General design
This study reused phenotypic data obtained in pigs from the three originally separated experiments that were previously published [19–21], to avoid the needs of new sampling in living animals while obtaining a high number of animals allowing robust predictions. The application of ML procedures on the merged dataset (n = 148 pigs) avoided the overfitting often observed when simple classification or regression procedures are used for a limited number of animals and high number of dependent variables, and the leave-one-out method was an additional way to resampling the datasets. Thus, this study fits with the 3R (Replacement, Reduction and Refinement) principles.
Pigs And Blood Samples
The three originally independent datasets referred to purebred French Large White pigs produced in the course of a divergent selection experiment for RFI. The selection program was described in full details elsewhere [41], including the equation to calculate RFI from a regression between observed feed intake and that expected based on requirements for maintenance (based on the metabolic BW) and performance (average daily gain, backfat thickness). From birth to weaning, all pigs were reared in the selection farm of INRAE (UE Genesi, Le Magneraud & Rouillé, France; https://doi.org/10.15454/1.5572415481185847E12). All pigs were weaned at 28 days (d), and were first fed ad libitum with standard starter and weaner diets. During subsequent test periods in dedicated buildings, pigs have underwent different feeding conditions depending on the experiments as described below. As indicated in the referenced publications [19–21], the three experiments were conducted in accordance with the French legislation on animal experimentation, and the protocols were approved by regional ethical committees evaluating the research question, design, plan analysis, animal care and monitoring, and ways to minimize pain and consider limit points (especially regarding jugular blood sampling). At the end of each experiment, pigs were slaughtered using approved procedures, including electronarcosis followed by jugular exsanguination.
The first dataset [19] included 21 castrated males from the 7th generation of selection (n = 10 low RFI pigs and n = 11 high RFI pigs) housed at thermo-neutrality (24 °C) and reared at the INRAE experimental pig facility at Saint-Gilles, France (UE3P, https://doi.org/10.15454/1.5573932732039927E12). At 80 d of age, pigs were transferred in individual cages, and were fed a standard diet that met nutritional requirements for growth. At 87 d of age (59.2 kg BW on average), blood was collected from the jugular vein and prepared for RNA extraction. The feed conversion ratio (FCR) was calculated from individually measured daily feed intake and average daily gain for the 14 d of the trial (i.e., from 87 d to 100 d of age).
The second dataset [20] included 48 castrated males from the 8th generation of selection (n = 24 low RFI pigs and n = 24 high RFI pigs). Pigs were reared at the INRAE experimental pig facility at Saint-Gilles, France (UE3P, https://doi.org/10.15454/1.5573932732039927E12). At 74 d of age, pigs were transferred in individual cages and after 2 d of transition, the first half was fed a standard diet and the second half was fed a high-fiber high-fat diet during the growing and finishing phases. At 132 d of age (average BW of 75.6 kg), blood was sampled from the jugular vein and prepared for RNA extraction. The FCR was then calculated from 76 d to 132 d of age.
The third dataset [21] included 79 castrated males and females from the 9th generation of selection (n = 37 low RFI pigs and n = 42 high RFI pigs). Pigs were reared at the experimental INRAE pig facility at Le Magneraud, France (UE Genesi; https://doi.org/10.15454/1.5572415481185847E12). Blood was sampled at 40 d of age from the jugular vein. At 70 d of age, pigs were transferred in group-housing facilities equipped with single-place electronic feeders. The first half of the pigs was fed standard diets, whereas the second half was fed a high-fiber diet during the growing-finishing phases [21]. The FCR was then calculated from 90 d to 161 d of age.
In the three datasets, the reference to low or high RFI line was indicated for each pig, and FCR value was individually attributed. Other factors (sex, season, generation, diet) were not taken into account.
Microarrays Data
Microarray data considered in the current study were obtained from the referenced publications in the first [19] and second [3] experiments, and were newly acquired from RNA extracted from the stored blood samples in the third experiment. All experiments followed the same procedures for RNA extraction and expression data generation. The porcine commercial Agilent-026440 microarray (V2, 44K, GPL15007, Agilent Technologies, Massy, France) has been used in the first experiment. The custom porcine microarray (8 × 60K, GPL16524 Agilent Technologies) that contained the same probes as the Agilent-026444 and an additional set of probes enriched with immune system, muscle and adipose tissue genes, has been used in the second and third experiments. In the three transcriptomic datasets, raw spot intensities have been submitted to quality filtration based on four criteria: background intensity value, diameter, saturation and uniformity of the spot, and intensities of filtered spots were log2 transformed and median-centered to correct for microarray effect.
For the current study, the three microarray datasets were then merged into a single new dataset. There was no exclusion of any animals in this merged dataset. To obtain consolidated expression values across the three originally separated datasets, the molecular data have been normalized by mean centering, i.e. subtracting the mean value across all probes from all raw values for each pig sample in the merged dataset. The merged dataset also included meta-data such as the experiment of origin (1, 2, and 3), RFI group (n = 71 pigs of low RFI line, n = 77 pigs of high RFI line) and FCR value (n = 148 pigs). All data were deposited in a publicly available repository at https://doi.org/10.15454/J4XOPD.
Supervised methods to identify important variables for prediction of FE traits
The merged dataset was used to search the most important molecular predictors for RFI group and FCR value, by using ML methods. The experimental unit was the pig. Among the panel of ML methods for dimensionality reduction, classification and regression used in livestock breeding [14], the RF and GTB procedures were chosen in the current study and were compared for performance in classification (RFI group) and regression (FCR value) procedures. These two ML methods use decision trees, but RF uses a large number of trees combined by averaging or "majority rules" at the end of the process [42], whereas GTB starts the combining process of decision trees at the beginning [27, 43, 44]. Other differences include how trees are built: RF builds each tree independently, while GTB builds one tree at a time but in an additive model proceeding in a forward stage-wise sequential error–correcting process to combine results along the way and converge to an accurate model [27]. Sequential steps for learning, validation, and finally, selection of the best models were performed according to standards described by Fernandez-Lozano and colleagues [45]. Models were generated from RF and GTB algorithms with Salford Predictive Modeler 8.0 (SPM 8.0®).
The RF models were generated with about 1,500 trees for classification of RFI and regression for FCR. For that, a randomly selected bootstrap sample set was created by using 50% of the original dataset for learning (n = 74 pigs). Consequently, each bootstrap sample called “out-of-bag” data (OOB) excluded 50% of the data that were further used for validation (n = 74 pigs), and the leave-one-out method assessed the performance by resampling the training set. The test dataset allowed a cross-validation ensuring that the training of the model was not biased. To split branches of a tree, a random sample of m variables was chosen from the full set of p variables. Partition of probes between learning and validation datasets was shown in Supplementary Fig. 1. We checked that the three experiments of origin were included in both training and validation datasets.
The GTB prediction models were also generated using 1,500 small decision trees for classification or regression, and using a randomly selected bootstrap sample set for learning (n = 74 pigs) and the remaining data (n = 74 pigs) for validation. As recommended, each tree typically contained about six terminal nodes. The model was similar to Fourier or Taylor series, which is a sum of factors that becomes progressively more accurate as the expansion continues. After each step of boosting, the algorithm scaled the newly added weights, which balanced the influence of each tree. The accuracy of the algorithm was improved by introducing randomization through training the base learner on different randomly selected samples at each iteration.
In both procedures, significant variables were selected using the Gini index to evaluate the discriminant ability of the potential selected feature, defined as:
Gi = 1 − ∑𝑗 𝑝2 (𝑗 | 𝑡)
Where p2 (𝑗 | 𝑡) is the estimated class probability for feature 𝑡 or node 𝑡 in a decision tree and 𝑗 is an output data or class. Only the variables that improved Gini index and minimized the OOB error rate were retained as very important variables in prediction (VIP).
Multiple runs for each ML methods were performed (ten times) to take into account variations in the observations used for the training step (using permutations and leave-one-out procedures) and the stability of the techniques. The iteration steps were also applied to reduce the number of VIP in the selected models. At each run, the accuracy of classification models was estimated with the proportion (%) of good classification and the optimal models were selected according to ROC curve. In regression, RMSE was calculated as the square root of the difference between the realized and the predicted observation within the OOB data after permuting each predictor variable in the training dataset divided by the number of trees for regression procedure. The adjusted coefficient of determination (R²) was also computed. The predicted (X) values for FCR obtained by the best GTB model and the observed (Y) values measured on the pigs were compared (X-Y) using the GLM procedure. The model was considered unbiased when the intercept obtained by the GLM model was not different from 0 and the slope was not significantly different from 1. The quality of the relationships was evaluated on the basis of RMSE of prediction (RMSEP) obtained by a leave-one-out cross-validation from the value of the predicted residual sum of squares.
Pathway Enrichment Analysis
Gene-annotation enrichment analyses among the VIP identified for binary classification of pigs on RFI and for prediction of FCR were performed on encoded genes by using DAVID bioinformatics tool on default settings [46].