Introduction

Flavor, texture, and appearance are all vital aspects of our eating experience, which have become central focuses in modern fruit breeding programs (Klee & Tieman, 2018). Breeding for these sensory aspects not only impacts consumer acceptance, but also affects other segments of the chain by increasing fruit demand, marketability, and grower profitability. Despite its significance, progress in this field has been slow, primarily due to the complexity of identifying key fruit-quality targets that align with consumer preference. Considered as multi-sensory traits, our preferences regarding colors, shapes, and flavor are shaped by multiple factors of the fruit, including its texture and chemical composition (El Hadi et al., 2013; Klee & Tieman, 2018). Human factors such as prior consumer experience, cultural background, and other attributes also affect our predilection. This complexity renders the objective assessment of consumer preferences challenging. For plant breeders, it makes routine evaluation a costly and time-consuming process, limiting formal sensory analyses to a few potential varieties, and impossible to be performed on a large scale.

In recent years, plant breeders have actively pursued new alternatives to assess and predict sensory traits. A modern approach relies on the integration of metabolite quantification, sensory panels, and machine learning methods to perform ‘metabolomic selection’ for consumer preferences (Colantonio et al., 2022). Promising results were reported for blueberry and tomato regarding the traits that affect consumer experience. However, fruit breeders now face the challenge of defining the target level or the minimum requirement that these traits should have to impact consumer experience. Using blueberry (Vaccinium corymbosum and hybrids) as an example, determining the minimal requirements for fruit-quality traits that significantly impact consumer overall liking had remained elusive. For instance, what is the minimum berry size or level of sweetness needed to satisfy consumers? By answering these questions, breeders could redefine priorities and avoid compromising some decisions on improving nontraditional target traits (i.e. aroma and health component attributes) to reach certain thresholds for fruit appearance and texture. Additionally, many of the standards reported in the fruit literature for blueberry are dated and have been defined by the industry using nonempirical methods (Beaudry, 1992; Retamales & Hancock, 2018).

Here we propose using data-driven methods to update breeding targets based on consumer data. Our primary hypothesis is that the interplay between fruit-quality attributes and consumer preferences might reach critical values that, when well estimated, may be used as empirical targets. The so-called ‘threshold effect’ occurs when the relationship between an outcome variable (the sensory trait) and a predictor variable (fruit-quality attribute) undergoes a sharp and statistically significant shift when the predictor value crosses a certain changing point (or threshold). As a numerical illustration, we can consider berry sweetness. If we assume that consumers, on average, have more enjoyable experiences when consuming berries with soluble solid content (SSC) rates higher than 11%; this value could be used as an empirical target to guide future decisions in a breeding program. The key challenge is how to estimate these thresholds accurately. The use of a proper statistical model and large datasets combining sensory and fruit quality will be needed.

Different methods can be applied to model the relationship between two variables and define such trends. The use of multivariate adaptive regression splines (MARS) stands out in the literature due to their flexibility in detecting interactions between predictors, robustness in dealing with outliers, and efficiency in handling multiple predictors (Friedman, 1991). Written as an extension of linear models, the MARS algorithm could be used to estimate empirical thresholds between sensory and fruit-quality traits using a two-stage approach. In the first stage, a range of predictors is partitioned into distinct groups, each of which is modeled with a separate linear regression and its slope. Leveraging spline theory, connections between the different regression lines are automatically explored to identify the optimal thresholds. In the second step, MARS estimates a least-squares model with its essential functions as independent variables. Therefore, MARS is a powerful tool for nonlinear regression analyses that can provide accurate predictions while maintaining interpretability and flexibility (Friedman & Roosen, 1995).

Alongside the adoption of novel statistical methods, genomic-assisted breeding is another critical strategy for flavor improvements. When framed as a prediction problem, genomic selection (GS) uses all available markers – regardless of the magnitude of their effects –to predict the genetic merit of an individual (Meuwissen et al., 2001). GS has been successfully applied for different traits and crops (Hickey et al., 2017), but they have been less exploited for fruit flavor improvement. Herein we propose to evaluate GS under a different objective: after defining the threshold values, how fruit targets (i.e. genotypes with fruit-quality attributes above or below the estimated thresholds) are predictable using genome-based methods? To answer this question, this study uses one of the most extensive genomic and sensory datasets reported in the fruit literature. We envision that methodology and results presented here could influence the food industry and breeding programs aiming to develop more flavorful cultivars, from any crop, by using a consumer-centric approach to define defined targets for flavor-related traits.

中文翻译：

---

一种以消费者为导向的方法，用于定义分子育种的育种目标

 介绍

风味、质地和外观都是我们饮食体验的重要方面，这些已成为现代水果育种计划的重点（Klee & Tieman， 2018）。针对这些感官方面的育种不仅会影响消费者的接受度，还会通过增加水果需求、适销性和种植者的盈利能力来影响链条的其他部分。尽管它很重要，但该领域的进展一直很缓慢，主要是因为确定符合消费者偏好的关键水果质量目标的复杂性。被认为是多感官特征，我们对颜色、形状和味道的偏好是由水果的多种因素决定的，包括它的质地和化学成分（El Hadi 等 人，2013 年;Klee & Tieman，2018 年）。人为因素，如先前的消费体验、文化背景和其他属性也会影响我们的偏好。这种复杂性使得对消费者偏好的客观评估具有挑战性。对于植物育种者来说，这使得常规评估成为一个昂贵且耗时的过程，将正式的感官分析局限于少数潜在品种，并且不可能大规模进行。

近年来，植物育种家积极寻求新的替代方案来评估和预测感官特性。一种现代方法依赖于代谢物定量、感官面板和机器学习方法的集成，以根据消费者的偏好进行“代谢组学选择”（Colantonio et al.， 2022）。据报道，蓝莓和番茄在影响消费者体验的性状方面取得了可喜的结果。然而，水果育种者现在面临的挑战是定义这些性状应具有的目标水平或最低要求，以影响消费者体验。以蓝莓（Vaccinium corymbosum 和杂交种）为例，确定对显着影响消费者整体喜好的水果品质性状的最低要求仍然难以捉摸。例如，满足消费者所需的最小浆果大小或甜度是多少？通过回答这些问题，育种者可以重新定义优先事项，并避免在改善非传统目标性状（即香气和健康成分属性）以达到水果外观和质地的某些阈值方面做出一些决定。此外，水果文献中报道的蓝莓标准品中的许多标准都是过时的，并且是由行业使用非实证方法定义的（Beaudry， 1992;Retamales & Hancock，2018）。

在这里，我们建议使用数据驱动的方法，根据消费者数据更新育种目标。我们的主要假设是，水果质量属性和消费者偏好之间的相互作用可能会达到临界值，如果估计得当，可以用作实证目标。当预测变量值超过某个变化点（或阈值）时，结果变量（感官特征）和预测变量（水果质量属性）之间的关系发生急剧且具有统计意义的变化时，就会发生所谓的“阈值效应”。作为一个数字示例，我们可以考虑浆果的甜度。如果我们假设消费者在食用可溶性固形物含量 （SSC） 高于 11% 的浆果时平均会有更愉快的体验;该值可以用作指导育种计划未来决策的经验目标。关键挑战是如何准确估计这些阈值。需要使用适当的统计模型和结合感官和水果质量的大型数据集。

可以应用不同的方法来对两个变量之间的关系进行建模并定义此类趋势。多元自适应回归样条曲线 （MARS） 的使用在文献中脱颖而出，因为它们在检测预测变量之间的交互方面具有灵活性，在处理异常值方面具有鲁棒性，并且在处理多个预测变量方面具有效率（Friedman，1991）。MARS 算法是作为线性模型的扩展编写的，可用于使用两阶段方法估计感官和水果品质性状之间的经验阈值。在第一阶段，将一系列预测变量划分为不同的组，每个组都使用单独的线性回归及其斜率进行建模。利用样条理论，可以自动探索不同回归线之间的联系，以确定最佳阈值。在第二步中，MARS 估计一个最小二乘模型，其基本函数作为自变量。因此，MARS 是非线性回归分析的强大工具，可以在保持可解释性和灵活性的同时提供准确的预测（Friedman & Roosen，1995 年）。

除了采用新颖的统计方法外，基因组辅助育种是风味改进的另一个关键策略。当被框定为预测问题时，基因组选择 （GS） 使用所有可用的标记 - 无论其影响的大小 - 来预测个体的遗传价值（Meuwissen et al.， 2001）。GS 已成功应用于不同的性状和作物（Hickey et al.， 2017），但它们在改善水果风味方面的开发较少。在此，我们建议在不同的目标下评估 GS：在定义阈值后，如何使用基于基因组的方法预测水果目标（即水果质量属性高于或低于估计阈值的基因型）？为了回答这个问题，本研究使用了水果文献中报道的最广泛的基因组和感官数据集之一。我们设想，这里介绍的方法和结果可以影响食品行业和育种计划，旨在通过使用以消费者为中心的方法为风味相关性状定义明确的目标，从任何作物中开发更美味的品种。