The current objective is to discuss if and where machine learning (ML) and artificial intelligence (AI) fit in the genetic prediction toolbox. Critical to the discussion is recognizing that, by definition, models currently employed for routine genetic prediction are part of the ML family, albeit linear as opposed to deep, and that ML fits under the more general umbrella of AI. Henderson’s Mixed Model Equations (MME) have served the animal breeding and genetic community well for several decades and have proven useful at predicting the additive genetic merit of animals. Moreover, extensions of the MME have clearly demonstrated the ability to fit non-additive genetic effects such as dominance and epistasis. The literature contains multiple examples of comparisons of genetic prediction between ML and mixed linear models, with a general consensus that ML does not outperform the MME when the trait(s) analyzed are largely additive in nature. With this in mind, the question of where ML fits is appropriate to ask. Given not all data are equally informative due to potential heterogeneity in the data generation process it is reasonable to employ ML to categorize data based on quality. Examples of using deep learning exist as a means of avoiding rules-based approaches to data editing and to categorize and weight data based on quality prior to inclusion in routine genetic evaluations. Additionally, the phenotypes used in genetic evaluations themselves could be predicted by ML. This might be the case in the use of animal tracking to garner data on social behavior, animal health and welfare, or metrics of libido and fertility; parasite loads predicted from camera data; or animal body weight gain or feed intake predicted from partial measurements. In all cases, the size of the raw data and lack of biological interpretation associated with it could necessitate the use of a predicted value. However, the error structure associated with the predicted values and the potential for compounding errors in genetic prediction need to be considered. Given the phenotype of an animal arises due to genetic and environmental effects and the complex interactions among and between these two broad sources of variation, the use of ML could be envisioned to predict phenotypic outcomes to aid in management and marketing decisions. To exploit this possibility, and outperform mixed linear models, the environmental data used would need to be sufficiently granular and high dimensional. There also exists the potential to employ ML to perform feature selection, as in the case of selecting variants from whole genome sequencing where feature subsets may be informed by multiple layers of external information. The use cases for AI also include decision support to help practitioners make use of enumerable genetic predictions given their unique objectives and circumstances.

中文翻译：

---

461 机器学习和人工智能改进肉牛的遗传预测：潜在用途和误用


目前的目标是讨论机器学习 （ML） 和人工智能 （AI） 是否适合遗传预测工具箱以及在何处适合。讨论的关键是认识到，根据定义，目前用于常规遗传预测的模型是 ML 家族的一部分，尽管是线性的而不是深度的，并且 ML 属于更普遍的 AI 范畴。几十年来，Henderson 的混合模型方程 （MME） 一直很好地服务于动物育种和遗传学界，并已被证明可用于预测动物的加性遗传价值。此外，MME 的扩展清楚地证明了拟合非加性遗传效应的能力，例如显性和上位性。文献包含 ML 和混合线性模型之间遗传预测比较的多个示例，人们普遍认为，当分析的性状在很大程度上是加性的时，ML 不会优于 MME。考虑到这一点，提出 ML 适合的问题是合适的。由于数据生成过程中可能存在异质性，因此并非所有数据都具有同等的信息量，因此使用 ML 根据质量对数据进行分类是合理的。使用深度学习的示例是作为避免基于规则的数据编辑方法的一种手段，以及在纳入常规遗传评估之前根据质量对数据进行分类和加权。此外，遗传评估本身使用的表型可以通过 ML 进行预测。使用动物追踪来收集有关社会行为、动物健康和福利或和生育能力指标的数据可能就是这种情况;从相机数据预测的寄生虫负荷;或通过部分测量预测的动物体重增加或采食量。 在所有情况下，原始数据的大小和缺乏与之相关的生物学解释都可能需要使用预测值。然而，需要考虑与预测值相关的误差结构以及遗传预测中复合误差的可能性。鉴于动物的表型是由于遗传和环境影响以及这两个广泛的变异来源之间和之间的复杂相互作用而产生的，可以设想使用 ML 来预测表型结果，以帮助管理和营销决策。为了利用这种可能性并超越混合线性模型，使用的环境数据需要足够精细和高维。还有可能使用 ML 来执行特征选择，例如从全基因组测序中选择变体，其中特征子集可能由多层外部信息通知。AI 的使用案例还包括决策支持，以帮助从业者根据其独特的目标和情况利用可枚举的遗传预测。