Synthesis of descriptive sensory attributes and hedonic rankings of dried persimmon (Diospyros kaki sp.)

Abstract This work aimed to characterize the sensory attributes of hot air‐dried persimmon (Diospyros kaki) chips, correlate these attributes with consumer hedonic information, and, by doing so, present recommendations for cultivars that are most suitable for hot‐air drying. A trained sensory panel evaluated dried persimmon samples (representing 40 cultivars) for flavor, taste/aftertaste, and texture. In addition, in each of two tests conducted in different years, more than 100 consumers provided hedonic evaluations of 21 unique samples in a ranking task with a balanced incomplete block design. A partial least squares regression model correlating the mean hedonic ranking to the trained panel data was developed using the data from the first consumer panel. The predictions from the model were correlated with the second panel to verify the model. It was found that including taste, aftertaste, and texture data (but not specific flavor attribute data) produced a predictive model (Spearman's ρ=0.83). This indicates that flavor is likely secondary to taste and texture in dried persimmon chips. Using the validated predictive model, 6 of the 40 persimmon cultivars tested are recommended for a dried chip product; these cultivars are ‘Fuyu’, ‘Lycopersicon’, ‘Maekawa Jiro’, ‘Nishimura Wase’, ‘Tishihtzu’, and ‘Yotsumizo’.

deastringency process known as "mellowing" or "bletting" in order to be palatable in raw form. A common astringent persimmon cultivar grown in the United States is 'Hachiya'.
In addition to their consumption in fresh form, some astringent persimmon cultivars are amenable to drying into "hoshi-gaki"-a confectionary delicacy with its origins in East Asia. Hoshi-gaki are prepared by tying ripe astringent persimmons on a string and allowing them to dry outdoors for 2-4 weeks; hand kneading of the drying fruit is sometimes performed to facilitate even moisture distribution. The resulting product (30%-50% moisture content) has a texture similar to that of jelly candy and a naturally-occurring powdery sugar coating; the astringency is also completely removed by the drying process (Sugiura & Taira, 2009).
Persimmons are a rich source of Vitamin C, carotenoids, and polyphenolic compounds. In vivo and in vitro studies of these dietary components suggest a relevant role of this fruit in protection against free radicals and prevention of some human diseases (Giordani, Doumett, Nin, & del Bubba, 2011). The overall aim of this study is to encourage more consumption of persimmons. Developing a dried chip-style product provides persimmon growers an option for preserving and marketing their fruit without using the (labor-and time-intensive) hoshi-gaki process. Dried apple chips are an analogous product that has seen widespread distribution, and the sensory properties of this product have been well-characterized (Konopacka & Plocharski, 2007;Sham, Scaman, & Durance, 2001;Velickova, Winkelhausen, & Kuzmanova, 2014).
However, evaluation of the dried products by consumers was not reported. Also, each of these studies involved persimmons collected at a single point during the harvest season; it is possible that the quality of the dried products would have been different for earlyand late-harvest source fruit.
Thus, the purpose of the present work was to assess the suitability of 40 cultivars of persimmon (harvested at multiple time points and from multiple sources, when possible) for hot-air drying into a chipstyle product. Assessments of the taste/aftertaste, flavor, and texture of the dried products were obtained from a trained sensory panel, and these results were correlated with the hedonic rankings of the products by 150 consumers in each of 2 years. The challenges of the large sample set and the timing of the consumer panels in the middle of the harvest season were addressed by the methods of a balanced incomplete block design and predictive partial least squares regression model, respectively.

| Persimmon samples
Fifty-four fresh persimmon samples, consisting of ~200 fruit each, were harvested in Fall 2015 and dried for this study ( . These sources are denoted as R, C-1, C-2, C-3 in Table 1. For additional detail, the accession numbers for the NCGR samples and California counties of origin for the commercial samples are listed in Supplemental Table S1. For some cultivars and sources, there was enough fruit available to collect multiple samples throughout the season; in these cases, the sample harvests were spaced apart by a minimum of 12 days. Persimmons were hand-harvested when commercial ripe-that is, when the exterior color had changed from green to yellowish-green, yellow, orange, or reddish-orange (cultivar dependent). Persimmons were packed directly into boxes with plastic liners that separated and cushioned each fruit. Within 24 hr of harvest, the boxes of fruit were transported to the USDA-ARS laboratory in Albany, CA, USA via pickup truck (Source R) or overnight commercial shipping (Sources C-1, C-2, and C-3).

| Drying method
Upon receipt, the persimmon samples were hand-sorted to remove visibly damaged fruit and then stored for an average of 8 days.
Following best practices for this commodity (Crisosto, 1999), nonastringent cultivars were stored in an incubator set at 18°C, and astringent and variant cultivars were stored in a refrigerator set at 2°C.
On the day of processing, persimmons were washed in tap water to remove surface soil and then sanitized in a 200 ppm chlorine solution. Slices of 5 mm thickness were cut with a commercial meat slicer (Model 1612P, Hobart, Troy, OH, USA). If present, seeds and seed fragments were left in the slices, since practice runs revealed that the seeds were easier to remove after drying than before drying. The slices were arranged in a single layer on the trays of a commercial dehydrator (Model 2924T, Excalibur Dehydrator, Sacramento, CA, USA) and dried at 52°C (125°F) for 18 hr. The dried slices were stored at ambient temperature in sealed metallized polyester film pouches.
Before-and after-drying photos of a typical persimmon sample are shown in Figure 1.
T A B L E 1 Sources, astringency types, number of harvests, and consumer evaluation status of the persimmon cultivars in this study

| Trained panel sensory evaluation
Descriptive sensory analysis was conducted on the 54 samples of   (Bourne, 2002;King et al., 2012); the attributes for flavor and astringency were benchtop tested and experimentally determined. Definitions for the texture attributes can be found in Table   S3. Panelists were considered adequately trained when they could correctly identify all 21 flavor attributes on the first try and correctly identify and rank 4 different concentrations of each of the five taste and aftertaste attributes. Additionally, the panelists practiced the testing procedure and use of the scales twice prior to evaluating the persimmon samples. The panelists received no compensation other than snacks at the end of each session.
Evaluation of the dried persimmon samples took place in isolated booths. The panelists were given one whole slice plus one "wedge" (1/4 to 1/3 of a slice) to assess the eight texture attributes. They were then presented with an additional half slice with which to evaluate the individual flavor attributes, taste, aftertaste, and overall flavor intensity. The panelists were instructed to taste the skin and flesh of each sample and expectorate all samples. Panelists received 12 products per session, constituting four samples evaluated in triplicate. Products were randomized and presented in black soufflé cups labeled with three-digit random codes. The panelists were instructed to cleanse their palates between samples with filtered water and unsalted water crackers (Carr's, Carlisle, UK)-a palate cleansing approach recommended for high-astringency foods like wine (Ross, Hinken, & Weller, 2007). To reduce fatigue, no more than 12 samples were evaluated in any given session, and all sessions lasted a maximum of 1 hr.
The texture attributes were rated on a 15 cm unstructured line scale with specific product anchors throughout the scale. The taste, astringency, and overall flavor intensity attributes were also rated on a 15 cm unstructured line scale, but using only "low" and "high" at

Astringency type Source
Yeddo "+" for the two consumer panel columns indicates that the sample was included in the indicated consumer hedonic ranking portion of the study. (All samples were evaluated by the trained sensory panel). In those same columns, "+" with no modifier indicates that the first harvest of the cultivar was used; the modifiers "1st," "2nd," and "3rd" indicate cases where multiple harvests were used. Astringency Type: A, astringent; N, nonastringent; V, variant. Source: R, research plot (National Clonal Germplasm Repository, Davis, CA, USA); C-X, commercial source X.
a There was not a sufficient amount of cultivar 'Brazzale' available for either consumer panel. However, this cultivar was evaluated by the trained panel. b Evidence from this study suggested that cultivar 'Yeddo' is a nonastringent cultivar; this differs from the classification of "variant" given by other publications (Camp & Mowry, 1929;Ryerson, 1927).
This method was chosen since it was thought that the persimmon flavors would be more readily assessed as "present" or "absent," versus having the intensity of these attributes indicated on a scale. Although, under this method, an individual panelist marks only the presence or absence of a specific attribute for a specific sample, aggregating repeated CATA assessments across multiple panelists leads to relative intensity information (Campo et al., 2010)-for example, an attribute that is selected 90% of the time is clearly more intense than an attribute that is selected only 5% of the time.

| Consumer tests
At the time of the first consumer test (early November 2015), there were only 25 dried persimmon samples available. The remaining samples had not yet been harvested. As it was not feasible for each consumer to evaluate all 25 samples, a balanced incomplete block design was used. This type of design has been used for other food products when it is desirable to reduce the assessment load on consumers (Bower & Whitten, 2000). The balanced incomplete block design used was (v=25, b=30, k=5, r=6, λ=1) where v is the number of products, b is the number of panelists in a block, k is the number of products each panelist evaluates, r is the number of times each sample appears across all blocks, and λ is the number of times each pair of samples appears across all blocks. For the consumer tasting, five full replicates (150 total panelists) of the balanced incomplete blocked design were prepared, and the presentation order for each block was randomized.
The same experiment design and number of samples (25)  For CT1, the consumer group was 58% female/42% male and ranged in age from 10 to 70. In terms of fresh persimmon consumption during the September-to-December harvest season, the group was nearly evenly divided among the six frequency categories offered-from "never" (16%) to "daily" (12%). For dried fruit consumption, the majority (57%) of the group consumed dried fruit between 1-3 times/month and 3-5 times/week. The CT2 consumer group had similar distributions of age, gender, and dried fruit consumption to that of the CT1 group. In terms of fresh persimmon consumption frequency, however, the CT2 group had a more concentrated subset in the "1-3 times/ month" category (27%) compared to the earlier group.
During the test, each consumer panelist was given a bag containing their five samples and the order in which they were to taste the It was known beforehand that the attendees at the event would vary widely in age and product evaluation experience, and ranking can be performed even by consumers who are unfamiliar with product rating scales. In addition, ranking using preprinted stickers helped simplify the data collection and reduced the risk of panelists' failing to evaluate 1 or more samples out of the set of 5 (a particularly important issue for a balanced incomplete block design). However, this method had inherent limitations. It captured consumer preference information but not necessarily consumer acceptance information, since the task was a forced ranking of five samples, whether all the samples were wellliked, disliked, or somewhere in between. To capture both preference and liking information, the task was altered for CT2. In the CT2 scoresheet, 15 possible sticker-placement spaces were distributed evenly along a line anchored with "Dislike" at the far left, "Neither Dislike Nor Like" at the center, and "Like" at the far right. Thus, panelists were still forced to rank the five samples, but a liking rating of 1 ("Dislike") to 15 ("Like") was obtained simultaneously. In short, CT1 was comprised solely of a ranking task while CT2 was comprised of a combined ranking/rating task.
For CT1, the 25 samples were, by necessity, the first harvests of 25 cultivars that were available at the time of the test. For CT2, however, all 54 samples were available (and had been evaluated by the trained panel earlier in 2016-a timeline of the study is given as Supplemental Figure S1). So, a subset of 25 samples was chosen for CT2 based on the following criteria: • cultivars that had not been evaluated by consumers in CT1 (16 samples) • a sample of cultivar 'Fuyu' from a commercial source that had not been evaluated by consumers in CT1 (one sample) • "anchor" cultivars known to be low-, medium-, and high-preference from the results of CT1 (three samples) • Second harvests (when available) of the "anchor" cultivars (two samples) • First, second, and third harvests of cultivar 'Mishirasu'-a cultivar whose first harvest was low-preference in CT1 but whose preference at third harvest was predicted to be high (see Section 3.4 for additional details) (three samples) The samples used in CT1 and CT2 are indicated in Table 1 with "+" symbol.

| Statistical analyses
Trained panel sensory data were summarized using multi factor analysis (MFA) and hierarchical clustering on principal components (HCPC).
The MFA and HCPC were carried out using the FactoMineR package in R (Husson, Josse, Le, & Mazet, 2015) with the flavor, taste, and texture data being considered different groups of variables. The total count for each flavor attribute for each product was used as the response and was considered frequency data. The mean ratings for the taste and texture data were treated as continuous data and were scaled before analysis. Based on visual examination of the scree plot, the first five components from the MFA were used for the HCPC. The number of clusters for the HCPC was determined using the default method for FactoMineR. By this method, a hierarchical tree is built.
The sums of the within-cluster inertia are then calculated for each partition. The suggested partition is the one with the higher relative loss of inertia (i [clusters n+1]/i [cluster n]) (Husson et al., 2015).
The consumer ranking data from CT1 were initially analyzed using Durbin's test with a least significant difference (LSD) test to determine which samples were different at p<.05 (Conover, 1999).
The trained panel descriptive data and the CT1 hedonic data were correlated using partial least squares regression (PLSR) using the pls package in R (Mevik, Wehrens, & Liland, 2015). The PLSR model correlated the mean rank of the dried persimmon samples to the scaled mean values for all texture and taste data and the scaled frequency for all flavor data except for the "chocolate" and "coconut" attributes, which were removed as they had a frequency count of 0 for the products tested. The model used the first two components of the PLSR; this number of components was chosen to minimize the root mean square error of prediction (RMSEP). In addition, a second PLSR model was constructed using the trained panel attributes without the specific flavor attributes; this sparse model used only the first component of the PLSR, again minimizing RMSEP. Both models were applied to both CT1 and CT2 data and validated using leave-oneout cross validation. All graphs were constructed using the R package ggplot2 (Wickham, 2016); all other statistics were conducted in R (R Core Team 2015).

| MFA on trained panel sensory data
The MFA with HCPC identified three sensory clusters (Figure 2). Figure 2a shows the products on the first two dimensions of the MFA representing 41% of the total variance in the dataset; sample cluster is denoted by both shape and color of the point. Figure 2b shows the corresponding placement of the sensory attributes for the MFA, with the darkness of the text of the attribute representing how well the biplot represents the attribute (cosine squared). The HCPC used the first five dimensions representing 63% of the variance of the dataset; as this is considerably larger than the first two dimensions, the Figure 2c visually plots all attributes that differ between a cluster at p<.05.
The three clusters identified primarily differed by attributes associated with ripeness and texture. Cluster 3 is the smallest cluster (n=7) and has traits most related to unripe fruit-for example, vegetal/green and grassy flavors and astringent aftertaste. Cluster 2 is the next largest cluster (n=21) and is primarily classified by its negative textural attributes (e.g., hardness, roughness, toughness) and lack of strong flavors. Cluster 1 is the largest cluster (n=26) and is primarily classified by characteristics most related to ripe fruit-for example, "stone fruit/ peach," "floral/citrus," and "toffee" flavors.

| Durbin test separation on hedonic data from CT1
The ranking data from CT1 were analyzed using the Durbin test with LSD separation; the results are summarized in
It is possible to use the combined characterization and hedonic data presented in Figures 3 and 4 to predict which persimmon cultivars would yield a dried chip-style product that would be preferred by consumers, even if a particular cultivar was not represented in the set of 25 samples presented in CT1. In Figure 3, there are 17 samples in Cluster 1 (the best-preferred cluster) that were not evaluated by the consumer group in CT1 (depicted by unfilled square data points).
While all of these represented potentially preferred dried products, some were from the second or third harvests of cultivars whose first and/or second harvests yielded dried samples in Clusters 2 and 3. Put another way, a persimmon cultivar may yield a preferred dried product, but only from later in the harvest season. If a dried product were made from fruit harvested commercial ripe, but early in the season, that dried product would not be preferred by consumers. Thus, the more useful subset of the 17 consumer-untested samples in Cluster 1 is comprised of cultivars whose trained panel MFA score placed them into Cluster 1 at the first harvest. There are 10 cultivars in this subset; they are listed below and highlighted on the MFA biplot in Supplemental Figure   S3. In alphabetical order (no ranking information implied), the 10 cultivars that were not tested by consumers in Consumer Test 1 [CT1] but whose trained panel multi-factor analysis [MFA] score placed them into Cluster 1 at the first harvest are 'Akoumanzaki', 'Fennio', 'Fuyu Imoto', 'Fuyu Jiro', 'Giant Fuyu', 'Gofu', 'Hachiya', 'Maru', 'Suruga', and 'Tishihtzu'. It should be noted that all of the commercial samples examined in this study fell into Cluster 1 (regardless of whether they were tasted during CT1). This indicates that persimmon cultivars currently on the Higher value for Average Rank corresponds to more liking of the dried product (highest possible rank would be 5.00; lowest possible rank would be 1.00). Astringency Type: A, astringent; N, nonastringent; V, variant. Source: R, research plot (National Clonal Germplasm Repository, Davis, CA, USA); C-X, commercial source X.
market are either known to be or predicted to be a good starting material for production of a dried chip-style product. 'Fuyu' persimmons from source C-3 were in Cluster 1 and preferred in dried form by the consumer group. It is encouraging to see the closely-related cultivars 'Fuyu Imoto', 'Fuyu Jiro', and 'Giant Fuyu' appear in the list.
In the same pattern, 'Jiro' and 'Maekawa Jiro' chips from the research F I G U R E 4 Predicted mean rankings vs. measured mean rankings of the samples for the full partial least squares regression (PLSR) model with all trained panel attributes included (a) and the sparse PLSR model, which excluded specific flavor attributes (b). The four samples which were common to both years are represented with filled data points source were preferred by the CT1 consumer group, and the closelyrelated cultivar 'Fuyu Jiro' from a commercial source appears in the list. So, there is some consistency in the cultivar groups that appear in Cluster 1.

| Confirming predictions using data from CT2
While the sample set was limited (due to harvest timing) for CT1 in November 2015, the entire sample set was available for CT2 in November 2016. The set of 25 samples for the latter test included 16 cultivars that had not been evaluated in the former test. Of these, nine cultivars had been confirmed to be preferred by consumers (the "a" group at top of Table 2), and 10 cultivars had been predicted to be preferred by consumers. The Durbin test separation results for CT2 are given in Table 3, along with the mean ratings. (Recall that both ranking and rating data were gathered in CT2, while only ranking data were gathered in CT1.) In Table 3, the 10 predicted-to-be-preferred-at-first-harvest cultivars are emphasized in bold font, and the high-('Nishimura Wase'), medium-('Fujiwaragosho', first harvest), and low-preference ('Ichikeijiko', first harvest) "anchor" samples are underlined.
In CT2, all 10 of the predicted-to-be-preferred-at-first-harvest cultivars from CT1 had a mean rating above 7.5 (possible range of 1-15), indicating that these cultivars were indeed preferred and liked by consumers. However, some cultivars in this set fared better than  Higher value for Average Rank corresponds to more liking of the dried product (highest possible rank would be 5.00; lowest possible rank would be 1.00).
Higher value for Average Rating corresponds to more liking of the dried product (highest possible rating would be 15.00; lowest possible rating would be 0.00). Tukey's least significant difference results (lower case lettering) are based on ranking data. The predicted-to-be-preferred-at-first-harvest cultivars (based on CT1 data) are given in bold font. The "anchor" samples (based on CT1 data) are underlined. Cultivars for which multiple harvests were tested are indicated with the "1st," "2nd," or "3rd" modifier. Astringency Type: A, astringent; N, nonastringent; V, variant.
In a more general way, the performance of the PLSR model can be assessed for all 25 samples presented to the consumers in each year.  Figure S4. In brief, the attributes of "astringency," "crispness," "skin toughness," "fibrousness," "hardness," "bitterness," and "sourness" are seen to be strong negative drivers of preference, while "moistness," (overall) "flavor intensity," and "sweetness" positively drive preference.
Using this refined PLSR model, the initially-identified-from-

| Trajectory of cultivars over multiple harvests
In this study, some persimmon cultivars were harvested 2 or 3 times throughout the season. Figure 5 depicts the MFA biplot with all samples still shown at the same coordinates as in Figures 2a and 3  Indeed, from the results of CT2 shown in Table 3, the mean rankings for 'Mishirasu' first and second harvest are at the very bottom of the list, while the mean ranking for the third harvest is at the very top.
Despite the resulting conclusion to only use later-harvest fruit for drying, this practice must be weighed against the possible decreased ease of slicing as the fruit get riper and softer. In this study, some of the latest-harvest samples were nearly impossible to slice (even with a commercial meat slicer) because they were too soft. Thus, we recommend that persimmon growers use cultivars that yield consumerpreferred dried products as soon as the fruit are commercial ripe (and still firm).
It should be noted that not all cultivars followed the pattern of clearly increasing in preference as the harvest season progressed.
For the low-and medium-preference anchors ('Ichikeijiko' and 'Fujiwaragosho', respectively), consumer preferences for the two harvests were not statistically different. It is possible that a third harvest of these cultivars would have yielded a more highly-preferred product, but the issue of over-ripeness hindering slicability would again be a concern.
The harvest-timing effect is likely the cause of the conspicuously lower-than-expected ranking of 'Fuyu' fruit from Source R; see Table 2, where this sample had an average ranking of 2.79. This was a mediocre ranking, especially in comparison to that of the same cultivar from Source C-3 (ranking=3.63, statistically tied with the topranking sample in CT1). Although both 'Fuyu' samples-and all other samples in the study-were harvested when the fruit were commercial ripe, these two particular samples were clearly at different maturity levels at harvest, as evidenced by the greater amount of green skin color of the fruit from Source R. This is shown in the top two panels of Supplemental Figure S5. Indeed, the second harvest of 'Fuyu' from Source R, shown in the bottom panel of the figure, has much more even orange color. Although this sample was not part of the set tested in CT2, it is one of the samples projected to be in Cluster 1 in the MFA biplot ( Figure 5), versus the Source R first harvest sample, which was in Cluster 2. So, 'Fuyu' remains a recommended cultivar for hot-air drying, with the stipulation that-beyond commercial maturity-even orange skin color should be a prerequisite for 'Fuyu' fruit selected for this process.

| CONCLUSIONS
The astringency type (astringent, nonastringent, variant) did not appear to inherently predict whether the dried chips made from a given persimmon cultivar would be preferred by consumers, since examples of all astringency types could be found throughout the ranking lists in both years. Thus, this attribute should not be used to screen persimmon cultivars for their suitability for hot-air drying. Regarding harvest timing, the general trajectory of the samples over the harvest season was into (or further into) the most-preferred cluster. However, we recommend using cultivars that are harvest-date-independent in their liking.
Comparison of the full and sparse PLSR models indicates that flavor is likely secondary to taste and texture in dried persimmon chips. Based on the sparse model, the six persimmon cultivars most suited for hot-air drying (for fruit harvested commercial ripe at any time during the season) are the following: 'Fuyu', 'Lycopersicon', 'Maekawa Jiro', 'Nishimura Wase', 'Tishihtzu', and 'Yotsumizo'. This list includes cultivars that are already established in the U.S. market as well as cultivars that have not yet seen widespread commercial propagation.