Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 https://doi.org/10.1186/s12711-021-00618-1 Ge n e t i c s Se l e c t i o n Ev o l u t i o n OPINION Open Access Phenomes: the current frontier in animal breeding Miguel Pérez‑Enciso1,2* and Juan P. Steibel3,4 Abstract Improvements in genomic technologies have outpaced the most optimistic predictions, allowing industry-scale application of genomic selection. However, only marginal gains in genetic prediction accuracy can now be expected by increasing marker density up to sequence, unless causative mutations are identified. We argue that some of the most scientifically disrupting and industry-relevant challenges relate to ‘phenomics’ instead of ‘genomics’. Thanks to developments in sensor technology and artificial intelligence, there is a wide range of analytical tools that are already available and many more will be developed. We can now address some of the pressing societal demands on the industry, such as animal welfare concerns or efficiency in the use of resources. From the statistical and computational point of view, phenomics raises two important issues that require further work: penalization and dimension reduc‑ tion. This will be complicated by the inherent heterogeneity and ‘missingness’ of the data. Overall, we can expect that precision livestock technologies will make it possible to collect hundreds of traits on a continuous basis from large numbers of animals. Perhaps the main revolution will come from redesigning animal breeding schemes to explicitly allow for high-dimensional phenomics. In the meantime, phenomics data will definitely enlighten our knowledge on the biological basis of phenotypes. Background Looking forward, we argue that some of the most sci- For the last two decades, genotyping and sequencing entifically disrupting and industry-relevant challenges technologies have outpaced the most optimistic predic- relate to ‘phenomics’ instead of ‘genomics’, as introduced tions, allowing industry-scale application of genomic in the premonitory words of Mike Coffey at the 2011 selection. Today, genomics is a mature technology but ICAR meeting, “In the age of the genotype, phenotype unfortunately the momentum may be fading away. is king”. Phenomics, which is defined as ‘the acquisition Increasing marker density follows the law of diminish- of high-dimensional phenotypic data on an organism- ing returns, unless the causative mutations are identi- wide scale’ [3], has flourished thanks to the development fied. Simulation and empirical results have shown that of all kinds of electronic devices and internet availabil- only marginal gains in genetic prediction accuracy can ity. Currently, sensors are able to inexpensively record be expected by genome sequencing data instead of high- images, videos, sounds, or a multitude of environmental density genotyping data [1, 2]. This small advantage will parameters, making large-scale, continuous phenotyping probably vanish when the extra cost of computer storage possible. Improvements in this vast area occur at breath- and power that sequence analyses require compared to taking pace. genotyping arrays are accounted for. However, there are some caveats in the ‘phenome’ concept [4]. While the genome, i.e., the DNA sequence, is finite and can in principle be fully characterized, the *Correspondence: miguel.perez@uab.es phenome is not a closed, fully-defined entity, and it will 2 Centre for Research in Agricultural Genomics (CRAG), CSIC-IRTA-UAB-UB, never be. An infinite number of phenotypes can be imag- Bellaterra, 08193 Barcelona, Spain Full list of author information is available at the end of the article ined: simply consider all the mathematical combinations © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat​iveco​ mmons​.org/licen​ses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creat​iveco​mmons​.org/publi​cdoma​in/ zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 2 of 10 of measured traits that can be defined. Furthermore, and are a major component of animal welfare. Given the phenotypic measurements may involve several individu- difficulty of measuring behavior before the ‘phenomics als simultaneously, as in many welfare and behavioral era’, in genetic evaluations the effect of behavior on pro- traits. Therefore, a phenome will always be a subset of duction traits has been either ignored or accounted for an infinite number of possible measurements that may indirectly, e.g., by using social genetic effects models [13]. span several individuals. The difference with a ‘stand- However, selection to modify behavior is possible since ard’ breeding setting is that, in the new paradigm, both some behaviors, especially those related to aggression, the genome and the phenome are high-dimensional vari- are partially heritable [14]. Understanding and modify- ables. Furthermore, whereas genome data are relatively ing the genetic factors that induce fighting between pairs homogeneous, phenome measurements can be highly of animals is valuable not only for selection, but also for heterogeneous and time-dependent. An example is the management purposes, as groups could be formed based composition of microbiota, which change from birth on the genetic makeup of animals that are expected to to adult stages and may vary with health status. In this fight less with each other. As a result, welfare and pro- sense, it is important to realize that phenomics ‘big data’ ductivity should increase. problems arise because of the heterogeneity and rapid Today, behavioral traits can be measured with wear- change over time of the data, not because of their size. able sensors and computer vision techniques [15–17]. Compared to plant breeding or human genetics [3], Behavior metrics can affect single individuals but often animal phenomics has received somewhat less atten- involve pairs or larger groups of animals. If behaviors tion. This is surprising since new technologies allow the are measured at the individual level, multi-trait direct assessment of new phenotypes that are in high demand genetic effects models can be used to obtain breeding by the society, such as those related to animal welfare, values for the behaviors of interest together with produc- resilience, disease incidence or resource use efficiency tion and welfare traits [18]. A less explored option is the [5–8]. Fortunately, recent work indicates that animal phe- modeling of behavior at the dyadic level (i.e., in individ- nomics is becoming popular in the animal sciences too, ual pairs). This type of data has been used to parameter- e.g., [9–12], and as reflected in public initiatives such as ize social genetic effects models [19], but it could also be the USDA Agricultural Genomes to Phenomes Initiative analyzed as a response matrix in a quantitative genetic AG2PI (https​://www.ag2pi​.org/). study to determine, e.g., which genetic factors affect The novelty around high-throughput phenotyping a group mate to be attacked. This type of analysis has in farm animal populations comes from two angles: (1) not been performed yet, but dyadic behavioral data are novel traits can be defined and measured that could not being collected in livestock species and have been used be recorded before, and (2) classical traits can now be to build social networks [20]. A typical interaction behav- observed on an almost continuous basis and in a non- ior is aggression (e.g., post-mixing attacks), where we dis- invasive way on large numbers of animals under normal tinguish between an individual measure and a group or production environments. However, we can expect the dyadic measures. For instance, the total amount of time datasets to be partially incomplete, noisy and partially that each animal spends engaged in fighting represents redundant, especially when traits are recorded on a con- an individual-level observation. But when the amount of tinuous basis. time spent fighting is annotated for each pair of animals With this opinion paper, our aim is to foster discus- in a social group, a dyadic phenotypic dataset is gener- sion and further research in the area. Towards this, we ated. This type of datasets allows the development of spe- briefly recall some of the most relevant traits that can be cific, new modeling strategies (Fig. 1). captured by modern technologies, and then discuss the future needs in terms of methods and algorithms and Gas emissions the possible long-term impact of phenomics in breeding. Genetic selection for reduced greenhouse gas emissions The focus of this note is on the use of sensors to collect is possible, given that the trait is heritable and geneti- measures on animals themselves, but note that sensors cally correlated with other traits such as milk produc- can and are used to record all sorts of environmental var- tion and residual feed intake [21]. Traditional measuring iables (climate, pathogen exposure, etc.), which are also methods that involve respiration chambers and analysis highly relevant for animal management and breeding. of tracer elements are not readily scalable for applica- tion in high-throughput phenotyping. However, recently Standard and new traits (re)visited developed spectroscopy technology for “breath analysis” Behavior at the individual and group levels can be applied on a large scale to measure methane emis- Behavior and social interactions between animals can sions. This technology can be incorporated into on-farm greatly affect production and performance phenotypes devices such as feeding kiosks or milking robots and used Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 3 of 10 a b d c Fig. 1 Emerging behavioral data in the phenomics era and the need for new models. a Behavioral phenotyping for social interactions results in a matrix of dyadic interactions, Zs, that can be collapsed to individual behavioral data (w and y). b Existing genomics and phenomics data can be integrated with behavioral phenotypes. c Classic genomic evaluation models focus on multi-trait analyses of individual behaviors or on social genetic effects models where the interaction matrix is used as a predictor of existing phenotypes. d In novel models, multi-trait analyses have to include full behavioral matrices to be able to predict the dyadic interactions from rker data to measure instantaneous methane emissions on thou- heterogeneity [24]. Because different devices are used sands of animals as they voluntarily approach measuring to record and pre-process feed intake data, the meas- stations several times per day [22]. ured traits may be slightly different. Sometimes, due Like other phenotyping data described in this paper, to malfunction, there may be partially missing data, incorporating high-throughput measures of emissions for instance, meal is recorded but no animal identity into genetic evaluations will require the integration of (ID) is attached to it or an animal’s visit is recorded heterogeneous data sources. In this case, the heterogene- but the feed intake is clearly wrongly measured or not ity will arise not only from a variety of measuring devices, measured at all. These peculiarities of the data record- but also from the different sampling schemes and meas- ing process need to be accounted for in the analysis ures. Highly processed and distilled data such as esti- pipelines, including raw data processing, cleaning, and mated total emissions per day will be available for some imputation. individuals together with raw instantaneous measure- On the data modeling side, there are opportuni- ments of emission for other animals. ties to extract extra information and novel traits from automatic feeding station data. The sequences of the Feed efficiency visits to the feeders, the time between visits and the Feed efficiency largely determines farm exploitation co-occurrence of animals in multi-space feeders can be costs and has been measured on farms on a limited mined to uncover interactions between animals. Also, scale for many years now. As with behavior, measur- meal characteristics extracted from automatic feeders ing feed efficiency accurately and massively has only have been used to parameterize social genetic effects been possible due to sensor technology, via automatic models [25]. Combining automatic feeding data with feed recording devices [23]. An inherent characteris- computer vision algorithms will likely shed further light tic of feed intake and feeding behavior data obtained on these problems. with automatic feeders is its incompleteness and its Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 4 of 10 Traditional phenotypes revisited their productive life. Computer vision algorithms for ani- Sensor technology allows the measurement of new phe- mal identification use different techniques, such as natu- notypes but is also disrupting how ‘standard’ traits, such ral variation in appearance (conformation, coat color, as weight or conformation, are recorded. In the case of etc.) of the whole animal, visual marking or tags that are weight and condition score, traditional methods typi- permanently or temporarily attached to the animals [27– cally require moving the animals and are labor inten- 29]. Ideally, a computer vision algorithm will work using sive; for that reason, only a limited number of measures images of panoramic views of the animal space (pen or can be taken in each production cycle. With the advent barn), which are taken with ceiling or high-wall mounted of phenomics, these traditional phenotypes can be col- cameras that include images of several animals in the lected automatically on a continuous basis on a massive same picture frame. number of animals, without the need to disturb them. Automatic, reliable animal identification is not a solved For instance, weight can be accurately measured using issue. Among the challenges that need to be addressed 3D imaging [17], or conformation features can be meas- are: (1) improving current computer vision algorithms ured from images [26]. These precision livestock farming for animal ID and tracking using top-view cameras in devices record under normal production conditions of normal production conditions, (2) the integration and market animals, and not only on elite animals in nucleus synchronization of multiple data streams (e.g., RFID farms. Together with these measures, environmental detection logs with video streams from more than one records of similar temporal and spatial resolution can be video camera), and (3) the ascertainment of animal iden- obtained from weather stations or from barn environ- tification or accounting for uncertain ID. For instance, ment controlling devices. what is to be done if the animal identification algorithm Nevertheless, the incorporation of automatically- produces two likely ID with a roughly similar probability measured body condition score or weight gain into exist- for a single animal image? Shall the ID be treated as miss- ing genomic evaluations imply unique challenges. Data ing data? Or should the uncertainty be propagated into streams from continuously-measured live-weight on mil- the animal genetic evaluation by using methods similar lions of animals will have to be summarized and cleaned to those proposed for dealing with uncertain paternity before feeding them into existing genetic evaluations. [30–32]? These questions will be revisited when these Cleaning data by detecting outliers before fitting growth kinds of data streams become common. curves to individual records may not be efficient in the phenomics era; instead, robust data analysis techniques High dimensionality such as non-linear quantile regression can be used for At the end of the day, phenomics technologies deliver growth curves using all available data points but avoid- highly-dimensional data that need to be processed and ing the effect of outlying observations. Finally, having full incorporated into breeding and management decisions. growth curves from each animal will allow the evaluation Two main, related statistical issues are relevant in this of new traits around these traditional phenotypes. context: dimension reduction and penalization. Dimen- sion reduction consists of obtaining new ‘synthetic’ vari- Traceability and identification ables that are combinations of the original dimensions. Environmental data collected by sensors allow the study A usual justification of dimension reduction is that only of genotype-by-environment interactions at a higher res- a few dimensions are actually relevant, and that dimen- olution scale than has been so far possible and the incor- sionality is artificially high. Penalization refers to setting poration of production farm records into the evaluation constraints of the parameter solutions in the predictive of elite animals. However, for such uses, the data need model. to be linked in some way: phenotypic records of produc- In practice, dimension reduction techniques are mainly tion animals need to be linked to parental genotypes and used for visualization. New variables in the reduced to environmental and production records. In symmetry, dimensional space are derived to retain the original data continuous individual identification throughout the pro- pattern as faithfully as possible. In principal component duction cycle will open the opportunity to record many analysis (PCA), the new variables are the linear combina- new traits. Thus, the needs for animal identification and tions of the original phenotypes that explain the maxi- tracking as well as the synchronization of real-time data mum possible variance, with the additional constraint streams will increase. of being orthogonal between them. PCA, together with Currently, individual identification is attained through multidimensional scaling (MDS), are perhaps the most computer vision of marked or unmarked animals or widely used dimension reduction tools, but interesting using wearable radio frequency identification (RFID) and less well-known options exist and can be good alter- devices that remain with each animal for the duration of natives. Some of these provide nonlinear approximations, Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 5 of 10 in contrast to linear PCA. An ‘autoencoder’ (AE) is one of However, t-SNE offers a series of advantages over MDS; such nonlinear alternatives [33]. Autoencoders are ‘deep in particular, it reduces the tendency of samples to clus- learning’ (DL) algorithms, i.e., they are based on several ter together, which is caused by a large number of dimen- stacked layers of neurons (Fig. 2). However, in contrast to sions and results in increased resolution. It seems that a typical DL network, where output and input are differ- neither t-SNE nor autoencoders are popular in animal ent, both input and output are the same in AE. Thus, they phenomics, but they are techniques worth exploring are unsupervised techniques, as they are mainly dimen- since they offer complementary information to standard sionality reduction methods. If no restriction is set, the linear methods. Figure 3 illustrates how different algo- optimum AE solution is the identity vector and recon- rithms can provide dramatically different representations structed output is identical to input. It is then necessary in the low dimensional space. to set some constraints, i.e., penalizations, to optimize Penalization refers to setting constraints on variables the AE network. In the specific case of one single layer to avoid collinearity and overfitting problems when the and a linear activation function, it has been shown that number of variables is large. As well known, penaliza- AE and PCA yield basically the same solution [34]. tion is needed to avoid the method ‘learning’ irreproduc- Another interesting dimension reduction algorithm is ible noise, i.e., one of the ‘curses of dimensionality’ [36]. t-distributed stochastic neighbor embedding (t-SNE, van Two main types of regularization have been proposed: der Maaten and Hinton [35]). The goal of t-SNE is to find L1 and L2. L1 consists in setting a constraint on the sum a low dimensional representation whereby similar data- of absolute values of the solutions, whereas L2 refers to points in the original space are shown together and dis- the sum of squared solutions [37]. Concepts such as prior tant samples are shown far apart. The most important information in a Bayesian framework are equivalent to difference between PCA and t-SNE is that the former is penalization. Although numerous methods with differ- a projection onto a lower dimensional space, whereas the ent names have been proposed in the Bayesian literature, latter is a representation strategy. Besides, by construc- most of them can be unified by realizing that they simply tion, PCA is aimed at maximizing distances when sam- differ in the prior chosen [38]. In addition, deep learning ples are plotted in the low dimensional space. MDS is a technologies have developed specific additional regulari- generalized approach for PCA and, similar to t-SNE, it zation strategies. One of them is called ‘dropout’, which is designed for preserving distances between samples. consists of randomly removing ‘neurons’ from the inner Fig. 2 Representation of an Autoencoder. Autoencoders (AE) are deep neural networks where the input and output are the same (in this case, multi-channel pixel intensity values from images of livestock). They consist of an encoder that codes the input in a low dimensional latent space and a decoder that transforms back the input into a regularized version. Variational autoencoders (VAE) generate a probability function instead of a point latent space. Then, random numbers are drawn and transformed into simulated images by the decoder. Applications of AE and VAE to phenomics remain to be explored, but they can be used for unsupervised learning and imputation. The figure of the cow is from www.dream​stime​ .com Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 6 of 10 Fig. 3 Comparison between PCA and t-SNE dimensional reduction methods using the 3D ‘S’ shape in left panel. Note that PCA is a projection that aims at maintaining the maximum variance, whereas t-SNE keeps local similarity in the low dimensional space. For instance, PCA projection clearly maintains the original ‘S’ shape, where the third dimension is lost. In contrast, the plot produced by t-SNE is better at representing local relative distances, where the first dimension reproduces the contour of the shape and the second dimension, the relative position within that part of the contour. As a result of different targets, very different pictures emerge. Implications for phenomics are to be explored. Plot using scikit [53], slightly modified from J. Vanderplas code (https​://sciki​t-learn​.org/stabl​e/auto_examp​les/manif​old/plot_compa​re_metho​ds.html) layers to force the system to use fewer parameters. In our field. One reason is that having many highly corre- spite of its distinct definition, dropout can be interpreted lated variables helps to smooth out noise. A further, more from a Bayesian point of view, and thus can be viewed relevant advantage is that increasing the number of vari- within the usual penalization framework [39]. Another ables usually results in penalized models with improved approach used in some DL models is direct L1 or L2 predictive performance [40]. Finally, increasing dimen- penalization on the neuron weights, i.e., a constraint is sionality allows a gain in biological insight. added on the sum of the absolute or square value of the weights. The way ahead Complexity, heterogeneity, and especially data size will Strangely, one of the first obstacles that will need to be notably increase in the phenomics era, which will have solved for routine phenome collection is access to broad- repercussions in the modeling approaches. It is some- band internet. Even in the USA, as much as 40% rural times erroneously believed that the influence of the farms lack reliable access to broadband [9]. Aside from prior vanishes with large datasets, but this is not the case infrastructure issues, here we wish to focus on methodo- since the prior will always have an effect on the solution, logical issues. regardless of the amount of data [38]. Thus, it is worth Quality control and visualization techniques should studying the impact of alternative regularization strat- be a first step in a phenomics pipeline. As phenotyping egies, as we cannot expect that a single strategy will be becomes a large-scale objective, reliability can be com- optimum—from the predictive performance point of promised and heterogeneity in environmental conditions view—in all cases. may increase. However, bias can be a much more serious As just mentioned, the term ‘curse of dimensional- danger than error because phenome data will not be col- ity’ has become popular in statistics and by extension in lected randomly. Likely, special attention will be paid to breeding to mean that increasing ‘unnecessarily’ the com- elite farms or breeding nucleus, and certain phenotypes plexity of a model leads to poor predictive performance will be collected preferentially on specific farms. We can [37]. The term ‘unnecessarily’ can be read as ‘without also expect differences in accuracy for data from elite penalization’. According to Donoho [36], the term was farms compared to data from commercial farms, which initially coined to reflect the impossibility of enumerat- will have to be accounted for through proper modeling. ing all possible models as the number of variables grows. As a result, phenome data will also be highly unbal- However, high dimensionality is a ‘blessing’ for many anced: the kind and amount of data available will vary purposes, an aspect that is less widely acknowledged in largely across individuals, even if sensors are widely Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 7 of 10 spread and collect information routinely. It is unlikely output is a regularized representation of the original that identical phenotypes are recorded on different input. This is accomplished via an ‘encoder’ that trans- farms or in different periods, whether milk recording or forms the data into a ‘latent space’ and a ‘decoder’ that behavior measurements. This can be an important obsta- takes the latent space coordinates and outputs a regu- cle, since it will require either removal of samples and/ larized image. Instead of recoding the input into latent or imputation of missing values. We need to develop space coordinates, variational autoencoders (VAE) gener- efficient and accurate imputation tools or use methods ate a probabilistic function to describe an observation in dealing with missing data directly. In this aspect, plant the latent space. As a result, realistic data-points can be phenomics measurements can often be more systemati- generated. For instance, VAE have been used to increase cally compared and measured on a larger scale than in the resolution of pictures or to restore damaged pictures, livestock. a problem that is conceptually identical to imputation. In Given that missing data is unavoidable, imputation will the same vein, generational adversarial networks (GAN) be needed. This is a wide area and numerous approaches are DL algorithms that can reproduce high-dimensional exist depending on the specific problem, e.g., [41, 42]. variables. So far, GAN have been mostly used to gener- However, it should be noted that most imputation tech- ate images, e.g., to generate very high-resolution pic- niques assume that missing data is at random, a condition tures out of incomplete or low-resolution images, or that phenome data will unlikely fulfill, as discussed above. even videos [43, 44]. Figure 4 represents a scheme of A further issue with phenomics data is their heterogene- a GAN for ‘inpainting’, i.e., completing missing parts in ity and thus no general imputation rule can be given. In an image. Compared to VAE, GAN are much more flex- addition to standard imputation techniques [42], alter- ible, yet they are more difficult and slower to train. They native approaches exist based on deep learning that, to also require larger amounts of data. Applying GAN and our knowledge, have not been used in this area and can VAE concepts to imputation in phenomics is a promis- be promising. For instance, autoencoders can be used to ing area of research, given their flexibility and lack of dis- fill ‘holes’ in data, in particular those with a spatial pat- tributional assumptions. However, caution applies, since tern such as image and video. The standard autoencoders these methods have been tested mainly with image data, Fig. 4 Representation of a generative adversarial network (GAN). This GAN aims at filling holes in an image (‘inpainting’, i.e., imputation). The generator network simulates a new image out of the input picture of the cow, which does not have front legs. The discriminator is trained with images that are true or fake and learns how a real cow looks like. The generator discriminates whether the generated image is true or fake, the result is passed on to the generator so that it can improve the quality of the output image. Each of the rectangles in the generator and the discriminator represent a group of neuron layers as in Fig. 3, the shape is approximately proportional to its dimension. The figure of the cow is from www.dream​ stime​.com Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 8 of 10 and performance in other scenarios has not been much argued that the future of artificial intelligence lies in explored. unsupervised learning (https​://www.youtu​be.com/watch​ However, even with complete data of reasonable size, ?v=Ount2​Y4qxQ​o&t=1072s​, NIPS conference, 2016). there are important gaps in the current methodology This is because unlabeled data (unsuited for super- and software that should be filled. There is an enormous vised learning) are far more abundant than labeled data range of analytical tools to be developed. Automatic indi- and, more importantly, because unsupervised learning vidual identification allowing for free range, or at least resembles more precisely how the human brain actually movement in group-housed animals under confinement works. As phenome data are collected over the years in is needed. This can be accomplished, e.g., using continu- the same or similar breeding schemes, by using unsuper- ous video recording and tracking. Algorithms that auto- vised learning methods, we will gain invaluable knowl- matically extract phenotypes from image, video, sound edge on the effects of selection on the whole organism. records are a very active area of research. Standard met- For instance, unsupervised learning may uncover unsus- rics (e.g., Hausdorff, Euclidean, etc.) to measure similari- pected relationships between traits or between traits and ties between images should be adapted to livestock, and environmental variables. We could, e.g., discover how new metrics for videos should be implemented as they some selection barriers can be overcome or how to opti- are required to compare animal behaviors. Automatic mize economic weights dynamically. conformation measures [26] must be improved to be per- Phenomics is a hot, promising area but is not exempt formed in vivo with minimal human intervention. A fur- from risks and cannot be considered a panacea. As Cole ther challenge is to incorporate all this information into et al. [12] warn us, ‘these new approaches have their own genome predictions. challenges, ranging from bias to interpretability and there Not all species and breeding programs will benefit is a temptation to oversell outcomes’. One serious issue equally from phenomics. For instance, the aquaculture is that, in contrast to genotypes, phenome data may not industry is highly advanced technologically and many be easily transferable or comparable across farms. Often, measurements are difficult to obtain manually. Here, the measurement technology is proprietary and several sys- use of phenomic technologies is far more widespread tems, e.g., to measure methane emissions, coexist. Algo- than in other species. Small ruminants in extensive farm- rithms to transform raw sensor data into meaningful ing may represent the opposite extreme. Yet, sensor tech- measurements and sensors themselves rapidly change nology can provide accurate animal tracking and remote over time, making it difficult to analyze data longitudi- measurement of physiological and environmental varia- nally. Standards and open-source algorithms in the sen- bles outdoors, which could boost productivity and health sor industry are necessary to fully unravel the potential precisely in traditional, low-input agriculture condi- of phenomics. A way to foster the development of novel tions. In general, those production systems where preci- algorithms in this field is the distribution of relevant sion agriculture and precision livestock management are datasets to the research community and the organization being implemented will be better equipped to collect rel- of activities (hackathons, competitions and journal spe- evant phenomics data, and the limitation in those cases cial issues) around the analysis of such datasets. This is will be on data storage and transmission as well as data being done successfully in the fields of autonomous vehi- labeling and individual identification. cles, computer vision [49] and in many other areas (e.g., Although model interpretability is an issue [45, 46], https​://www.kaggl​e.com/compe​titio​ns). experience shows that opening the ‘black box’ is not Perhaps the main revolution will come from redesign- needed for accurate prediction [40]. Nevertheless, phe- ing animal breeding schemes to explicitly allow for highly nomics data will definitely enlighten the biological basis dimensional phenomics. To begin with, do we need new of phenotypes and will complement genotype data. As definitions of breeding values? Genotype × environment geneticists, we are often in the quest of DNA causative (G × E) interactions can be an inspiring concept here. polymorphisms. However, focusing on this search may In essence, including G × E interactions in the model is often impede the detection of non-genetic factors that equivalent to providing a function for the breeding value have a stronger effect on the phenotypic expression than instead of a single value, a reaction norm. In the phenomics the causative mutations. This has been observed, e.g., in era, traditional point breeding values might be replaced by expression quantitative trait loci (QTL) studies [47]. We high-dimensional generative functions. It is not clear at this contend that large-scale phenotyping is important per point how this will be accomplished. A natural approach is se, independently of whether genome data are available to use phenomics data to integrate mechanistic biological or not. In this context, structural equation modeling [48] models into genetic evaluation. Examples are growth crop but also unsupervised learning gains relevance. Influ- models in plants such as those developed by Totir et al. [50, ential DL pioneers, such as Yann LeCun, have actually 51]. More generally, we hypothesize that phenomics-based Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 9 of 10 genomic evaluations will likely be a combination of stand- References 1. Pérez-Enciso M, Rincón JC, Legarra A. Sequence- vs. chip-assisted ard statistics methods with generative machine-learning genomic selection: accurate biological information is advised. Genet Sel and simulation tools. In a recent work, de los Campos Evol. 2015;47:43. et al. [52] applied large-scale simulation conditioned on 2. van Binsbergen R, Calus MPL, Bink MCAM, van Eeuwijk FA, Schrooten C, Veerkamp RF. Genomic prediction using imputed whole-genome genotype and environmental variables to predict future sequence data in Holstein Friesian cattle. Genet Sel Evol. 2015;47:71. performance but, instead of a point prediction, a whole 3. Houle D, Govindaraju DR, Omholt S. Phenomics: the next challenge. Nat distribution over uncertain, future climate conditions was Rev Genet. 2010;11:855–66. 4. Mahner M, Kary M. What exactly are genomes, genotypes and pheno‑ generated. We can imagine that future phenomics-assisted types? And what about phenomes? J Theor Biol. 1997;186:55–63. breeding schemes will be able to simulate expected com- 5. Brito LF, Oliveira HR, McConn BR, Schinckel AP, Arrazola A, Marchant-Forde plex phenotypes under a range of potential environmental JN, et al. Large-scale phenotyping of livestock welfare in commercial production systems: a new frontier in animal breeding. Front Genet. conditions for each target genotype. 2020;11:793. 6. Rosenfeld A. Image analysis and computer vision: 1992. CVGIP Image Conclusions Underst. 1993;58:85–135. 7. Berghof TVL, Poppe M, Mulder HA. Opportunities to improve resilience in The influence of phenomics in livestock breeding has only animal breeding programs. Front Genet. 2019;9:692. begun and much work remains to be done. High dimen- 8. Pooley CM, Marion G, Bishop SC, Bailey RI, Doeschl-Wilson AB. Estimat‑ sionality is a ‘blessing’ rather than a ‘curse’ to improve ing individuals’ genetic and non-genetic effects underlying infectious disease transmission from temporal epidemic data. PLoS Comput Biol. prediction [40] and we should not be afraid of it. High 2020;16:e1008447. dimensionality should also help breeders to fine tune 9. Koltes JE, Cole JB, Clemmens R, Dilger RN, Kramer LM, Lunney JK, et al. A which are the most relevant phenotypes, and what are vision for development and utilization of high-throughput phenotyping and big data analytics in livestock. Front Genet. 2019;10:1197. the expected constraints. Making progress in phenomics 10. Baes C, Schenkel F. The future of phenomics. Anim Front. 2020;10:4–5. depends on the fast-developing field of sensor technol- 11. Rexroad C, Vallet J, Matukumalli LK, Reecy J, Bickhart D, Blackburn H, et al. ogy and machine learning. This reinforces the idea that Genome to phenome: improving animal health, production, and well- being—a new USDA blueprint for animal genome research 2018–2027. the breeders of the future will require sound agronomic Front Genet. 2019;10:327. and biological backgrounds as well as a solid training in 12. Cole JB, Eaglen SAE, Maltecca C, Mulder HA, Pryce JE. The future of phe‑ statistical and machine learning. This may seem paradoxi- nomics in dairy cattle breeding. Anim Front. 2020;10:37–44. 13. Bijma P. The quantitative genetics of indirect genetic effects: a selective cal, but it will be the case because easy-to-use, powerful review of modelling issues. Heredity. 2014;122:61–9. programming libraries and code will be widely available, 14. Turner SP, D’Eath RB, Roehe R, Lawrence AB. Selection against aggres‑ whereas interpretation of results and application in breed- siveness in pigs at re-grouping: practical application and implications for long-term behavioural patterns. Anim Welf. 2010;19:123–32. ing schemes require specific biological and agronomic 15. Chen C, Zhu W, Steibel J, Siegford J, Wurtz K, Han J, et al. Recognition of knowledge. Last but not least, breeders trained in phenom- aggressive episodes of pigs based on convolutional neural network and ics who can effectively collaborate with biologists, produc- long short-term memory. Comput Electron Agric. 2020;169:105166. 16. Brown-Brandl TM, Rohrer GA, Eigenberg RA. Analysis of feeding behavior ers, engineers and computer scientists will have increased of group housed growing-finishing pigs. Comput Electron Agric. chances of succeeding in the job market. 2013;96:246–52. 17. Fernandes AFA, Dórea JRR, de Rosa GJ. Image analysis and computer Acknowledgements vision applications in animal sciences: an overview. Front Vet Sci. We are grateful for the comments and suggestions from L.M. Zingaretti, 2020;7:551269. and GSE editors for endorsing this note. MPE is funded by MINECO Grants 18. Turner SP. Breeding against harmful social behaviours in pigs and PID2019-108829RB-I00/AEI/10.13039/501100011033 and “Centro de Excelen‑ chickens: state of the art and the way forward. Appl Anim Behav Sci. cia Severo Ochoa 2016-2019” award SEV-2015-0533 (Spain). JPS is funded by 2011;134:1–9. NIFA award 2017-67007-26176 (USA). 19. Angarita BK, Cantet RJC, Wurtz KE, O’Malley CI, Siegford JM, Ernst CW, et al. Estimation of indirect social genetic effects for skin lesion count in Authors’ contributions group-housed pigs by quantifying behavioral interactions. J Anim Sci. Both authors conceived, discussed and wrote the work jointly. Both authors 2019;97:3658–68. read and approved the final manuscript. 20. Foister S, Doeschl-Wilson A, Roehe R, Arnott G, Boyle L, Turner S. Social network properties predict chronic aggression in commercial pig sys‑ Compliance with ethical standard tems. PLoS One. 2018;13:e0205122. 21. Lassen J, Løvendahl P. Heritability estimates for enteric methane emis‑ Competing interests sions from Holstein cattle measured using noninvasive methods. J Dairy The authors declare that they have no competing interests. Sci. 2016;99:1959–67. 22. Negussie E, Lehtinen J, Mäntysaari P, Bayat AR, Liinamo AE, Mäntysaari Author details EA, et al. Non-invasive individual methane measurement in dairy cows. 1 ICREA, Passeig de Lluís Companys 23, 08010 Barcelona, Spain. 2 Centre Animal. 2017;11:890–9. for Research in Agricultural Genomics (CRAG), CSIC-IRTA-UAB-UB, Bellaterra, 23. Lu D, Jiao S, Tiezzi F, Knauer M, Huang Y, Gray KA, et al. The relationship 08193 Barcelona, Spain. 3 Department of Animal Science, Michigan State between different measures of feed efficiency and feeding behavior traits University, East Lansing, MI 48824, USA. 4 Department of Fisheries and Wildlife, in Duroc pigs. J Anim Sci. 2017;95:3370–80. Michigan State University, East Lansing, MI 48824, USA. 24. Casey DS, Stern HS, Dekkers JCM. Identification of errors and factors associated with errors in data from electronic swine feeders. J Anim Sci. Received: 8 December 2020 Accepted: 22 February 2021 2005;83:969–82. 25. Ragab M, Piles M, Quintanilla R, Sánchez JP. Indirect genetic effect model using feeding behaviour traits to define the degree of interaction Pérez‑Enciso and Steibel Genet Sel Evol (2021) 53:22 Page 10 of 10 between mates: an implementation in pigs growth rate. Animal. 41. Browning BL, Browning SR. Improving the accuracy and efficiency 2019;13:231–9. of identity-by-descent detection in population data. Genetics. 26. Nye J, Zingaretti LM, Pérez-Enciso M. Estimating conformational traits in 2013;194:459–71. dairy cattle with DeepAPS: a two-step deep learning automated pheno‑ 42. van Buuren S, Groothuis-Oudshoorn K. mice: multivariate imputation by typing and segmentation approach. Front Genet. 2020;11:513. chained equations in R. J Stat Softw. 2011;45:1–67. 27. Psota ET, Mittek M, Pérez LC, Schmidt T, Mote B. Multi-pig part detection 43. Karras T, Aila T, Laine S, Lehtinen J. Progressive growing of GANs for and association with a fully-convolutional network. Sensors. 2019;19:852. improved quality, stability, and variation. arXiv​:1710.10196​; 2017. 28. Chen C, Zhu W, Steibel J, Siegford J, Han J, Norton T. Recognition of 44. Alqahtani H, Kavakli-Thorne M, Kumar G. Applications of generative feeding behaviour of pigs and determination of feeding time of each adversarial networks (GANs): an updated review. Arch Comput Methods pig by a video-based deep learning method. Comput Electron Agric. Eng. 2019;28:525–52. 2020;176:105642. 45. Biecek P, Burzykowski T. Explanatory model analysis: explore, explain, and 29. Thomasen JR, Lassen J, Nielsen GGB, Borggard C, Stentebjerg PRB, examine predictive models. London: Chapman & Hall/CRC Data Science Hansen RH, et al. Individual cow identification in a commercial herd using Series; 2021. 3D camera technology. In: Proceedings of the 11th World Congress on 46. Castelvecchi D. Can we open the black box of AI? Nature. 2016;538:20–3. Genetics Applied to Livestock Production: 7–11 February 2018; Auckland. 47. Pérez-Enciso M, Quevedo JR, Bahamonde A. Genetical genomics: use all 2018. data. BMC Genomics. 2007;8:69. 30. Cardoso FF, Tempelman RJ. Bayesian inference on genetic merit under 48. Wu XL, Heringstad B, Gianola D. Bayesian structural equation models for uncertain paternity. Genet Sel Evol. 2003;33:469–87. inferring relationships between phenotypes: a review of methodology, 31. Cardoso FF, Tempelman RJ. Genetic evaluation of beef cattle accounting identifiability, and applications. J Anim Breed Genet. 2010;127:3–15. for uncertain paternity. Livest Prod Sci. 2004;89:109–20. 49. Caesar H, Bankiti V, Lang AH, Vora S, Liong VE, Xu Q, et al. nuScenes: a mul‑ 32. Perez-Enciso M, Fernando RL. Genetic evaluation with uncertain parent‑ timodal dataset for autonomous driving. arXiv​:1903.11027​; 2019. age: a comparison of methods. Theor Appl Genet. 1992;84:173–9. 50. Cooper M, Technow F, Messina C, Gho C, Totir LR. Use of crop growth 33. Hinton G, Salakhutdinov R. Reducing the dimensionality of data with models with hole-genome prediction: application to a maize multienvi‑ neural networks. Science. 2006;313:504–7. ronment trial. Crop Sci. 2016;56:2141–56. 34. Goodfellow I, Bengio Y, Courville A. Deep learning. Cambridge: The MIT 51. Technow F, Messina CD, Totir LR, Cooper M. Integrating crop growth Press; 2016. models with whole genome prediction through approximate Bayesian 35. van der Maaten L, Hinton GE. Visualizing data using t-SNE. J Mach Learn computation. PLoS One. 2015;10:e0130855. Res. 2008;9:2579–605. 52. Los Campos G, Pérez-Rodríguez P, Bogard M, Gouache D, Crossa J. A 36. Donoho DL. High-dimensional data analysis: The curses and blessings of data-driven simulation platform to predict cultivars’ performances under dimensionality. In: Proceedings of the AMS Conference on Mathemati‑ uncertain weather conditions. Nat Commun. 2020;11:4876. cal Challenges of the 21st Century: 7–12 August 2000; Los Angeles. 53. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, 2000;1–33. et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 37. Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: Data 2011;12:2825–30. mining, inference, and prediction. 2nd ed. New York: Springer; 2009. 38. Gianola D. Priors in whole-genome regression: the Bayesian alphabet returns. Genetics. 2013;194:573–96. Publisher’s Note 39. Gal Y, Ghahramani Z. Dropout as a Bayesian approximation: representing Springer Nature remains neutral with regard to jurisdictional claims in pub‑ model uncertainty in deep learning. arXiv​:1506.02142​; 2016. lished maps and institutional affiliations. 40. Breiman L. Statistical modeling: the two cultures. Statist Sci. 2001;16:199–231. Ready to submit your research ? Choose BMC and benefit from: • fast, convenient online submission • thorough peer review by experienced researchers in your field • rapid publication on acceptance • support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations • maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions