-CoA and alcohol molecules as substrates (25, 35) to produce a variety of straight- and branched- chain esters including six key esters found in orange. The function prediction of CsAAT1 via phylogenetic analysis is congruent with results from fruit overexpression assays and marker tests. Ester biosynthesis is also regulated by the availability of substrates, which may depend on the activity of various catabolic pathways (e.g., lipid breakdown) (35), which could explain our observation of an increased amount of hexanoic acid esters, but not butanoic acid esters in small green overexpressed oranges. A previous study has shown that orange was distinguished from other citrus species by the lowest levels of oleic and palmitoleic and by the highest levels of linolenic and arachidic acids (38). Future studies are needed to explore genetic variation in the early steps of ester biosynthesis among citrus germplasm, as well as fatty acid availability during fruit development.

Current juice regulations limit adding fruits from mandarin hybrids not classified as C. sinensis to 10% to maintain the “orange juice” label, leaving the citrus juice industry vulnerable to disease epidemics such as HLB when growing oranges in a monoculture (3). Our findings show that some poncirus-introgressed hybrids such as US Sundragon and FF11061 are organoleptically similar to orange, are more HLB-tolerant than orange, and can potentially be used in orange juice, widening selections that can substitute for traditional orange cultivars. Using a large collection of accessions from multiple citrus types, we determined esters were the key compounds for perceiving orange flavor and differentiating orange from other citrus types. We discovered a previously undescribed alcohol acyltransferase, CsAAT1, which catalyzes the production of both straight- and branched chain esters in orange fruit. Our work will greatly accelerate the breeding of orange-like hybrids. With the aid of the DNA marker for orange flavor, seedlings can be screened at an early stage, long before they produce fruits, and orange flavor can be more rapidly recovered with fewer generations of backcrossing.

MATERIALS AND METHODS

Fruit sampling

Fruits were sampled from mature trees grown at the A.H. Whitmore Citrus Research Foundation Farm, Groveland, FL (28.687502, −81.886090) or at the US Department of Agriculture (USDA)/Agricultural Research Service (ARS) Research Farm, Fort Pierce, FL (27.433215, −80.427057). Trees were self-rooted for hybrids or grafted on rootstocks for named cultivars. Hybrids were selected on the basis of tree health and general quality, spanning a large range in shapes, colors, seed numbers, and flavors, especially during the first season (2016–2017) (39). Hybrids that produced fruit that were too sour or bitter in the first 2 years were eliminated from further observations. Hybrids of interest (healthy trees and acceptable flavor) were harvested multiple times at 2- to 4-week intervals to determine optimum maturity. Fruits were then taken to the USDA/ARS laboratory in Fort Pierce, FL, washed, sanitized, and manually juiced using a reamer-type juicer (Oster Model 3183, Household Appliance Sales and Service, Niles, IL, USA; or Vinci Hand Free Juicer, Vinci Housewares, La Mirada, CA, USA). Fruits were split into four batches of equal number of pieces to account for four biological replications. Aliquots were taken for measurements of SSC, TA, limonin, nomilin, and volatiles. A minimum of 1 L was used for sensory evaluation. All juice samples were stored at −20°C until analysis. A total of 198 harvest/accession combinations from 59 accessions were analyzed for volatile production and/or evaluated for orange and mandarin flavors, with 179 samples from 57 accessions having both evaluations. Information about sample ID, harvest time, harvest location, and pedigree for individual samples is shown in data S2 and table S1.

Flavor evaluation

The project involves human subjects and is exempted on the basis of 45 CFR 46.104 (d) (6): 46 “Protection of Human Subjects”; 104 (d) “Exempt Research”; (6) “Taste and food quality evaluation and consumer acceptance studies” (i) “if wholesome foods without additives are consumed” or (ii) “if a food is consumed that contains a food ingredient at or below the level and for a use found to be safe, or agricultural chemical or environmental contaminant at or below the level found to be safe, by the Food and Drug Administration or approved by the Environmental Protection Agency or the Food Safety and Inspection Service of the US Department of Agriculture.”

Most panelists had been trained and practiced descriptive sensory evaluation of citrus juices for over 10 years. In this study, 10 to 12 of these panelists rated citrus hybrids for orange and mandarin flavors using a linear scale with numeral anchors from 0 to 15 and word anchors at 1 = “low,” 7.5 = “medium,” and 15 = “high.” A reference standard was provided for each of these descriptors at each session. Unpasteurized orange juice locally produced (Al’s Family Farm, Fort Pierce, FL) and tangerine juice [“gourmet pasteurized” from Natalie’s Orchid Island Juice Company (Fort Pierce, FL)] were the references for orange and mandarin flavor, respectively. Orange and mandarin juice standards were given an intensity value of 12 on the scale of 0 to 15, as they were considered the representative flavor for those descriptors. Panelists met each year for five or six sessions to practice and refresh their memories about descriptors. They tasted samples that were evaluated from the previous year and that represented typical orange and mandarin flavors and then discussed their evaluations. The panel leader gathered and averaged their ratings from the discussion and entered the values using the Feedback Calibration Method (FCM) feature in Compusense Cloud (Compusense, Guelph, ON, Canada). Panelists then returned to the tasting booths and evaluated the same samples presented in coded cups, using the FCM method. Samples were served as 45 ml in 118-ml (4-oz) plastic soufflé cups (Solo Cups Co., Urbana, IL) at 14°C, in a randomized order across panelists. Panelists tasted up to 40 samples per season, with no more than four samples per day, repeated in two daily sessions.

Volatile identification and quantification

Volatiles were analyzed using a headspace–solid-phase microextraction (SPME)–gas chromatography–MS system as previously reported (40). Briefly, 6 ml of juice was sealed in 20-ml vials and stored at −20°C until analysis. The juice sample was incubated at 40°C for 30 min, and, then, a 2-cm tri-phase SPME fiber (50/30 μm DVB/Carboxen/ PDMS, Supelco, Bellefonte, PA, USA) was inserted to the headspace to collect and concentrate volatiles for 30 min. The SPME fiber was then inserted into the injector of an Agilent 7890 GC coupled with a 5975 MS detector (Agilent Technologies, Palo Alto, CA, USA) for 15 min at 250°C. The column was a DB-5 [60 m–by–0.25 mm inside diameter, 1.00-μm film thickness, J&W Scientific, Folsom, CA, USA). Mass units were monitored from 30 to 250 mass/charge ratio (m/z) and ionized at 70 eV. Volatile identification and quantification of peak areas were conducted with MassHunter Workstation software (version 10.0; Agilent Technologies). Initial identification was done by mass spectra searches with the NIST library (version 14, match score > 0.9). The identification was then confirmed by comparing the retention indices generated by running standard C6-C17 alkane mixture under the same conditions as the samples with online resources (NIST Chemistry WebBook and Flavornet.org). Each sample had four replicates.

SSC, TA, and limonoid quantification

Juice was centrifuged, and SSC, TA, and limonoids were measured using the supernatant. SSC was determined by a digital refractometer (Atago RX-5000cx, Tokyo, Japan). TA was measured by titration of 10 ml of supernatant with 0.1 mM sodium hydroxide (NaOH) to the final pH of 8.1 using a titrator (Dosino model 800, Metrohm, Herisau, Switzerland). Quantification of limonin and nomilin was performed via liquid chromatography–tandem MS (LC-MS/MS) with a 1290 Infinity II UPLC coupled with a 6470 triple quadrupole MS (Agilent, Santa Clara, CA, USA). The MS was operated in multiple reaction monitoring (MRM) mode, and nomilin was detected with a precursor ion with m/z of 515.3 and a product ion with m/z of 411.2 with fragmentor and collision energy voltages set to 135 and 14 V, respectively. Limonin was detected with a precursor ion with m/z of 471.2 and a product ion with m/z of 425.2 with fragmentor and collision energy voltages set to 135 and 19 V, respectively. Quantification of limonoids was performed by integrating the area under the chromatographic peak and calculating the amount of each compound based on standard curves (R² ≥ 0.99 with range 0.006 to 25 mg liter⁻¹). Each sample had four replicates that were averaged.

Prediction models and statistical analyses

Model training and testing were conducted using tidymodels package in R (41). A total of 160 samples evaluated from 2016 to 2021 was used to train models. Seven prediction models including generalized linear model, least absolute shrinkage and selection operator model, Bayesian model, partial least square model (PLS), XGBoost model, and RF model were used to predict scores for orange, relative orange (orange score − mandarin score), and mandarin flavors. Two hyperparameters were tuned for the PLS model including the maximum proportion of original predictors and the number of PLS components to retain. Three hyperparameters were tuned for XGBoost and RF including the number of predictors that randomly sampled at each split, the number of trees contained in the ensemble, and the minimum number of data points in a node required for the node to be split further. Tuning for hyperparameters was conducted using a grid searching method, and the best values were determined using 10-fold cross-validation. To evaluate the performance of the final models, we used two testing datasets: a 10-fold cross-validation set and a true validation set including 19 samples collected in the 2021–2022 season. Because RF performed the best across both validation sets, it was used to infer important compounds for prediction of flavor attributes. Variable importance in projection (VIP) scores were extracted from RF models using VIP package in R. Compounds with VIP > 6 were identified as important compounds. PCA and hierarchical clustering of chemical compounds were constructed with absolute distance matrix using prcomp and dist functions in R.

RNA extraction, sequencing, and alignment

RNA was extracted from fruit juice samples of 11 accessions (tables S1 and S2) having distinct volatile profiles. RNeasy Plant Mini kits (QIAGEN Co., Hilden, Germany) were used for extraction. Illumina (San Diego, CA, USA) 150-bp pair-end sequencing was performed on the Illumina NovoSeq platform by Novogene Co. (Sacramento, CA, USA). On average, 7.1-Gb data were obtained for each sample. Reads were aligned to sweet orange genome V2 (22) using STAR aligner with the default parameters (42). Reads for each transcript were counted using htseq-count (43) in Union mode excluding nonunique reads.

WGCNA and differential gene expression analysis

Weighted correlation network for gene expression was conducted using WGCNA package in R (44). Transcript counts for 11 accessions were converted into fragments per kilobase of transcript per million mapped reads (FPKM). The input transcripts were filtered with total FPKM across 11 samples > 1 and coefficient of variation > 0.58; a total of 4092 transcripts remained for the analysis. The value of softPower was set to 8, chosen by the “pickSoftThreshold” function. The minimum module size was set to 10, and deepsplit parameter was set to 2. The rest of the parameters in WGCNA were set to default. To find relationships between co-expression modules and chemical compounds, two correlation matrices were built. The first one included the first eigenvectors of all co-expression modules and abundances of all chemical compounds; the second one consisted of the first eigenvectors of all co-expression modules and chemical clusters. The distance matrix was computed as 1 − correlation matrix. Hierarchical clustering analysis was performed to determine relationships between co-expression modules and chemical compounds. Differentially expressed genes (DEGs) comparing between ester producers (n = 5) and low producers (n = 6) were identified using DESeq2 package in R (45). The input list of transcripts was pre-filtered using only those with total counts > 10. The final list of DEGs was determined with FDR adjusted P < 0.05 and fold change > 2.

Quantitative real-time PCR, Sanger sequencing, gene tree, and HRM marker

Quantitative real-time PCR was conducted for 15 samples including five additional sweet orange cultivars (table S1). These sweet orange cultivars were not included in RNA-seq or statistical modeling for flavor. The experimental conditions followed previous study (46) except that annealing temperature was set to 56°C. Primer pairs were designed for CsAAT1 (CsAAT1_1,087F: GCCTTGAAATTTTCCAGTTGGG, CsAAT1_1,175R: TCTCCAAAAATGCCGCTCCA) and CsGADPH (CsGADPH_F: GGAAGGTCAAGATCGGAATCAA, CsGADPH_R: CGTCCCTCTGCAAGATGACTCT) genes. Three technical replicates were tested for each sample. Relative expression was computed using 2^−ΔΔCT method (47). Total RNA of Valencia orange juice was used to synthesize cDNA. To amplify the full length of CsAAT1 cDNA, two primers (CsAAT1_1F: ATGGAAATTGGCATTGTCTCAAGAG, CsAAT1_1,340R: TCAGAATCGACAAACGTCAAACAATA) were designed according to the curated coding region of CsAAT1 on the basis of evidence of genome sequences of SWOv2 and SWOv3 (4) and RNA-seq alignment. cDNA of CsAAT1 was amplified with touchdown PCR, and a gel confirmed the size of amplicon. Sanger sequencing for unpurified PCR product was conducted at Azenta Inc. using two primers (CsAAT1_1,340R, CsAAT1_658R: TGTACTTGGCTTCTCTGAACCA). The sequencing result confirmed the CDS sequence of CsAAT1 [National Center for Biotechnology Information (NCBI) accession: OR003937]. To build a gene tree of acyltransferases including CsAAT1, protein sequences of 23 acyltransferases including alcohol acyltransferases involved in the biosynthesis of esters from strawberry (35), melon (36), banana (37), and apple (48) were downloaded from NCBI (table S3). Protein sequences were aligned using Clustal Omega V1.2.2. A maximum likelihood tree with 100 fast bootstrapping was built with GAMMA GTR model using RAxML. Sequences of CsAAT1 orthologs in genus Citrus were obtained through blasting CsAAT1 to genome assemblies (21). Two alleles of CsAAT1 were retrieved from the phased assembly of C. sinensis (49). The approximately maximum-likelihood gene tree was constructed using Fasttree (50). A HRM marker (CsAAT1_1,087F, CsAAT1_1,175R) was designed to target the SNP 1122A->C. PCR and melting curve genotyping was conducted according to previous studies (46). The HRM marker was tested for 27 accessions that had volatile data from multiple harvests (table S4). Sensitivity was calculated as $\frac{\text{True positive}}{\text{True positive}+\text{False negative}}$ , and accuracy was measured as $\frac{\text{True positive}+\text{True negative}}{\text{Total number of samples}}$.

Transient fruit overexpression assay

The full-length cDNA of CsAAT1 flanked by attB sequences was synthesized by Twist Bioscience Inc. (San Francisco, CA, USA). The target sequence was first introduced into pDONR/Zeo vector (Thermo Fisher Scientific Inc. Waltham, MA, USA) via Gateway cloning and then inserted into overexpression vector pMDC32 (51). The empty vector pMDC32 and overexpression vector pMDC32::CsAAT1 were transformed into Agrobacterium tumefaciens strain EHA105. Transformation to citrus fruits was performed according to Shen et al. (15). Fruits from three commercial mandarin brands (“Peelz,” “Cuties,” and “Halos”) were purchased at a local grocery store, and small green fruits of “Hamlin” orange (1 to 2 cm in diameter), in which esters should not be produced based on tissue-specific expression results (fig. S9), were harvested from the research farm. Two treatments were performed on separate fruits. Five biological replicates with one fruit per repetition were performed for each treatment. Four injections with 3 ml of Agrobacterium suspension each were made for mandarin fruits. For orange fruits, the injection stopped when the whole fruit was wet. Pulp tissue (whole fruit for orange) was collected after 7 days of incubation in dark at room temperature. All samples were analyzed with the GC-MS under the aforementioned conditions.

Supplementary Materials

This PDF file includes:

Figs. S1 to S13

Tables S1 to S8

Legends for data S1 to S4

Other Supplementary Material for this manuscript includes the following:

Data S1 to S4