The role of machine learning in decoding the molecular complexity of bovine pregnancy: a review
Marilijn van Rumpt A and M. Belen Rabaglino A *A
Abstract
Pregnancy establishment and progression in cattle are pivotal research areas with significant implications for the industry. Despite high fertilization rates, ~50% of bovine pregnancies are lost, pinpointing the need to keep studying the biological principles leading to a successful pregnancy. The increasing access to and generation of omics data have aided in defining the molecular characteristics of pregnancy, i.e. embryo and fetal development and communication with the maternal environment. Large datasets generated through omics technologies are usually analyzed through pipelines that could lack the power to deeply explore the complexity of biological data. Machine learning (ML), a branch of artificial intelligence, offers a promising approach to address this challenge by effectively handling large-scale, heterogeneous and high-dimensional data. This review explores the role of ML in unraveling the intricacies of bovine embryo–maternal communication, including the identification of biomarkers associated with pregnancy outcome prediction and uncovering key genes and pathways involved in embryo development and survival. Through discussing recent studies, we define the contributions of ML towards advancing our understanding of bovine pregnancy, with the final goal of reducing pregnancy losses and enhancing reproductive efficiency while also addressing current limitations and future perspectives of ML in this field.
Keywords: bovine, embryo, endometrium, epigenomics, fetus, machine learning, metabolomics, molecular data, omics technologies, pregnancy, pregnancy outcome prediction, transcriptomics.
Introduction
Understanding the complex mechanisms underlying successful embryogenesis and fetal growth is crucial for the dairy and beef industry, impacting economic efficiency and genetic improvement programs. Embryo mortality is a significant factor affecting reproductive success. Despite achieving high fertilization rates upwards of ~80%, birth rates significantly lag behind, with most embryos lost within the first month of pregnancy (Berg et al. 2010; Wiltbank et al. 2016; Reese et al. 2020). The primary reasons for this discrepancy are diverse, such as genetic factors, poor oocyte or embryo quality, suboptimal uterine receptivity, inadequate embryo–uterus interaction, failure of conceptus elongation, and abnormal fetal growth (Lonergan et al. 2016; Wiltbank et al. 2016).
Pregnancy is a complex process following tightly regulated steps. After mating or artificial insemination (AI) at estrus, fertilization of the oocyte occurs in the oviduct (day 0), followed by mitotic cleavages of the embryo. These cleavages result in a 16-cell stage, developing into a morula, which enters the uterus by day 4. The embryo remains transcriptionally silent during these first cleavages and relies on maternal mRNA (Schulz and Harrison 2019). The ‘maternal-to-embryonic transition’ occurs around the 8- to 16-cell stage (Frei et al. 1989; Kopecny et al. 1989). It involves degradation of maternal transcripts and activation of the embryonic genome (Memili and First 2000). By day 6 to 7, the morula develops into a blastocyst, containing an inner cell mass, which will become the fetus, and outer trophectoderm cells, which will give rise to the placenta. At this point, the blastocyst interacts with the endometrium by activating local genes in the uterus favoring an optimal intrauterine environment (Sponchiado et al. 2017; Passaro et al. 2018). At day 9 to 10, the blastocyst hatches from the zona pellucida. The embryo starts changing its morphology on day 12 to 14 from an ovoid to a tubular and finally a filamentous conceptus by day 16 to 17 (Degrelle et al. 2005). The elongation and surface increase are required for elevating interferon tau (IFNT) secretion by the embryo to trigger the signal for maternal pregnancy recognition and thus prevention of luteolysis to maintain progesterone (P4) production by the corpus luteum. Trophectoderm cells start attachment to the uterine epithelium by day 16, initiating placentation at approximately day 20 (Assis Neto et al. 2010).
The first 7 days of gestation represent the period of most embryo loss (Sartori et al. 2010; Reese et al. 2020). This period is bypassed in case of embryo transfer (ET), which involves transferring a 7-day-old fertilized embryo to the uterus of a recipient cow. Contradictorily, pregnancy success is generally not higher when ET is used, compared to AI (Hansen 2020). Furthermore, there is a difference in pregnancy rates between embryos generated in vivo and in vitro, with the latter having an average pregnancy rate of ~25% lower, according to the outcomes of studies done between 1992 and 2014 (Ealy et al. 2019). In vivo produced embryos are usually obtained by ovarian stimulation, resulting in multiple ovulations, after which embryos are collected and transferred to the recipient (MOET). For in vitro produced (IVP) embryos, oocytes are either retrieved from ovaries after slaughter or collected by transvaginal aspiration of follicles in a living cow, followed by maturation and fertilization of the cultured oocyte. Another reason likely to be responsible for a significant amount of embryo loss is the failure of the conceptus elongation process (Moraes et al. 2018; Sánchez et al. 2019), which is not avoided by ET, as it is completely maternally driven and has not been reproduced in vitro.
As both AI and ET pregnancy successes are suboptimal, research has focused on uncovering mechanisms involved in determining embryo quality and production (in vivo vs IVP) and endometrial receptivity in pregnancy losses. With the rising availability and collection of omics data, progress has been made in defining the molecular characteristics of pregnancy, i.e. embryo and fetal development and communication with the maternal environment. Transcriptomic analyses of the uterus revealed a different response to short versus long conceptuses (Sánchez et al. 2019), and to in vivo versus IVP embryos (Mathew et al. 2019), suggesting a sensitive uterine response to embryos with varying developmental competence (Bauersachs et al. 2009; Mansouri-Attia et al. 2009). Similarly, the embryonic transcriptome differs between embryos able to sustain a pregnancy or not, and between MOET versus IVP embryos (Rabaglino 2023). Proteomic studies on early embryos have uncovered differently expressed proteins in each developmental stage (Deutsch et al. 2014; Banliat et al. 2022), and metabolomics research has shown different metabolic profiles before ET between pregnant and open cows (Gómez et al. 2020b; Gimeno et al. 2023). Even though these high throughput data have advanced reproductive research, it is hard to explore relationships hidden in these noisy, complex and high-dimensional datasets using traditional methods, such as differential analysis. In this context, the emergence of machine learning (ML) offers a promising approach to further explore the biological molecules involved in a successful pregnancy in the cow, and to complement traditional methods, considering its ability to handle large-scale, heterogenous and high-dimensional data.
Machine learning, a branch of artificial intelligence, encompasses algorithms that enable computers to learn from data patterns and make predictions without explicit programming. ML can be divided into two broad classes: supervised learning and unsupervised learning. In supervised learning, the dataset is randomly divided into training and validation datasets with labeled data such as pregnancy outcomes. The supervised model trains on the dataset and learns to recognize patterns associated with the label and test its predictive ability to assign correct labels on new data in the validation dataset. This is mainly used for problems in classification (in case of discrete labels) or regression (in case of continuous label values), on which the model ranks the sample features according to their importance in predicting the label. One practical application of supervised learning is in biomarker discovery, where, for example, models are trained to predict pregnancy outcomes in cattle based on metabolites or transcripts and select the top predictive markers (Rabaglino and Kadarmideen 2020; Gómez et al. 2021; Gimeno et al. 2023; Rabaglino et al. 2023a; Hoorn et al. 2024). In contrast, unsupervised models operate without labeled data, identifying data patterns and searching for similarities between data samples, with the main goal of performing clustering or dimensionality reduction. Unsupervised methods such as principal component analysis (PCA) and hierarchical clustering have been broadly applied in the past decades in reproductive science. The use of supervised ML, however, has surged in recent years, with a growing application in evaluating and predicting embryonic viability and potential interaction with the maternal environment, with the ultimate goal of reducing pregnancy losses.
This review aims to explore the role of supervised ML applied to omics data in understanding the intricacies of bovine embryonic and fetal development, and how ML may be implemented to predict the risk of pregnancy losses. We first briefly introduce commonly used supervised ML models, followed by discussing their applications in recent studies in the field of bovine reproductive biology. The studies are organized into those conducted before day 7 (before the first interactions between embryo and uterus) and after day 7 of gestation. Finally, we discuss the limitations, current challenges, and perspectives of ML in studying pregnancy in cattle.
Supervised machine learning models
This section introduces the supervised ML models applied in the articles discussed in the next section.
Decision tree (DT), random forest (RF) and extreme gradient boosting
A DT is a flowchart-like structure of an inverted tree, consisting of a root, nodes, branches and leaves. The hierarchical nodes represent a series of feature tests connected by branches, with the root being the first test and the leaves being the final class labels (Kingsford and Salzberg 2008). The aim of using DT analyses is creating the best model for allocating all samples into the right segment. After learning the best DT from the training dataset, new instances are passed through the tree to predict their class.
For RF (Breiman 2001), multiple hundreds to thousands of DTs are ensembled, each grown on a smaller random subset of samples and features of the original dataset. Hence, each DT will differ slightly. A new sample is passed down all the trees and its class is predicted by the majority vote. RF methods work well with high dimensional data, having many predictor variables, even when the number of samples is small (‘large p, small n’) (Chen and Ishwaran 2012). RF often performs with higher accuracy compared to DT, lowering the prediction variance and bias associated with a single DT. Furthermore, RF is less prone to overfitting and is efficient in estimating missing values. Among other models used for classification tasks, RF takes a top position in predictive accuracy (Fernández-Delgado et al. 2014). However, interpreting a large RF may be difficult, and they can be computationally expensive when working with large datasets.
Another method based on DTs is extreme gradient boosting (XGboost), which builds multiple decision trees sequentially, with each new tree correcting the errors made by the previous ones (Chen and Guestrin 2016). This boosting technique focuses on optimizing the gradient of the loss function, leading to highly accurate and efficient models, while regularization is also incorporated to avoid overfitting.
K-nearest neighbor (KNN)
For KNN models, the distances between samples in the (high dimensional) feature space are calculated, after which the class of a new sample is predicted by considering the class of the majority of its closest neighbor samples (Mucherino et al. 2009). KNN may be computationally expensive, mostly because of the distance calculations. The value of k (number of evaluated neighbors) and the distance method used are important for the model’s performance. If k is too small, noise and outliers can have negative effects, resulting in overfitting and high variance. However, if k is too large, overrepresented classes can overwhelm smaller classes, resulting in bias and underfitting. KNN is severely affected by ‘the curse of dimensionality’ (phenomena that arise when analyzing data in high-dimensional spaces) and its accuracy decreases if more sample features (or dimensions) are used, as all datapoints tend to be far away, without meaningful neighbors (Seidl 2009).
Bayesian network (BN)
A BN is a probabilistic graphical model, specifically a ‘directed acyclic graph’, consisting of nodes depending on each other, connected with edges that have direction, but do not form a loop (Stephenson 2000). Each node represents a variable and has a conditional probability attached, representing the chance of the node being in a given state, influenced by the given state of its parents. The node dependencies propagate through the network, influencing the probabilities of other nodes. However, BN assumes that, given its parents, a node is conditionally independent of its non-descendants. BNs can work well with missing data and small datasets, preferring small sets of parent variables. They are less suited for large datasets and data with feedback loop relationships affecting the outcome.
Partial least squares discriminant analysis (PLS-DA)
PLS-DA is a tool for dimensionality reduction but is also widely applied for classification. High dimensional data is projected into a lower dimensional space, preserving distances between samples. Unlike PCA, which identifies principal components that explain the highest variance in the data, PLS-DA focuses on maximizing the covariance between the data and class labels, making it a ‘supervised’ version of PCA (Ruiz-Perez et al. 2020). This approach allows PLS-DA to effectively handle high-dimensional data, but it can be susceptible to overfitting, necessitating the use of cross-validation. In sparse PLS-DA (sPLS-DA), a variant of PLS-DA, a sparsity constraint is incorporated, selecting only the most informative variables for class prediction through using penalties (Lê Cao et al. 2011). This makes sPLS-DA particularly useful for identifying key features in complex datasets while maintaining model simplicity and interpretability.
Support vector machine (SVM)
The goal of SVM is to find the optimal hyperplane that separates the datapoints into different classes while maximizing the margin between classes (Pisner and Schnyer 2020). SVM uses the training data to define this optimal hyperplane, and when separation cannot be done linearly, a kernel method is used to transform and map the data into a high-dimensional space where a simple hyperplane separates the data. Finding the optimal hyperplane can be computationally costly. A soft margin boundary is used that allows misclassification and lowers the effect of outliers on the boundary. SVM works effectively with high-dimensional data and in cases where the number of dimensions exceeds the number of samples but is less suited for noisy data and when datapoints with different classes are overlapping.
Artificial neural networks (ANN)
ANNs are inspired by the structure and function of the neural networks in the human brain and consist of connected hierarchical layers of neurons. Each neuron receives connection weights of other neurons and a bias, which are summed up, after which an activation function (often sigmoid) is applied to determine if the neuron gets activated (Priddy and Keller 2005). ANN consists of an input layer where the raw data enters, one or more hidden layers, and a final output layer performing the last prediction. The output of a neuron of one layer is the input to neurons in the next layer. During training of the ANN, the connection weights are updated, with the goal to make the output of the last layer as close as possible to the desired output. The updating of biases and weights is often done by backpropagation: propagating the error back through the neurons. ANNs are difficult to interpret, can be prone to overfitting, are computationally costly and often require a lot of training data. They perform better when the dataset is larger and are good at modeling complex and non-linear relationships.
Machine learning to study embryonic or maternal molecular features before day 7 of gestation
As most embryonic losses occur in the first week of pregnancy (Reese et al. 2020), even before uterus–embryo interaction, much research applying the predictive abilities of ML has focused on this critical period (see Table 1). Successful pregnancy requires both a competent embryo and a receptive endometrium. Hence, both embryonic and maternal omics are deployed for understanding early pregnancy.
| Reference | Goal of ML | ML methods | Data type | Data set | Result/highest performance | |
|---|---|---|---|---|---|---|
| Rabaglino et al. (2023a) | Biomarker discovery | sPLS-DA | Embryo transcriptomics | 89 blastocysts, 48 elongated conceptuses | 8 biomarkers | |
| Pregnancy outcome prediction | BLR, ANN | >85% accuracy | ||||
| Hoorn et al. (2024) | Biomarker discovery | RF (BORUTA), sPLS-DA | Endometrium transcriptomics | 193 endometrial cytobrush samples | 57 biomarkers | |
| Pregnancy outcome prediction | SVM | 77% accuracy | ||||
| Diniz et al. (2022) | Biomarker discovery | Ordinal learning method and models based on SVM, KNN | Endometrium transcriptomics | 43 endometrial samples | 9 biomarkers | |
| Pregnancy outcome prediction | 53.85–61.54% (using nine genes) and >80% accuracy (using 225 genes) | |||||
| Rabaglino and Kadarmideen (2020) | Biomarker discovery | Models based on BN and logistic regression | Endometrium transcriptomics | 52 endometrial samples | 50 biomarkers | |
| Pregnancy outcome prediction | SVM | 96.1% accuracy | ||||
| Moorey et al. (2020) | Pregnancy outcome prediction | Parallelized RF | White blood cell transcriptomics | 23 blood samples, 198 genes | >90% accuracy | |
| Muñoz et al. (2014b) | Pregnancy outcome prediction | KNN | Embryo CM and blood metabolomics | 49 MOET embryos | 64.4% accuracy | |
| 49 blood samples | 74.2% accuracy | |||||
| Muñoz et al. (2014a) | Pregnancy outcome prediction | KNN | Embryo CM and blood metabolomics | 69 IVP fresh or VW embryos | 71.9% accuracy | |
| 69 blood samples | 74.9% accuracy | |||||
| Gómez et al. (2020b) | Biomarker ranking | RF | Blood metabolomics | 67 blood samples (Holstein) | NA | |
| Gómez et al. (2020a) | Biomarker ranking | RF | Blood metabolomics | 74 blood samples (Asturiana de los Valles) | NA | |
| Gómez et al. (2021) | Biomarker ranking | SVM, PLS-DA, RF | Embryo CM and blood metabolomics | 36 VW embryos, 36 blood samples | NA | |
| Gimeno et al. (2023) | Biomarker ranking | RF | Embryo CM and blood metabolomics | 70 IVP fresh or FT embryos, 107 blood samples | NA |
ML, machine learning; (s)PLS-DA, (sparse) partial least squares discriminant analysis; BLR, Bayesian logistic regression; ANN, artificial neural network; RF, random forest; SVM, support vector machine; KNN, k-nearest neighbor; BN, Bayesian network; CM, culture media; MOET, multiple ovulation embryo transfer; IVP, in vitro produced; VW, vitrified–warmed; FT, frozen–thawed; NA, not applicable
Machine learning in embryo transcriptomics
Embryo quality evaluation and selection is conventionally done by visual inspection of morphological features by embryologists, grading the embryos generally following the guidelines from the International Embryo Technology Society (IETS) (Bó and Mapletoft 2018). However, this method is subjective and lacks accuracy and repeatability (Hansen 2020). Combining embryo transcriptomics with ML offers a more precise approach to predicting embryonic competence and identifying molecular differences between viable and non-viable embryos.
Based on the knowledge that an inherent molecular signature defines the bovine embryo ability for pregnancy establishment, we integrated seven transcriptomic datasets of both pre-ET blastocyst biopsies and elongated conceptuses of varying competence to identify biomarkers predictive of pregnancy outcome (Rabaglino et al. 2023a). Differently expressed genes (DEG) were identified between competent blastocysts (blastocysts resulting in pregnancy and long conceptuses) and incompetent blastocysts (blastocysts not resulting in pregnancy and short conceptuses), after which eight biomarker genes most discriminative between pregnant and non-pregnant samples were selected by sPLS-DA and linear discriminant analysis. The predictive ability of these biomarkers was tested on an independent dataset, consisting of competent embryos (resulting in pregnancy, cultured in normal conditions, or long conceptuses), and incompetent embryos (not resulting in pregnancy, cultured in suboptimal conditions, or short conceptuses). Using Bayesian logistic regression (BLR) and ANN, prediction accuracies from 85 to 100% were achieved, depending on the validation dataset, with ANN having the same or higher accuracies than BLR. Upregulated biomarkers in competent embryos were involved in cellular metabolism, including glycolysis/gluconeogenesis, whereas downregulated biomarkers were related to cell cycle processes (Rabaglino et al. 2023a). Glucose metabolism is known to be a critical process in embryo survival, with glucose deprivation leading to apoptosis (Riley and Moley 2006). Hence, activation of glycolysis/gluconeogenesis pathways could have a positive effect on embryo development. The downregulation of cell cycle processes in competent embryos is in agreement with the ‘quiet embryo hypothesis’, which states that viable embryos have a less active metabolism (Leese 2002). Furthermore, we have recently developed a formula to estimate an embryonic competence index based on the expression of the identified eight biomarker genes, which is available as a function for the R software (Rabaglino and Hansen 2024). Estimation of a quantitative index value can be employed in experiments to objectively identify interventions in embryo production that could increase embryo survival after transfer.
Machine learning in maternal transcriptomics
Transcriptomic data of the endometrium before embryo interaction have been used to uncover different gene expression profiles between cows that will become pregnant or not. Hoorn et al. (2024) identified biomarkers capable of predicting pregnancy status at day 30 through analysis of the transcriptome of endometrial cells of Holstein cows at day 0 before AI. For this, we applied the BORUTA algorithm (a wrapper built around the RF algorithm; Kursa and Rudnicki 2010) to identify biomarkers among transcripts differently expressed between cows that became pregnant or not, after which sPLS-DA was used to determine the combination of transcripts most discriminative for pregnancy status at day 30. Transcript combinations were evaluated by applying SVM with linear kernels, which resulted in a set of 57 transcripts with an average prediction accuracy of 77%. The functional analysis of these biomarkers indicated that uterine immunological condition may be important for maternal fertility, with cows experiencing less immune activation being more likely to become pregnant (Hoorn et al. 2024).
Similar research was performed by Diniz et al. (2022), using endometrial transcriptomics data collected 3 days before ET from Angus-Brahman crossbred cows. Using BioDiscML (software that automizes ML steps in feature and model selection in omics data; Leclercq et al. 2019), 225 genes and five ML models based on SVM, KNN and an ordinal learning method were selected for further biomarker identification and day 30 pregnancy status prediction. Prediction accuracies on all 225 genes ranged from 80% to higher than 90%, while prediction accuracies using nine genes selected as potential biomarkers ranged from 53.85% to 61.54%, with the models based on SVM and KNN having the greatest accuracies. The lower prediction ability of the nine genes is likely due to the smaller dataset size and only nine genes not being able to reflect the complex process involved in pregnancy loss or success. Contrary to the study of Hoorn et al. (2024), immune pathways were not found to be significantly associated with analyzed genes; however, one of the nine selected biomarker genes, PDCD1, is likely to be involved in immune inhibition during pregnancy (Taglauer et al. 2008; Diniz et al. 2022). Other pathways related to the nine biomarker genes were focal adhesion, remodeling of endometrial tissue and embryonic development.
We conducted another study focused on predicting pregnancy outcome using day 6–7 endometrial transcripts from cows of four different European cattle breeds (Rabaglino and Kadarmideen 2020). Using five ML models, of which three were based on BN and two on logistic regression, 50 genes overlapping between methods were selected as potential biomarkers. An average accuracy of 96.1% was achieved, by applying SVM to predict pregnancy outcome with these biomarkers, training on all samples from all but one breed, and using the left-out breed for validation. Among pathways in which upregulated genes related to the biomarkers were involved are embryonic development, circadian rhythm and Wnt pathways.
The three studies discussed above all identified different endometrial biomarkers for pregnancy outcome prediction (except for one biomarker in common between two studies). A potential factor influencing this lack of overlap among the biomarkers is the timing of sampling, since samples were collected on day 0, 4 and 6–7, according to the study. Transcriptomic research has shown differences in endometrial gene expression at different timepoints in the periestrous period, likely through fluctuations in estradiol and progesterone (Alfattah et al. 2024), which may lead to different biomarkers associated at each timepoint. Furthermore, cow breeds differed between the studies, which may affect biomarker selection. Only the study of Rabaglino and Kadarmideen (2020) used multiple breeds for training and validation of the ML model. Thus, the identified biomarkers can be used for pregnancy outcome prediction across different breeds, avoiding the selection of breed specific biomarkers.
A different approach for pregnancy outcome prediction applied by multiple studies is the use of blood omics data. This has the benefit of blood being easily available and collection being minimally invasive. Moorey et al. (2020) used peripheral white blood cell RNA at day 0 to predict pregnancy outcome. Based on 198 DEG between pregnant and non-pregnant cows, a prediction accuracy of >90% was achieved using a parallelized RF. More blood-based research has focused on metabolomics, as discussed in the next section.
Machine learning in maternal and embryonic metabolomics
Several studies have analyzed metabolites from recipient blood and/or culture medium (CM) of embryos used for ET, providing insights into the metabolic status during early pregnancy and embryonic development. Blood and CM have the advantage of being easily accessible and minimally invasive, compared to endometrial samples. Furthermore, using CM for embryo assessment does not interfere with embryo developmental competence, unlike embryonic biopsy. However, CM composition might be impacted by external factors such as contact with plastic dishes or variations in temperature, pH and osmolarity (Sciorio and Rinaudo 2023). Therefore, ML models based on CM metabolites might not perform similarly among different laboratories, as mentioned below.
Muñoz et al. (2014b) found recipient blood metabolites to have a higher predictive accuracy for day 60 pregnancy compared to CM metabolites of MOET embryos when employing KNN for prediction. The study also revealed variability in pregnancy status prediction when CM samples were processed in two separate laboratories. Using the cumulative CM data from both laboratories, a predictive accuracy of 64.4% was achieved, while the separate accuracies were 74.6% and 74.8%. This difference might be due to divergent laboratorial procedures altering embryo-produced metabolites or their measurement, indicating that the used training dataset may alter predictions and must be selected carefully, considering the origin of the data. In contrast, predictive accuracies based on blood were highest when using cumulative data compared to separate data (74.2% versus 59.6% and 69.1%). A second study performed by Muñoz et al. (2014a) also showed higher prediction accuracies for blood metabolites compared to CM metabolites. Metabolites in CM of fresh and vitrified–warmed (VW) embryos, and recipient blood, were analyzed by employing a similar approach using KNN for pregnancy outcome prediction. Birth prediction accuracies based on fresh embryo CM and blood of fresh embryo recipients reached 71.9% and 74.9% respectively. Interestingly, increased birth accuracies were obtained when only expanded blastocysts were considered (CM: 82.8%, blood: 85.0%), so the CM of more developed embryo stages might contain a metabolite profile with higher predictive ability (Muñoz et al. 2014a). For VW embryos, only metabolites in blood plasma and not in CM were able to give a relevant prediction with 69.3% accuracy for birth. Considering the higher achieved prediction accuracies for fresh embryos compared to VW embryos, the embryo treatment likely affects the embryo’s metabolome.
Further underpinning the possible effect of embryo treatment on its metabolome and pregnancy outcome prediction, Gómez et al. (2020b) revealed distinct blood biomarkers in cows that received differently produced embryos. RF was applied for ranking potential metabolite biomarkers based on their importance in pregnant and non-pregnant classification for Holstein cows having either fresh or VW embryos transferred. Metabolite ranking varied between different ET embryos, with ornithine and oxoglutaric acid being among top metabolites for fresh embryos, and L-glutamine and L-lysine being among top metabolites for VW embryos. The same approach in a study with Asturiana de los Valles (AV) also ranked oxoglutaric acid highly for pregnancy outcome prediction in recipients receiving fresh embryos, whereas dimethylamine and 2-hydroxybutyric acid were top metabolites for VW embryos (Gómez et al. 2020a). In both studies, only metabolites from day 7 and not day 0 had predictive ability for pregnancy outcome. Oxoglutaric acid, an intermediate in the Krebs cycle and an antioxidant, was found to have positive effects on embryo development in mice, in which oxogluratic acid treatment of embryos before ET increased blastocyst development rate and fetal growth (Zhang et al. 2019). If oxoglutaric acid blood levels reflect its uterine levels, this metabolite may positively affect embryo development and survival in the bovine uterus.
Taking a different approach, Gómez et al. (2021) applied SVM, PLS-DA and RF on CM of VW embryos and recipient blood metabolomics for feature ranking and pregnancy outcome prediction. Among high-ranking features in CM were metabolites involved in lipid metabolism (stearic, capric and palmitic acid), with higher levels of non-esterified saturated fatty acids (NEFA) in the CM of non-viable embryos. NEFA exposure alters the epigenetic and transcriptomic profiles of oocytes and blastocysts (Desmet et al. 2016), and negatively affects embryo viability, shown by factors such as decreased cell number, elevated apoptosis and altered metabolism (Van Hoeck et al. 2011). Top ranking amino acids in recipient blood were similar to previous findings in Holsteins and AV (Gómez et al. 2020a, 2020b). Furthermore, NEFAs were among high ranked blood metabolites in AV, with lower levels among pregnant cows (Gómez et al. 2021). NEFAs are released into the blood during a negative energy balance (NEB), and its concentrations are related to the depth of NEB. Besides the negative effect of NEFAs on embryos, a negative influence also exists on endometrial cells, causing decreased cell viability and elevated levels of pro-inflammatory cytokines (Chankeaw et al. 2018).
In the most recent study, pregnancy outcome prediction and biomarker ranking for fresh and frozen–thawed (FT) embryos was done by Gimeno et al. (2023). They also applied SVM, PLS-DA and RF, using multiple iterations to reevaluate misclassified samples and improve the model, on both blood metabolites and embryo CM. More recipients than embryos were shown to be competent, which can lead to misclassifying viable recipients because they were matched with non-viable embryos. Taking viability of embryos into consideration during revaluation of recipients improved pregnancy outcome predictions. In concordance with Gómez et al. (2020b), L-glutamine and L-glycine were top ranked recipient biomarkers. For CM biomarkers, higher prediction ability was achieved with FT embryos than fresh embryos, and top biomarkers varied between fresh and ET embryos (Gimeno et al. 2023), further supporting production procedures of embryos affecting their metabolomes.
Machine learning to study embryonic, fetal or maternal molecular features after day 7 of gestation
Endometrium–embryo crosstalk starts playing an important role in embryo survival from the very first moments of interaction. The focus of research on embryo development from day 7 of gestation has not been pregnancy outcome prediction, because embryo or recipient selection needs to be done before this moment, as ET occurs at day 7. Instead, research has focused on understanding the molecular profile of both embryo and uterus and what processes are key to pregnancy maintenance during this period (see Table 2).
| Reference | Goal of ML | ML methods | Data type | Data set | Result/highest performance | |
|---|---|---|---|---|---|---|
| Rabaglino et al. (2021) | Prediction of type of embryo production (IVP vs MOET) | SVM | Embryo transcriptomics and epigenomics | 34 blastocysts and 40 conceptuses, 188 DEG/DMG | 90–100% accuracy | |
| Talukder et al. (2023) | Identification of ISG | RF (BORUTA) | Endometrium transcriptomics | 53 endometrial explants | 54 ISGs | |
| O’Callaghan et al. (2022) | Biomarker ranking | sPLS-DA | Endometrium transcriptomics | 32 endometrial explants | 200 genes | |
| Prediction of bull fertility status | SVM | 90% accuracy | ||||
| Rabaglino et al. (2023b) | Prediction of fetal weight | Extreme gradient boosting (XGboost) | Blood and fetal organ (heart, liver, gonads) transcriptomics | 10 blood samples, 8–9 samples per fetal organ, 35 genes | 0.4 root-mean-square error |
ML, machine learning; IVP, in vitro produced; MOET, multiple ovulation embryo transfer; (s)PLS-DA, (sparse) partial least squares discriminant analysis; RF, random forest; SVM, support vector machine; DEG, differently expressed genes; DMG, differently methylated genes; ISG, interferon-stimulated genes.
Machine learning in conceptus omics
More IVP embryos than MOET embryos have been transferred worldwide during the last eight years, with the last IETS report showing a significant divergence: 1,189,699 transferred embryos were IVP while 368,783 were produced in vivo (Viana 2022). However, as mentioned in the introduction, IVP embryos achieve lower pregnancy rates compared to MOET embryos (Ealy et al. 2019). To explore the molecular profile induced by the lack of the oviductal and uterine environment, regardless of variables such as the maturation media, technical procedures, or parental characteristics, we performed a meta-analysis of six transcriptomic and four epigenetic datasets from day 7 blastocysts to day 13 and day 16 conceptuses to define temporally DEG and differently methylated genes (DMG) between IVP and MOET embryos (Rabaglino et al. 2021). A SVM model was trained on conceptus expression data to predict the type of embryo production, which was validated on independent conceptus datasets. This test dataset consisted of two datasets, the first one consisting of transcriptomic data of short and long conceptuses, and the second dataset consisting of conceptuses of embryos treated or not with Dickkopf-related protein (DKK1, Wnt signaling inhibitor). It was hypothesized that short and untreated conceptuses would be predicted as IVP embryos and long and DKK1-treated embryos be predicted as MOET embryos.
Prediction accuracies of 90–100% were achieved using a cluster of 188 DEG/DMG, in which the gene expression profiles showed a clear difference at day 13 between IVP and MOET embryos, while using only DEG resulted in lower accuracies of 70%, underscoring the power of multi-omics analysis. A significantly related pathway to the DEG/DMG cluster was focal adhesion, including the ‘extracellular exosome’. Adhesion is important for connection of cells with their environment and the organization of the cytoskeleton within the embryo (Shawky and Davidson 2015), invasion of the embryo in the endometrium, and trophectoderm–endometrium communication at implantation (Kaneko et al. 2008, 2012). Additionally, exosomes, vesicles secreted by either the endometrium or the embryo, carrying bioactive molecules, play key roles in maternal pregnancy recognition, and uterus–embryo signaling in the peri-implantation period (Bridi et al. 2020). Given the different expression of key pathways in IVP and MOET embryos, and short conceptuses being predicted as IVP embryos, the genes important for conceptus elongation and focal adhesion may be de-regulated in IVP embryos, possibly affecting their viability (Rabaglino et al. 2021).
Machine learning in maternal transcriptomics
From day 7, IFNT secretion by the embryo poses an important way of interaction with the endometrium, altering gene expression and changing the uterine environment needed for maternal pregnancy recognition and prevention of luteolysis. Talukder et al. (2023) used the BORUTA algorithm to identify the most prominent interferon-stimulated genes (ISG), using transcriptomic data from day 15 endometrial explants exposed to nothing, IFNT or a conceptus. This resulted in 54 ISG being upregulated around the time of maternal pregnancy recognition, of which the majority are related to immunity regulation. Maternal immunomodulation is needed for controlling innate immune responses and avoiding rejection of the allogenic embryo. IFNT likely plays an important role in this process, creating a uterine environment suitable for pregnancy establishment (Rocha et al. 2021).
Less well researched is the contribution of the sire in conceptus-induced changes in the uterine environment. Fertility can vary significantly among bulls, which may greatly impact pregnancy rates as one bull might fertilize thousands of cows. O’Callaghan et al. (2022) compared the transcriptomic profiles of endometrial explants exposed to conceptuses conceived with semen from high fertile and low fertile bulls. We used three methods for selection of genes differing between fertility classes: identification of DEG, co-expression network analysis, and lastly sPLS-DA was applied to identify genes most discriminative between bull fertility classes. Subsequently, SVM was applied on selected genes to train a model for fertility status prediction. The highest accuracy of 90% was achieved using 200 genes selected by sPLS-DA. Upregulated genes predictive of conception with a high fertile bull were mostly involved in immune regulation, further underscoring the importance of proper immune regulation for survival of the peri-implantation embryo.
Machine learning in fetal transcriptomics
A topic further explored in human research is the use of maternal blood omics and ML to predict normal pregnancy or future pregnancy complications (Camunas-Soler et al. 2022; Rasmussen et al. 2022; Xiong et al. 2022), and prediction of gestational stage (Ngo et al. 2018). With the use of ML in the bovine field being mostly applied for predictions in the early stage of pregnancy, predictive studies involving fetal development are scarce. To shed light on this area, we explored the use of maternal blood transcriptomics for predicting fetal weight, by identifying co-expressed overlapping genes between the maternal blood and day 42 fetal organs (heart, liver, gonads) positively correlated to fetal weight (Rabaglino et al. 2023b). The overlapping genes between maternal blood and each organ were used for training a regression model, applying extreme gradient boosting (XGboost), after which the model was tested using the same genes in maternal blood. The most effective training dataset consisted of 35 genes overlapping between fetal heart and blood, achieving a root-mean-square error of 0.4. Furthermore, variance in fetal heart genes explained ~93% of gene expression variance in the maternal blood. The 35 selected genes enriched ontological terms related to energy metabolism processes, including oxidative phosphorylation. These results showed a relationship between the molecular profile of the developing fetal heart and fetal weight, which can also be measured in and is associated with transcripts in the maternal blood.
Discussion
Machine learning has been increasingly utilized in a wide range of medical and biological applications, demonstrating its significant value. While its application in cattle reproduction and fertility is relatively recent, ML has already shown its potential in enhancing our understanding of the biological mechanisms driving bovine pregnancy establishment. ML can play a crucial role in reducing pregnancy loss through its predictive capabilities, identifying relevant biomarkers for maternal receptivity and embryonic viability.
There is no universally optimal ML model for every situation. For ML method selection, the type and amount of data needs to be considered. Often, multiple models may be suitable for a given problem, and testing several models can help determine the most effective one. This approach allows for the comparison of their performance, and, in the context of biomarker identification, which biomarkers are selected and their relevance. Among the discussed articles, SVM and KNN were the most often used models for class prediction, and only one study applied a deep learning approach through ANN. While the use of deep learning is a hot topic in multiple fields, classical ML models have shown to often outperform deep learning methods when applied to tabular data (e.g. transcriptomics, metabolomics) (Eraslan et al. 2019). Furthermore, considering the small sample sizes in the discussed articles, deep learning methods may not have been a good fit as they generally perform better with large training data sets. However, we obtained high prediction accuracies (>85%) for embryo competence using ANN (Rabaglino et al. 2023a). Additionally, ANN outperformed classical ML methods in a study predicting embryo implantation in humans, achieving 100% accuracy (Cheredath et al. 2023).
The basis of creating a good predictive model is the selection of an appropriate training dataset. Low prediction accuracies can often be traced back to noisy or mislabeled data, affecting performance. Specifically for predicting pregnancy outcome, correct labeling of data is challenging. Only the pregnancy outcome can be measured, but the exact cause of embryo mortality may not be known, as it can be due to an incompetent embryo and/or an unreceptive endometrium. Thus, when predicting pregnancy outcome based on embryo omics, an embryo may be wrongly classified as incompetent, while the cause for embryo mortality was maternal. Reevaluating misclassified samples in multiple iterations during training improved the model’s performance in the study by Gimeno et al. (2023). Additionally, invasive data collection methods such as embryo biopsy may compromise embryo competence (Ponsart et al. 2013), possibly resulting in a competent embryo being mislabeled as non-competent. Furthermore, the training dataset affects the generalizability of the model, and thereby its practical applicability. When one wants to be able to apply a ML model on varying external datasets, variance also needs to be included in the training dataset, lowering the risk of overfitting on a specific dataset with a specific origin, considering that prediction accuracies and/or identified top biomarkers may differ between breeds (Gómez et al. 2020a, 2020b; Rabaglino and Kadarmideen 2020), laboratories (Gómez et al. 2020b) and embryo production techniques (Gómez et al. 2020a; Gimeno et al. 2023).
It has been suggested that maternal competence has a higher variability compared to embryo viability, with differing pregnancy outcomes being more associated with varying recipient competence (McMillan 1998; McMillan and Donnison 1999). This would imply that using maternal data could better predict pregnancy outcome. Supporting this, metabolomics studies showed higher pregnancy success prediction accuracies for blood metabolites compared to CM (Muñoz et al. 2014a, 2014b). Additionally, Gimeno et al. (2023) identified more competent recipients compared to embryos. Employing maternal omics may, thus, be a more effective approach in predicting pregnancy success and lowering embryo losses. A highly fertile cow will not become pregnant if matched with an incompetent embryo, and relevant prediction accuracies have also been achieved using embryo omics (Muñoz et al. 2014a, 2014b; Rabaglino et al. 2023a), indicating that the use of both maternal and embryo omics are relevant for improving fertility rates.
Not discussed in this review is the influence of sire fertility on embryo survival. Fertility varies significantly among bulls, resulting in different field fertility rates, so the importance of sperm quality for pregnancy establishment should not be ignored. While routine sperm evaluations can detect substantially low fertile bulls, identifying sub-fertile bulls producing apparently normal sperm remains a challenge, with no accurate diagnostic test available (Kastelic and Thundathil 2008). Hence, the use of ML to aid in fertility status prediction has also surged in this field in recent years. ML methods applied to multiple omics types and sperm variables have been utilized for predicting bull fertility (Bucher et al. 2019; Costes et al. 2022; Rabaglino et al. 2022; Costes et al. 2024) and embryo yield (Campanholi et al. 2023). Bull fertility is not only determined by the semen’s ability to fertilize an oocyte but also by its impact on embryonic characteristics. Top predictive features for bull fertility were found to be related to embryonic development (Costes et al. 2024), and the endometrial transcriptome differed after being exposed to conceptuses conceived with semen from high or low fertile bulls (O’Callaghan et al. 2022). However, the true influence of sire fertility on embryo survival and identification of affected pathways requires further research.
Currently, embryo quality classification relies on visual inspection of microscopic images, following IETS grading guidelines (Bó and Mapletoft 2018). Although higher pregnancy rates are achieved with high-quality graded embryos (Farin et al. 1999), this method suffers from subjectivity and low reproducibility (Hansen 2020). Together with the high pregnancy losses still experienced, this indicates that there is need for more reliable embryo viability prediction methods. Besides the use of omics data for embryo competence prediction as discussed in this review, ML has also been applied to microscopic embryo images, though it is more widely explored in human research than in bovine research. ANNs have been trained on bovine embryo images labeled by embryologists to predict embryo quality grades (Rocha et al. 2017). However, quality grades do not directly translate to embryo viability, so incorporating pregnancy outcomes may increase predictive importance for embryo competence. ML performs better when the correlation between the label and the data is stronger. Omics data, being more directly related to the molecular biology of the embryo, may provide a more reliable prediction of embryo viability than images and might be better suited for predicting embryo survival through ML. However, further research should be performed assessing the predictive value of images and omics, alone or combined, in relation to embryo competence.
Future perspectives
The future of ML in bovine embryonic and fetal development holds promising advancements, considering the use of ML has only surged in recent years. The articles discussed in this review have applied ML on relatively small datasets, which can lead to suboptimal performance, overfitting and limited generalizability. Therefore, a key focus of future studies should be optimizing ML performance and validation of the already obtained results. Nevertheless, with the increasing availability of omics data in public databases, and the ability to integrate and reuse existing datasets, ML models will soon be able to be trained on larger and more diverse datasets, enhancing their robustness and accuracy. ML for biomarker discovery can be applied on integrated datasets and validated on external datasets, as we did for the study identifying biomarkers of embryo survival (Rabaglino et al. 2023a). These markers are powerful and reliable and can reach high predictive accuracies. Furthermore, the involvement of the selected biomarkers in relevant biological pathways highlighted their importance for embryo survival.
The increased understanding of embryo development through validated biomarkers and key pathways can be employed to improve embryo cultures and maternal treatments enhancing endometrial function. For example, culturing embryos in CM with amino acid concentrations similar to the uterine fluid improved embryo development and freezing viability (Li et al. 2006). Additionally, ML models can assist in embryo treatment experiments by providing a more reliable method for embryo competence estimation, allowing the treatment effect to be more accurately defined (Rabaglino and Hansen 2024).
Nowadays, the collection of biological samples from the uterus or the embryo is not done routinely, except for some breeding companies performing embryo biopsies for genomic selection. Nevertheless, using these reproductive management practices can be justified if the application of ML will objectively select competent embryos for ET and identify receptive cows, effectively increasing pregnancy success rates. However, for these models to be suitable for clinical implementation, they must be trained with large datasets (considering the biological content of the data) and demonstrate high performance across different settings. Field validation and development of applicability standards are crucial to harnessing the power of ML and ensuring its reliability and practical utility as a supportive tool in bovine reproduction.
Conclusion
The application of ML in bovine reproductive research holds significant potential for reducing embryo mortality and improving pregnancy success rates. By employing the power of ML to analyze complex and high-dimensional omics data, researchers can gain deeper insights into the molecular mechanisms essential for successful embryonic and fetal development. ML models have demonstrated their ability to predict pregnancy outcomes and to identify critical biomarkers. Future efforts should focus on developing non-invasive and practical applicable assessment methods to enhance the precision and efficacy of reproductive strategies. As the cattle industry continues to evolve, ML will play a crucial role in optimizing reproductive efficiency and genetic selection, ultimately contributing to its economic sustainability.
Data availability
Data sharing is not applicable as no new data were generated or analyzed during this study.
References
Alfattah MA, Correia CN, Browne JA, Mcgettigan PA, Pluta K, Carrington SD, Machugh DE, Irwin JA (2024) Transcriptomics analysis of the bovine endometrium during the perioestrus period. PLoS ONE 19, e0301005.
| Crossref | Google Scholar | PubMed |
Assis Neto AC, Pereira FTV, Santos TC, Ambrosio CE, Leiser R, Miglino MA (2010) Morpho-physical recording of bovine conceptus (Bos indicus) and placenta from days 20 to 70 of pregnancy. Reproduction in Domestic Animals 45, 760-772.
| Crossref | Google Scholar |
Banliat C, Mahé C, Lavigne R, Com E, Pineau C, Labas V, Guyonnet B, Mermillod P, Saint-Dizier M (2022) Dynamic changes in the proteome of early bovine embryos developed in vivo. Frontiers in Cell and Developmental Biology 10, 863700.
| Crossref | Google Scholar | PubMed |
Bauersachs S, Ulbrich SE, Zakhartchenko V, Minten M, Reichenbach M, Reichenbach HD, Blum H, Spencer TE, Wolf E (2009) The endometrium responds differently to cloned versus fertilized embryos. Proceedings of the National Academy of Sciences of the United States of America 106, 5681-5686.
| Crossref | Google Scholar | PubMed |
Berg DK, van Leeuwen J, Beaumont S, Berg M, Pfeffer PL (2010) Embryo loss in cattle between Days 7 and 16 of pregnancy. Theriogenology 73, 250-260.
| Crossref | Google Scholar |
Bó GA, Mapletoft RL (2018) Evaluation and classification of bovine embryos. Animal Reproduction 10, 344-348.
| Google Scholar |
Breiman L (2001) Random forests. Machine Learning 45, 5-32.
| Crossref | Google Scholar |
Bridi A, Perecin F, Silveira JC (2020) Extracellular vesicles mediated early embryo–maternal interactions. International Journal of Molecular Sciences 21, 1163.
| Crossref | Google Scholar |
Bucher K, Malama E, Siuda M, Janett F, Bollwein H (2019) Multicolor flow cytometric analysis of cryopreserved bovine sperm: a tool for the evaluation of bull fertility. Journal of Dairy Science 102, 11652-11669.
| Crossref | Google Scholar | PubMed |
Campanholi SP, Garcia Neto S, Pinheiro GM, Nogueira MFG, Rocha JC, Losano JDA, Siqueira AFP, Nichi M, Assumpção M, Basso AC, Monteiro FM, Gimenes LU (2023) Can in vitro embryo production be estimated from semen variables in Senepol breed by using artificial intelligence? Frontiers in Veterinary Science 10, 1254940.
| Crossref | Google Scholar | PubMed |
Camunas-Soler J, Gee EPS, Reddy M, Mi JD, Thao M, Brundage T, Siddiqui F, Hezelgrave NL, Shennan AH, Namsaraev E, Haverty C, Jain M, Elovitz MA, Rasmussen M, Tribe RM (2022) Predictive RNA profiles for early and very early spontaneous preterm birth. American Journal of Obstetrics and Gynecology 227, 72.e1-72.e16.
| Crossref | Google Scholar |
Chankeaw W, Guo YZ, Båge R, Svensson A, Andersson G, Humblot P (2018) Elevated non-esterified fatty acids impair survival and promote lipid accumulation and pro-inflammatory cytokine production in bovine endometrial epithelial cells. Reproduction, Fertility and Development 30, 1770-1784.
| Crossref | Google Scholar | PubMed |
Chen T, Guestrin C (2016) XGBoost: a scalable tree boosting system. In ‘Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining’. (Association for Computing Machinery) doi:10.1145/2939672.2939785
Chen X, Ishwaran H (2012) Random forests for genomic data analysis. Genomics 99, 323-329.
| Crossref | Google Scholar |
Cheredath A, Uppangala S, A CS, Jijo A, R VL, Kumar P, Joseph D, GA NG, Kalthur G, Adiga SK (2023) Combining machine learning with metabolomic and embryologic data improves embryo implantation prediction. Reproductive Sciences 30, 984-994.
| Crossref | Google Scholar | PubMed |
Costes V, Chaulot-Talmon A, Sellem E, Perrier JP, Aubert-Frambourg A, Jouneau L, Pontlevoy C, Hozé C, Fritz S, Boussaha M, Le Danvic C, Sanchez MP, Boichard D, Schibler L, Jammes H, Jaffrézic F, Kiefer H (2022) Predicting male fertility from the sperm methylome: application to 120 bulls with hundreds of artificial insemination records. Clinical Epigenetics 14, 54.
| Crossref | Google Scholar | PubMed |
Costes V, Sellem E, Marthey S, Hoze C, Bonnet A, Schibler L, Kiefer H, Jaffrezic F, Moreira N (2024) Multi-omics data integration for the identification of biomarkers for bull fertility. PLoS ONE 19, e0298623.
| Crossref | Google Scholar | PubMed |
Degrelle SA, Campion E, Cabau C, Piumi F, Reinaud P, Richard C, Renard JP, Hue I (2005) Molecular evidence for a critical period in mural trophoblast development in bovine blastocysts. Developmental Biology 288, 448-460.
| Crossref | Google Scholar | PubMed |
Desmet KLJ, Van Hoeck V, Gagne D, Fournier E, Thakur A, O’doherty AM, Walsh CP, Sirard MA, Bols PE, Leroy JL (2016) Exposure of bovine oocytes and embryos to elevated non-esterified fatty acid concentrations: integration of epigenetic and transcriptomic signatures in resultant blastocysts. BMC Genomics 17, 1004.
| Crossref | Google Scholar | PubMed |
Deutsch DR, Fröhlich T, Otte KA, Beck A, Habermann FA, Wolf E, Arnold GJ (2014) Stage-specific proteome signatures in early bovine embryo development. Journal of Proteome Research 13, 4363-4376.
| Crossref | Google Scholar | PubMed |
Diniz WJS, Banerjee P, Rodning SP, Dyce PW (2022) Machine learning-based co-expression network analysis unravels potential fertility-related genes in beef cows. Animals 12(19), 2715.
| Crossref | Google Scholar |
Ealy AD, Wooldridge LK, McCoski SR (2019) BOARD INVITED REVIEW: Post-transfer consequences of in vitro-produced embryos in cattle. Journal of Animal Science 97, 2555-2568.
| Crossref | Google Scholar |
Eraslan G, Avsec Ž, Gagneur J, Theis FJ (2019) Deep learning: new computational modelling techniques for genomics. Nature Reviews Genetics 20, 389-403.
| Crossref | Google Scholar | PubMed |
Farin PW, Slenning BD, Britt JH (1999) Estimates of pregnancy outcomes based on selection of bovine embryos produced in vivo or in vitro. Theriogenology 52, 659-670.
| Crossref | Google Scholar | PubMed |
Fernández-Delgado M, Cernadas E, Barro S, Amorim D (2014) Do we need hundreds of classifiers to solve real world classification problems? Journal of Machine Learning Research 15, 3133-3181.
| Google Scholar |
Frei RE, Schultz GA, Church RB (1989) Qualitative and quantitative changes in protein synthesis occur at the 8-16-cell stage of embryogenesis in the cow. Reproduction 86, 637-641.
| Crossref | Google Scholar |
Gimeno I, Salvetti P, Carrocera S, Gatien J, García-Manrique P, López-Hidalgo C, Valledor L, Gómez E (2023) Biomarker metabolite mating of viable frozen-thawed in vitro-produced bovine embryos with pregnancy-competent recipients leads to improved birth rates. Journal of Dairy Science 106, 6515-6538.
| Crossref | Google Scholar | PubMed |
Gómez E, Muñoz M, Gatien J, Carrocera S, Martín-González D, Salvetti P (2020a) Metabolomic identification of pregnancy-specific biomarkers in blood plasma of BOS TAURUS beef cattle after transfer of in vitro produced embryos. Journal of Proteomics 225, 103883.
| Crossref | Google Scholar | PubMed |
Gómez E, Salvetti P, Gatien J, Carrocera S, Martín-González D, Muñoz M (2020b) Blood plasma metabolomics predicts pregnancy in holstein cattle transferred with fresh and vitrified/warmed embryos produced in vitro. Journal of Proteome Research 19, 1169-1182.
| Crossref | Google Scholar | PubMed |
Gómez E, Canela N, Herrero P, Cereto A, Gimeno I, Carrocera S, Martin-Gonzalez D, Murillo A, Muñoz M (2021) Metabolites secreted by bovine embryos in vitro predict pregnancies that the recipient plasma metabolome cannot, and vice versa. Metabolites 11, 162.
| Crossref | Google Scholar |
Hansen PJ (2020) The incompletely fulfilled promise of embryo transfer in cattle–why aren’t pregnancy rates greater and what can we do about it? Journal of Animal Science 98, skaa288.
| Crossref | Google Scholar |
Hoorn QA, Rabaglino MB, Amaral TF, Maia TS, Yu F, Cole JB, Hansen PJ (2024) Machine learning to identify endometrial biomarkers predictive of pregnancy success following artificial insemination in dairy cows. Biology of Reproduction 111(1), 54-62.
| Crossref | Google Scholar |
Kaneko Y, Lindsay LA, Murphy CR (2008) Focal adhesions disassemble during early pregnancy in rat uterine epithelial cells. Reproduction, Fertility and Development 20, 892-899.
| Crossref | Google Scholar | PubMed |
Kaneko Y, Lecce L, Day ML, Murphy CR (2012) Focal adhesion kinase localizes to sites of cell-to-cell contact in vivo and increases apically in rat uterine luminal epithelium and the blastocyst at the time of implantation. Journal of Morphology 273, 639-650.
| Crossref | Google Scholar | PubMed |
Kastelic JP, Thundathil JC (2008) Breeding soundness evaluation and semen analysis for predicting bull fertility. Reproduction in Domestic Animals 43(Suppl 2), 368-373.
| Crossref | Google Scholar | PubMed |
Kingsford C, Salzberg SL (2008) What are decision trees? Nature Biotechnology 26, 1011-1013.
| Crossref | Google Scholar | PubMed |
Kopecny V, Flechon JE, Camous S, Fulka J, Jr (1989) Nucleologenesis and the onset of transcription in the eight-cell bovine embryo: fine-structural autoradiographic study. Molecular Reproduction and Development 1, 79-90.
| Google Scholar | PubMed |
Kursa MB, Rudnicki WR (2010) Feature selection with the boruta package. Journal of Statistical Software 36, 1-13.
| Crossref | Google Scholar |
Lê Cao K-A, Boitard S, Besse P (2011) Sparse PLS discriminant analysis: biologically relevant feature selection and graphical displays for multiclass problems. BMC Bioinformatics 12, 253.
| Crossref | Google Scholar |
Leclercq M, Vittrant B, Martin-Magniette ML, Scott Boyer MP, Perin O, Bergeron A, Fradet Y, Droit A (2019) Large-scale automatic feature selection for biomarker discovery in high-dimensional OMICs data. Frontiers in Genetics 10, 452.
| Crossref | Google Scholar | PubMed |
Leese HJ (2002) Quiet please, do not disturb: a hypothesis of embryo metabolism and viability. Bioessays 24, 845-849.
| Crossref | Google Scholar | PubMed |
Li R, Wen L, Wang S, Bou S (2006) Development, freezability and amino acid consumption of bovine embryos cultured in synthetic oviductal fluid (SOF) medium containing amino acids at oviductal or uterine-fluid concentrations. Theriogenology 66, 404-414.
| Crossref | Google Scholar |
Lonergan P, Fair T, Forde N, Rizos D (2016) Embryo development in dairy cattle. Theriogenology 86, 270-277.
| Crossref | Google Scholar | PubMed |
Mansouri-Attia N, Sandra O, Aubert J, Degrelle S, Everts RE, Giraud-Delville C, Heyman Y, Galio L, Hue I, Yang X, Tian XC, Lewin HA, Renard JP (2009) Endometrium as an early sensor of in vitro embryo manipulation technologies. Proceedings of the National Academy of Sciences of the United States of America 106, 5687-5692.
| Crossref | Google Scholar | PubMed |
Mathew DJ, Sánchez JM, Passaro C, Charpigny G, Behura SK, Spencer TE, Lonergan P (2019) Interferon tau-dependent and independent effects of the bovine conceptus on the endometrial transcriptome. Biology of Reproduction 100, 365-380.
| Crossref | Google Scholar | PubMed |
McMillan WH (1998) Statistical models predicting embryo survival to term in cattle after embryo transfer. Theriogenology 50, 1053-1070.
| Crossref | Google Scholar | PubMed |
McMillan WH, Donnison MJ (1999) Understanding maternal contributions to fertility in recipient cattle: development of herds with contrasting pregnancy rates. Animal Reproduction Science 57, 127-140.
| Crossref | Google Scholar | PubMed |
Memili E, First NL (2000) Zygotic and embryonic gene expression in cow: a review of timing and mechanisms of early gene expression as compared with other species. Zygote 8, 87-96.
| Crossref | Google Scholar | PubMed |
Moorey SE, Walker BN, Elmore MF, Elmore JB, Rodning SP, Biase FH (2020) Rewiring of gene expression in circulating white blood cells is associated with pregnancy outcome in heifers (Bos taurus). Scientific Reports 10, 16786.
| Crossref | Google Scholar | PubMed |
Moraes JGN, Behura SK, Geary TW, Hansen PJ, Neibergs HL, Spencer TE (2018) Uterine influences on conceptus development in fertility-classified animals. Proceedings of the National Academy of Sciences of the United States of America 115, E1749-E1758.
| Google Scholar | PubMed |
Muñoz M, Uyar A, Correia E, Díez C, Fernandez-Gonzalez A, Caamaño JN, Martínez-Bello D, Trigal B, Humblot P, Ponsart C, Guyader-Joly C, Carrocera S, Martin D, Marquant Le Guienne B, Seli E, Gomez E (2014a) Prediction of pregnancy viability in bovine in vitro-produced embryos and recipient plasma with Fourier transform infrared spectroscopy. Journal of Dairy Science 97, 5497-5507.
| Crossref | Google Scholar | PubMed |
Muñoz M, Uyar A, Correia E, Ponsart C, Guyader-Joly C, Martínez-Bello D, Marquant-Le Guienne B, Fernandez-Gonzalez A, Díez C, Caamaño JN, Trigal B, Humblot P, Carrocera S, Martin D, Seli E, Gomez E (2014b) Metabolomic prediction of pregnancy viability in superovulated cattle embryos and recipients with fourier transform infrared spectroscopy. BioMed Research International 2014, 608579.
| Crossref | Google Scholar | PubMed |
Ngo TTM, Moufarrej MN, Rasmussen MH, Camunas-Soler J, Pan W, Okamoto J, Neff NF, Liu K, Wong RJ, Downes K, Tibshirani R, Shaw GM, Skotte L, Stevenson DK, Biggio JR, Elovitz MA, Melbye M, Quake SR (2018) Noninvasive blood tests for fetal development predict gestational age and preterm delivery. Science 360, 1133-1136.
| Crossref | Google Scholar |
O’Callaghan E, Sanchez JM, Rabaglino MB, Mcdonald M, Liu H, Spencer TE, Fair S, Kenny DA, Lonergan P (2022) Influence of sire fertility status on conceptus-induced transcriptomic response of the bovine endometrium. Frontiers in Cell and Developmental Biology 10, 950443.
| Crossref | Google Scholar | PubMed |
Passaro C, Tutt D, Mathew DJ, Sanchez JM, Browne JA, Boe-Hansen GB, Fair T, Lonergan P (2018) Blastocyst-induced changes in the bovine endometrial transcriptome. Reproduction 156, 219-229.
| Crossref | Google Scholar | PubMed |
Ponsart C, Le Bourhis D, Knijn H, Fritz S, Guyader-Joly C, Otter T, Lacaze S, Charreaux F, Schibler L, Dupassieux D, Mullaart E (2013) Reproductive technologies and genomic selection in dairy cattle. Reproduction, Fertility and Development 26, 12-21.
| Crossref | Google Scholar | PubMed |
Rabaglino MB (2023) Review: overview of the transcriptomic landscape in bovine blastocysts and elongated conceptuses driving developmental competence. Animal 17(Suppl 1), 100733.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, Hansen PJ (2024) Development of a formula for scoring competence of bovine embryos to sustain pregnancy. Biochemistry and Biophysics Reports 39, 101772.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, Kadarmideen HN (2020) Machine learning approach to integrated endometrial transcriptomic datasets reveals biomarkers predicting uterine receptivity in cattle at seven days after estrous. Scientific Reports 10, 16981.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, O’Doherty A, Bojsen-Møller Secher J, Lonergan P, Hyttel P, Fair T, Kadarmideen HN (2021) Application of multi-omics data integration and machine learning approaches to identify epigenetic and transcriptomic differences between in vitro and in vivo produced bovine embryos. PLoS ONE 16, e0252096.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, Le Danvic C, Schibler L, Kupisiewicz K, Perrier JP, O’Meara CM, Kenny DA, Fair S, Lonergan P (2022) Identification of sperm proteins as biomarkers of field fertility in Holstein-Friesian bulls used for artificial insemination. Journal of Dairy Science 105, 10033-10046.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, Salilew-Wondim D, Zolini A, Tesfaye D, Hoelker M, Lonergan P, Hansen PJ (2023a) Machine-learning methods applied to integrated transcriptomic data from bovine blastocysts and elongating conceptuses to identify genes predictive of embryonic competence. The FASEB Journal 37, e22809.
| Crossref | Google Scholar | PubMed |
Rabaglino MB, Sánchez JM, Mcdonald M, O’callaghan E, Lonergan P (2023b) Maternal blood transcriptome as a sensor of fetal organ maturation at the end of organogenesis in cattle. Biology of Reproduction 109, 749-758.
| Crossref | Google Scholar | PubMed |
Rasmussen M, Reddy M, Nolan R, Camunas-Soler J, Khodursky A, Scheller NM, Cantonwine DE, Engelbrechtsen L, Mi JD, Dutta A, Brundage T, Siddiqui F, Thao M, Gee EPS, La J, Baruch-Gravett C, Santillan MK, Deb S, Ame SM, Ali SM, Adkins M, DePristo MA, Lee M, Namsaraev E, Gybel-Brask DJ, Skibsted L, Litch JA, Santillan DA, Sazawal S, Tribe RM, Roberts JM, Jain M, Høgdall E, Holzman C, Quake SR, Elovitz MA, McElrath TF (2022) RNA profiles reveal signatures of future health and disease in pregnancy. Nature 601, 422-427.
| Crossref | Google Scholar | PubMed |
Reese ST, Franco GA, Poole RK, Hood R, Fernadez Montero L, Oliveira Filho RV, Cooke RF, Pohler KG (2020) Pregnancy loss in beef cattle: a meta-analysis. Animal Reproduction Science 212, 106251.
| Crossref | Google Scholar | PubMed |
Riley JK, Moley KH (2006) Glucose utilization and the PI3-K pathway: mechanisms for cell survival in preimplantation embryos. Reproduction 131, 823-835.
| Crossref | Google Scholar | PubMed |
Rocha JC, Passalia FJ, Matos FD, Takahashi MB, Ciniciato DS, Maserati MP, Alves MF, De Almeida TG, Cardoso BL, Basso AC, Nogueira MFG (2017) A method based on artificial intelligence to fully automatize the evaluation of bovine blastocyst images. Scientific Reports 7, 7659.
| Crossref | Google Scholar | PubMed |
Rocha CC, da Silveira JC, Forde N, Binelli M, Pugliesi G (2021) Conceptus-modulated innate immune function during early pregnancy in ruminants: a review. Animal Reproduction 18, e20200048.
| Crossref | Google Scholar | PubMed |
Ruiz-Perez D, Guan H, Madhivanan P, Mathee K, Narasimhan G (2020) So you think you can PLS-DA? BMC Bioinformatics 21, 2.
| Crossref | Google Scholar |
Sánchez JM, Mathew DJ, Behura SK, Passaro C, Charpigny G, Butler ST, Spencer TE, Lonergan P (2019) Bovine endometrium responds differentially to age-matched short and long conceptuses. Biology of Reproduction 101, 26-39.
| Crossref | Google Scholar | PubMed |
Sartori R, Bastos MR, Wiltbank MC (2010) Factors affecting fertilisation and early embryo quality in single- and superovulated dairy cattle. Reproduction, Fertility and Development 22, 151-158.
| Crossref | Google Scholar | PubMed |
Schulz KN, Harrison MM (2019) Mechanisms regulating zygotic genome activation. Nature Reviews Genetics 20, 221-234.
| Crossref | Google Scholar | PubMed |
Sciorio R, Rinaudo P (2023) Culture conditions in the IVF laboratory: state of the ART and possible new directions. Journal of Assisted Reproduction and Genetics 40, 2591-2607.
| Crossref | Google Scholar | PubMed |
Seidl T (2009) Nearest neighbor classification. In ‘Encyclopedia of database systems’. (Eds L Liu, MT Özsu) pp. 1885–1890. (Springer: Boston, MA, USA) doi:10.1007/978-0-387-39940-9_561
Shawky JH, Davidson LA (2015) Tissue mechanics and adhesion during embryo development. Developmental Biology 401, 152-164.
| Crossref | Google Scholar | PubMed |
Sponchiado M, Gomes NS, Fontes PK, Martins T, Del Collado M, Pastore AA, Pugliesi G, Nogueira MFG, Binelli M (2017) Pre-hatching embryo-dependent and -independent programming of endometrial function in cattle. PLoS ONE 12, e0175954.
| Crossref | Google Scholar | PubMed |
Stephenson TA (2000) An introduction to Bayesian network theory and usage. IDIAP Research Report. (IDIAP). Available at https://infoscience.epfl.ch/handle/20.500.14299/227920
Taglauer ES, Trikhacheva AS, Slusser JG, Petroff MG (2008) Expression and function of PDCD1 at the human maternal-fetal interface. Biology of Reproduction 79, 562-569.
| Crossref | Google Scholar | PubMed |
Talukder AK, Rabaglino MB, Browne JA, Charpigny G, Lonergan P (2023) Dose- and time-dependent effects of interferon tau on bovine endometrial gene expression. Theriogenology 211, 1-10.
| Crossref | Google Scholar | PubMed |
Van Hoeck V, Sturmey RG, Bermejo-Alvarez P, Rizos D, Gutierrez-Adan A, Leese HJ, Bols PE, Leroy JL (2011) Elevated non-esterified fatty acid concentrations during bovine oocyte maturation compromise early embryo physiology. PLoS ONE 6, e23183.
| Crossref | Google Scholar | PubMed |
Viana JHM (2022) 2022 statistics of embryo production and transfer in domestic farm animals: the main trends for the world embryo industry still stand. Embryo Transfer Newsletter 41, 20-38.
| Google Scholar |
Wiltbank MC, Baez GM, Garcia-Guerra A, Toledo MZ, Monteiro PLJ, Melo LF, Ochoa JC, Santos JEP, Sartori R (2016) Pivotal periods for pregnancy loss during the first trimester of gestation in lactating dairy cows. Theriogenology 86, 239-253.
| Crossref | Google Scholar | PubMed |
Xiong Y, Lin L, Chen Y, Salerno S, Li Y, Zeng X, Li H (2022) Prediction of gestational diabetes mellitus in the first 19 weeks of pregnancy using machine learning techniques. Journal of Maternal-Fetal & Neonatal Medicine 35, 2457-2463.
| Crossref | Google Scholar | PubMed |
Zhang Z, He C, Zhang L, Zhu T, Lv D, Li G, Song Y, Wang J, Wu H, Ji P, Liu G (2019) Alpha-ketoglutarate affects murine embryo development through metabolic and epigenetic modulations. Reproduction 158, 123-133.
| Google Scholar | PubMed |
