Introduction

In bovine reproduction, the success rate of artificial insemination (AI) remains a pivotal concern, with a pregnancy rate of approximately 50%1,2. Addressing this challenge is crucial for the economic sustainability of the cattle industry and for advancing our understanding of reproductive physiology in these animals. In Brazil, around 80% of the herd is the Nellore breed, and efficacy improvements in artificial insemination of Nellore cattle have been the focus of many studies2,3.

In reproductive management, selecting high-fertility females is crucial to enhance reproductive indices. However, the assessment of fertility profile can be a complex trait due to its multifactorial nature, such as hormonal differences, age-related factors, body condition score, weight, and even environmental factors4. The relationship between fertility and age is inherently linked to animals becoming less fertile as they age. Additionally, recognizing when a female reaches puberty—characterized by high fertility—is critical as it allows the animal a more extended reproductive cycle5,6. Considering the body condition score (BCS), the selection is based on the principle that females with appropriate BCS (3 on a scale of 1 to 5 and 5 on a scale of 1 to 9) have more significant energy reserves. This can be converted into increased hormonal production, leading to higher ovarian activity, greater conception rates, and even shorter calving intervals, reducing the time between births7,8. It also contributes to starting the best females into reproductive protocols and establishing nutritional management for adequate energy balance9.

An essential parameter to understand the reproductive health of female cattle is the evaluation of the population of antral follicles in the ovaries. Females with a high population of antral follicles reach puberty earlier, exhibit higher conception rates with lower embryonic loss, and have a higher concentration of progesterone from the corpus luteum. Furthermore, there is a positive correlation between the antral follicle population and embryo production10. Although these parameters are essential to define a fertility profile, the efficiency of reproductive management also relies on the early diagnosis of pregnancy and fertility, impacting the handling of animals undergoing a new reproductive cycle with shorter calving intervals. Currently, this diagnosis is conducted through transrectal palpation at 45 days after insemination, transrectal ultrasonography (conventional or Doppler mode) at 30 days after insemination, and the measurement of pregnancy-associated glycoproteins (PAGs)11,12.

A promising strategy to enhance the success of artificial insemination is to implement a hormonal protocol that includes progesterone and estradiol. This practice aims to induce heat in cows and synchronize the moment of insemination (Fixed-Time Artificial insemination—FTAI). This strategy has shown promise in enhancing the precision of artificial insemination protocols13. However, the comprehension of the intricate physiological changes induced by intravaginal progesterone remains under investigation, particularly in the blood serum composition of the treated animals.

The compositional alterations within Nellore cattle’s blood serum due to FTAI protocol can be used to unravel the biochemical changes associated with hormonal modulation, providing valuable insights into the factors influencing successful artificial insemination in this specific breed. Studies about the reproductive performance and hormonal changes in Nellore cattle undergoing fixed-time artificial insemination (FTAI) protocols demonstrate that these animals subjected to FTAI protocols exhibit changes in progesterone (P4) levels during the estrous cycle synchronization process. These hormonal manipulations likely lead to alterations in the blood serum composition of Nellore cows, including changes in steroid hormone levels14,15,16.

Successful early pregnancy diagnosis in sheep has been accomplished through the characterization of blood plasma using near-infrared spectroscopy (NIRS). The analysis of blood plasma spectra within the 400 to 2498 nm range, coupled with Partial Least Square Discriminant Analysis (PLS-DA), yielded results demonstrating comparable or even superior performance (with a sensitivity of 98% and specificity of 100%) compared to traditional pregnancy-associated glycoproteins (PAG) and progesterone (P4) ELISA tests conducted 18 days after artificial insemination17. The Vis–NIR spectroscopy in the 490 to 960 nm range was able to predict early pregnancy of rabbits by a non-invasive approach with 86% accuracy—an extra-corporeal measurement showed to be able to differentiate the gestational sac from other abdominal tissues with PLS-DA aid18,19. Other optical techniques, such as Laser Induced Breakdown Spectroscopy (LIBS), Laser Induced Fluorescence (LIF), and Raman spectroscopy, were used to evaluate buffalo and Karan Fries semen to improve the success rate of artificial insemination20,21. The NIR spectroscopy of urine in the 1100 to 2432 nm, coupled with PLS, could also determine the giant panda’s timing of ovulation22. Finally, the potential use of mid-infrared spectroscopy to identify more fertile dairy cows for insemination was demonstrated by milk analysis23.

Intravaginal progesterone treatment can potentially elicit changes in cows’ molecular spectral signature of blood serum. Cows likely providing more favorable conditions for conception will exhibit distinct responses to the treatment compared to those succeeding in artificial insemination protocols. Then, Fourier-transform infrared (FTIR) spectroscopy can be used to obtain vibrational molecular information from blood serum; these data can be interpreted with multivariate analysis (machine learning algorithms) aid to build a prediction model able to identify specific spectral signatures within the blood serum that correlate with a higher likelihood of successful pregnancy. By applying this innovative approach, we aim to distinguish, at an early stage, those Nellore cattle that are more likely to conceive from those that may not respond optimally to the FTAI protocol. Through this investigation, we can anticipate and enhance the artificial insemination success in this economically significant breed.

Materials and methods

This study was approved by the Animal Use Ethics Committee (CEUA) of the Federal University of Mato Grosso do Sul, UFMS, Campo Grande—MS, under protocol n 1273/2023.

Animals and fixed-time artificial insemination (FTAI) protocol description

Sixty Nellore (Bos indicus) female cattle, heifers and primiparous, aged 14 to 36 months, weighing on average 320 kg (± 20), with BCS of 3.25, were maintained in paddocks with uniform environmental conditions. The animals received mineral supplementation, and water was provided ad libitum.

All animals underwent the following ovulation synchronization protocol: on the initiation day of the protocol (10 days before the insemination: D-10), they were administered an intravaginal progesterone-releasing device (P4; Sincrogest®, Ourofino, SP, Brazil) along with 2 mg i.m. of estradiol benzoate (BE; Sincrodiol®, Ourofino, SP, Brazil). On D-2 (2 days before insemination), the P4 device was removed, and 500 µg i.m. of sodium cloprostenol (PGF; Sincrocio®, Ourofino, SP, Brazil), 1 mg i.m. of estradiol cypionate (CE; E.C.P.®, Pfizer, SP, Brazil), and 300 IU i.m. of equine chorionic gonadotropin (eCG; Novormon®, Intervet, SP, Brazil) were administered. Concurrent with removing the P4 implant, all females were visibly marked at the base of the tail using a marking stick to facilitate estrus expression identification. Their markings were subsequently evaluated for presence and intensity at the time of AI, indicating mating receipt. Blood sample collection and Artificial insemination (AI) were conducted after 48 h (insemination day: D0). All animals were inseminated with frozen semen from the same bull, whose macroscopic and microscopic characteristics were in the parameters established by the manual for the andrological examination and evaluation of animal semen24. The pregnancy diagnosis (PD) was performed 30 days after the AI (30 days after insemination: D30) by transrectal ultrasound. From this point, obtaining two groups for analysis was possible: (P) pregnant and (NP) non-pregnant groups with thirty samples each. The timeline of activities is summarized in Fig. 1.

Figure 1
figure 1

Timeline of procedures. Fixed-time artificial insemination (FTAI) protocol description. D-10, D-2 means 10 and 2 days before the artificial insemination, respectively. D0 is the day of artificial insemination and D30 means thirty days after artificial insemination.

Full size image

Sample collection and FTIR analysis

Blood samples were collected from all animals in the D0 before the artificial insemination procedure. Blood samples were collected (at least 4 mL) from the jugular vein in tubes with a clot activator (Vacutainer®, Becton–Dickinson & Company), and the samples were immediately stored in an isothermal box with ice until centrifugation at 600×g for 10 min. Subsequently, the serum was removed, aliquoted into microtubes containing 500 µL (in quadruplicates) and stored in an ultra-freezer at − 80 °C until the day of processing.

The blood serum was thawed at room temperature before the FTIR measurements. A drop of 20 µL of serum is placed directly onto the crystal of the Attenuated Total Reflectance (ATR) accessory of an Agilent spectrometer, model Cary 630. The measurements were taken in 2000 to 900 cm−1 range, with 12 scans and 4 cm−1 resolution, using deionized water as a background25.

The averaged FTIR spectra underwent preprocessing before analysis. Initially, the spectra intensity was normalized with the amide I (1636 cm−1) intensity as a reference. Subsequently, a Savitzky–Golay filter with a polynomial function degree of 2 and 11 data points was applied to reduce noise signals. This normalization process ensures comparability and meaningful comparisons between different spectra, especially when analyzing samples with different concentrations or physical states. The primary objective is to mitigate the impact of factors that could obscure critical spectral features, thereby facilitating accurate data interpretation26.

Typically, the normalization process entails dividing each data point in the spectrum by a normalization factor to minimize random fluctuations in experimental data. Subsequently, the data were subjected to principal component analysis (PCA), an unsupervised technique for dimensionality reduction27. This Python-based analysis utilized the Scikit-learn package (version 1.1.2) to segregate the data based on variance.

Outliers were removed from the data set by T2–hoteling28. Then, sample classification was carried out using machine learning (ML) algorithms, utilizing predictive models generated from the output data of PCA applied to 70% of the sample set. The Support Vector Machine (SVM) method was employed as the prediction algorithm, which organizes sample classes by optimizing a hyperplane capable of exhibiting linear or nonlinear characteristics29.

The SVM framework explored diverse functions to construct hyperplanes, including Linear, Quadratic, and Cubic. The hyperparameter C (cost parameter) was adjusted to balance fitting the training data well and avoiding overfitting. Fine-tuning of this hyperparameter was crucial, as their optimal values depend on the dataset’s characteristics30.

The investigation extended the 1800–900 cm−1 range to capture relevant sample data patterns and improve prediction model accuracy. The model’s performance was evaluated using classification accuracy in a Leave-One-Out Cross-Validation (LOOCV) test, wherein a sample is selected for testing. In contrast, the remaining samples are used for model training31. To enhance overall accuracy, an examination of the prediction model’s performance across various numbers of principal components (PCs) was conducted. Optimal accuracies were observed when additional high-order PCs with lower data variance were incorporated. The number of PCs leading to the first peak in overall accuracy during both LOOCV and Validation tests was selected, mitigating overfitting32,33.

After determining the optimal number of PCs and identifying the most influential ones for sample classification, a LOOCV test was executed using SVM and different functions. The most effective ML algorithm and PCs were pinpointed to construct a predictive model whose generalization capacity was evaluated through an external validation test using 30% of the sample set34.

Ethical approval

The present study was approved by the Animal Use Ethics Committee (CEUA) of the Federal University of Mato Grosso do Sul, UFMS, Campo Grande—MS, under protocol n 1273/2023. All procedures were performed in accordance with the ethical standards of the Brazilian regulation agencies.

Results and discussion

Figure 2 shows the FTIR-Normalized spectra for female Nellore blood serum submitted to insemination protocol ten days after the progesterone-releasing device implant. The superimposed FTIR-Normalized spectra for all samples from both groups (P and N), Fig. 2, exhibit remarkable similarity. The average normalized spectra for each group with respective standard deviations can be visualized in the Supplementary Fig. SM1. The main vibrational bands were assigned to proteins, carbohydrates, fatty acids, and lipids35.

Figure 2
figure 2

Superimposed FTIR-Normalized spectra for female Nellore blood serum submitted to insemination protocol 10 days after the progesterone-releasing device implant. Cows with positive (D0_P) (blue trace) and negative (D0_N) (red trace) pregnancy results 30 days after artificial insemination.

Full size image

The FTIR spectra obtained from different samples before the artificial insemination exhibit the same vibrational modes, which suggests that the alterations caused by the progesterone-releasing device implant in the blood serum are minor compositional changes below the equipment limit of quantification. There are two prominent vibrational bands; the first, centered around 1636 cm−1 (C=C and C=O stretching) assigned to Amide I from proteins and lipids molecules, is remarkable, followed by 1544 cm−1 (N–O stretching and N–H bending) assigned to Amide II. The second results from several superimposed vibrational modes, which can be assigned to carbohydrates, lipids, phospholipids, and protein molecules25,35. Progesterone and other hormones associated with fertility are intricately linked to the molecular groups of proteins and lipids.

Besides the two main prominent bands assigned to Amide vibrational modes, it is possible to identify minor bands at 1452 cm−1, assigned to Amide II from proteins and lipids (CH2 deformation; C–H scissoring bending; COO– symmetric stretching; C–H bending); 1397 cm−1, assigned to amino acids and proteins (O–H bending; COO– symmetric stretching); 1311 cm−1, assigned to Amide III and collagen from transferrin, glycoprotein and α1-acid; 1236 cm−1, assigned to lipid phosphate and Amide III (P–O stretching); 1163 cm−1, assigned to carbohydrates and proteins (C–O stretching; C=O stretching; C–O–H; C–O–C); 1126 cm−1, assigned to carbohydrates and proteins (C–O stretching); and 1075 cm−1, assigned to sugar (PO2)25,34. However, the vibrational modes from each group may differ due to subtle alterations in blood serum composition due to organism response induced by the artificial insemination protocol. These minor compositional changes may occur by adding pregnancy-associated glycoproteins and progesterone content in the blood17.

These minor blood serum FTIR spectra alterations can be highlighted with principal component analysis (PCA) to verify the group separation tendency. Figure 3 shows the score plot results and loadings for PC1 and PC2, which are responsible for 69.9% of the data variance (Supplementary Fig. SM2). The score plot exhibits the sample set used for machine learning prediction model training (full circles) and the sample set used for validation tests (hollow circles).

Figure 3
figure 3

Principal component analysis from FTIR-normalized spectra data for female Nellore blood serum submitted to insemination protocol ten days after the progesterone-releasing device implant. Cows with positive (D0_P) (blue trace) and negative (D0_N) (red trace) pregnancy results at 30 days after artificial insemination.

Full size image

The principal component analysis effectively distinguished between the positive (P) and negative (N) pregnancy groups, as illustrated in Fig. 3, using FTIR-normalized spectra data from female Nellore blood serum subjected to the insemination protocol 10 days after the progesterone-releasing device implant. Besides the great dispersion of the scores, it was possible to identify two distinct clusters using only the PC1 × PC2 score plot36. Additionally, a subtle clear separation boundary between each group is observed—a compelling indication of the successful classification of samples for machine learning algorithms.

The loading plot describes the correlation between the principal component information and the original data information (molecular vibrational modes from FTIR spectra), as can be observed in the Amide I and Amide II bands centered in the 1500 to 1650 cm−1 range are those which more contribute for data variance in PC1 and PC2 components. The minor band contribution was mitigated by the noise signal in the loading, which can be assigned to noise inherent to the measurement process (equipment and sample type limitations) since outliers were previously removed and diverse alternatives of data preprocessing were explored before the result showed here. The presence of loading noise can introduce instability in PCA results. Still, it probably will not significantly impact the results since we observe a great separation between the groups with only PC1 and PC2 components.

The machine learning (ML) algorithm utilized score plots as input data. Before conducting the ML tests, the scores were normalized to enhance classification performance, mitigate bias, and foster greater stability in the results, Fig. 4. These results for Normalized PCs demonstrate the significant contribution of PC2 for group separation with only 29.26% of data variance and the correlation with the original data, mainly 1636, 1544, 1452, and 1397 cm−1 bands—these bands assigned to proteins, lipids, and carbohydrates—for group clustering and separation.

Figure 4
figure 4

Normalized principal components used to build the prediction model by machine learning algorithms, and respective loading plot (green line) showing the correlation with the original data set (FTIR-Normalized spectra data for female Nellore blood serum submitted to insemination protocol 10 days after the progesterone-releasing device implant—gray line). Cows with positive (D0_P) (blue trace) and negative (D0_N) (red trace) pregnancy results at 30 days after artificial insemination.

Full size image

Then, we explore the performance of a Support Vector Machine (SVM) with different functions (Linear, Quadratic, and Cubic), and the hyperparameter C was adjusted (1, 10, and 100) to provide the best accuracy. The proper number of PCs for each function was chosen according to the overall accuracy achieved in the LOOCV and Validation test (Supplementary Fig. SM3), avoiding over and underfitting. The number of PCs is chosen when the LOOCV and Validation Test accuracy ratio reaches its first maximum. Figure 5 summarizes the overall accuracy achieved for each test; the highest overall accuracy was 100% for SVM with C = 10 for any function. The lowest number of PCs used to reach 100% accuracy was 4 PCs, which is responsible for 83.7% of data variance.

Figure 5
figure 5

Overall accuracy obtained in the LOOCV tests for support vector machine (SVM) algorithms. Different numbers of PCs were explored according to the function (Linear, Quadratic, and Cubic) and the hyperparameter C (1, 10, and 100) in the 1800 to 900 cm−1 range. The highest overall accuracy was 100% for SVM, with C = 10 for any function.

Full size image

Figure 6 shows the merit figure for sample classification of female Nellore submitted to the insemination protocol ten days after the progesterone-releasing device implant. The Linear SVM with C = 10 using 4 PCs, responsible for 83.7% data variance, achieved a 93.8% accuracy in the external validation test (blind test), exhibiting high generalization capacity of the prediction model. Here, only one cow with a negative pregnancy result after artificial insemination was misclassified.

Figure 6
figure 6

Confusion matrix for the performance of Linear SVM algorithm using 4-PCs and C = 10. Classification results were obtained in the LOOCV test (right side) with a 70% data set and the External validation (left side) test with a 30% data set. PCA data from FTIR blood serum spectra in the 1800 to 900 cm−1 range.

Full size image

To evaluate the robustness of FTIR spectroscopy with machine learning for real applications in selecting high-fertility females, we also tested the prediction probability of each sample in the LOOCV and Validation Test. Since we have both groups near each other in the score plot, the sample classification near the separation boundary can be hindered, and the results can be made by chance. Most of the samples from the N group showed a prediction probability above 75% in the LOOCV test (Supplementary Fig. SM4), and only one sample was around 60%. Similarly, the P group showed a prediction probability above 75% and only a sample above 65%. The P group also shows a broader prediction probability distribution, while the N group centers around 100%.

Figure 7 shows the prediction probability distribution for N and P groups obtained in the validation test. The negative group exhibits a bimodal prediction probability distribution; most samples are classified with probabilities above 90%, while only two are classified with probability around 70%. As shown in Fig. 5b, one sample from the N group was classified as a positive—sample identified by the X symbol in Fig. 7—despite having a 70% prediction probability, this sample was inaccurately categorized. To enhance the reliability of the method, it may be beneficial to introduce a cutoff parameter in the machine learning classification. With this parameter, samples with prediction probabilities below the specified cutoff would be deemed outliers and not classified. This adjustment could contribute to a more robust and accurate classification process.

Figure 7
figure 7

Prediction probability of each sample at validation test into the Linear SVM algorithm using 4-PCs and C = 10. Prediction probability for positive (Pos) group (right side—blue bars) and negative (Neg) group (left side—red bars). Sample distribution is shown in the middle (P = blue circle; N red circle).

Full size image

The positive group exhibited a wide prediction probability distribution from 60 to 90%, and all samples were correctly classified. If a cutoff parameter is inserted as suggested, three samples would be removed from the set, corresponding to 37% of the cows analyzed. This is worse than the original misclassification observed in Fig. 6. Here, we may analyze how this prediction model behaves with more samples in the validation test. Alternatively, many other improvements can be inserted into the protocol to enhance more parameters such as accuracy, sensitivity, and specificity.

Finally, this research contributes to the evolving field of reproductive management practices by offering a non-invasive, timely, and efficient method for high-fertile female cows’ selection for artificial insemination protocol. The identified spectral signatures associated with successful classification provide a foundation for further investigations around changes in the blood serum due to the progesterone-releasing device implant. As we continue to refine and expand this approach, it holds significant potential for enhancing artificial insemination’s overall success and efficiency in this economically significant breed, ultimately benefitting the cattle industry.

Conclusions

In conclusion, our study introduces a novel and promising approach for early high-fertility female Nellore cattle selection at 10 days after progesterone-releasing device implant before artificial insemination. The FTIR analysis of blood serum revealed consistent vibrational modes, particularly in the Amide I and II bands, and multivariate analysis offered valuable insights into molecular changes caused by FTAI protocol. Principal component analysis (PCA) effectively distinguished between pregnant and non-pregnant cows before artificial insemination, highlighting the potential of FTIR data and subsequent application of ML algorithms, specifically Support Vector Machine (SVM), for high accuracy in predicting pregnancy status, achieving 100% accuracy using Linear SVM with C = 10 and 4 PCs.