Mouse oocyte phenotypes differ in blastocyst response to resveratrol and vitamin B12

Abstract

Abstract

This experimental laboratory study examined whether associations between culture condition and blastocyst formation differed across radiomics-defined oocyte clusters and age groups. Mouse oocytes (n = 2,061) were assigned to control or five supplementation groups: resveratrol (0.1, 0.5, 1.0 μM) and vitamin B12 (50, 200 pM). The primary outcome was blastocyst formation across morphological clusters and age subgroups. Three clusters with distinct baseline blastocyst formation rates were identified: Cluster 1, 28.2% (44/156); Cluster 2, 42.4% (156/368); and Cluster 3, 34.9% (105/301). Cluster 2 comprised the largest and most spherical oocytes with heterogeneous cytoplasmic texture. Cluster 3 represented the smallest, brightest phenotype with the most uniform intensity profile, whereas Cluster 1 showed intermediate size, irregular contours, and high internal contrast. Associations between culture condition and blastocyst formation differed across radiomics-defined clusters. In Cluster 1, blastocyst formation increased from 28.2 to 45.1% with resveratrol 1.0 μM and to 48.7% with vitamin B12 200 pM. In Cluster 3, blastocyst formation increased from 34.9 to 60.4% with vitamin B12 50 pM and to 50.0% with resveratrol 0.1 μM. Cluster 2 showed a non-significant increase from 42.4 to 57.9% with resveratrol 0.5 μM. Older oocytes exhibited lower developmental competence than younger oocytes. Selecting the best performing medium for each cluster–age combination increased overall blastocyst formation from 36.97% (305/825) to 56.00% (126/225). These findings indicate associations among radiomics-defined clusters, culture conditions, and blastocyst formation. Prospective cluster-stratified studies are required to determine whether cluster-guided culture-condition selection improves developmental outcomes.

Lay summary

Not all eggs have the same ability to develop into embryos, and they may respond differently to the conditions in which they are cultured. In this study, we analysed images of more than 2,000 mouse eggs to determine whether features visible in the images could identify groups of eggs with different developmental potential. We first used eggs grown under standard culture conditions to define three groups based on their image characteristics. We then examined how eggs in these groups developed under standard conditions or with different amounts of resveratrol, an antioxidant, or vitamin B12 added to the special fluid used to support embryo development. The three groups differed in their rates of development to the blastocyst stage, an early stage of embryo development, and the relationship between culture condition and blastocyst development also varied between groups and between younger and older mice. These findings suggest that egg image characteristics may help us understand why eggs respond differently to their culture environment. However, further prospective experiments are needed before these findings can be used to guide culture conditions.

Introduction

In vitro fertilisation (IVF) is an advanced assisted reproductive technology that enables fertilisation and early embryonic development under controlled laboratory conditions. Nevertheless, IVF outcomes remain critically dependent on oocyte quality and adequacy of the culture environment. Over the past several decades, advancements in culture medium formulations, oxygen tension, and incubation systems have improved IVF outcomes. However, pregnancy and live birth rates remain inconsistent, particularly among patients of advanced maternal age or those with suboptimal oocyte quality.

Despite these progressive advances, most IVF laboratories still adhere to a ‘one-size-fits-all’ paradigm, wherein identical culture media and supplementation protocols are uniformly applied to all oocytes. This approach neglects substantial oocyte heterogeneity arising from maternal age, genetic background, or subtle morphological or metabolic variations (Cimadomo et al. 2018, Zhang et al. 2025b), which influence developmental potential and responsiveness to the culture environment (Yanez et al. 2016).

Recent advances in imaging analytics and computational biology offer opportunities to break this paradigm. Radiomics, originally developed in oncology, extracts high-dimensional quantitative features, such as shape, texture, and intensity, from standard microscopic images (Lambin et al. 2012, Gillies et al. 2016, Bera et al. 2021, Alabi et al. 2024, Lomer et al. 2025) and enables objective oocyte phenotyping by capturing subtle morphological and cytoplasmic variations that may not be apparent to the human eye.

Of the numerous potential supplements for IVF media, resveratrol and vitamin B12 were selected for this study because of their distinct and biologically relevant mechanisms. Resveratrol, a polyphenolic antioxidant, reduces oxidative stress, modulates mitochondrial function, and improves meiotic spindle integrity in aged oocytes, potentially alleviating age-related quality decline (Liu et al. 2018, Hayashi et al. 2019, Yoon et al. 2020). Previous studies in aged mouse models further demonstrated that resveratrol supplementation enhanced blastocyst development, improved implantation and pregnancy rates, and reduced reactive oxygen species production, while increasing mitochondrial membrane potential (Wang et al. 2017, Liu et al. 2018, Yoon et al. 2020). Several studies have reported that resveratrol improves oocyte maturation and embryo developmental competence in various species in vitro (Liu et al. 2013, Takeo et al. 2013, Liang et al. 2018, Madrid Gaviria et al. 2019).

Vitamin B12, a key cofactor in one-carbon metabolism, plays essential roles in DNA synthesis, methylation, and energy metabolism, all of which are critical for early embryonic development (Swelum et al. 2022). Cyanocobalamin, a form of vitamin B12, significantly enhances oocyte maturation, fertilisation, and embryo development in mice, whereas in sheep, it primarily improves oocyte maturation, with the most pronounced effect observed at 200 pM (Zacchini et al. 2017, Rostami et al. 2020). Similarly, in pigs, cobalamin supplementation during in vitro maturation improves oocyte quality and embryo development, likely through its antioxidant effects (Roy et al. 2020). Cyanocobalamin and resveratrol – compounds with antioxidant and mitochondrial-modulatory properties – are promising candidates for enhancing oocyte developmental competence in vitro.

Despite these promising candidates, the current application of IVF supplements remains largely empirical, with culture media and supplements typically applied uniformly across patients regardless of individual oocyte characteristics, neglecting substantial heterogeneity in oocyte morphology and quality differences that influence developmental potential and treatment responsiveness.

Traditional morphological evaluations are inherently subjective and are limited in their ability to capture heterogeneity. Radiomics has recently emerged as a powerful tool for the extraction of high-dimensional quantitative features from biomedical images. When applied to oocyte imaging, radiomics offers the potential to objectively classify oocytes based on their shape, texture, and intensity. However, existing radiomics studies have focused primarily on descriptive classification without linking phenotypic heterogeneity to functional outcomes under different environmental or culture conditions.

To address this gap, we aimed to move beyond static classification to actionable cultural guidance. Specifically, this study sought to determine whether radiomics-defined oocyte phenotypes exhibit differential developmental responses to targeted culture medium supplementation, providing a functional basis for phenotype-guided embryo culture.

Building on advances in radiomics and artificial intelligence-driven phenotyping, we applied an unsupervised clustering strategy to over 2,000 mouse oocytes to define distinct morphological subtypes. While our previous study demonstrated the feasibility of radiomics-based oocyte phenotyping under uniform culture conditions, the current study moves beyond descriptive classification to explore functional differences in culture medium responsiveness across phenotypes (Kim et al. 2025). We systematically examined how these phenotypes and their interactions with maternal age influenced blastocyst formation under six different culture conditions, including varying concentrations of resveratrol and vitamin B12. We aimed to determine whether radiomics-based phenotyping could inform personalised medium-selection strategies to enhance IVF outcomes.

To our knowledge, this is the first study to functionally link radiomics-defined oocyte phenotypes with differential responsiveness to specific culture medium supplements, thereby moving beyond descriptive classification to actionable individualised embryo culture strategies.

Materials and methods

Ethics approval

All animal procedures were performed in accordance with the National Institute of Health Guide for the Care and Use of Laboratory Animals. This animal study was approved by the Animal Research Ethics Committee of the Maria Fertility Research Institute (approval no. 2025-003).

Oocyte preparation and embryo culture

A total of 2,061 mouse oocytes were collected between January 2023 and December 2024. All oocytes were obtained from female B6D2F1 (C57BL/6 × DBA/2) mice (Koatech, Korea).

Two age groups were defined. For the young group, female B6D2F1 mice were purchased at 4 weeks of age and acclimated for 2–4 weeks before oocyte collection at 6–8 weeks of age. For the old group, oocytes were collected from female B6D2F1 mice maintained in-house at the Maria Fertility Research Institute until 66–70 weeks of age.

To induce ovulation, young mice received 5 IU of pregnant mare serum gonadotrophin (Prospec, Israel), followed 48 h later by 5 IU of human chorionic gonadotrophin (hCG; Sigma-Aldrich, USA). Old mice received the same hormonal regimen, except that the dose of both hormones was increased to 10 IU.

Cumulus–oocyte complexes were collected from the oviducts 14 h after hCG administration using pre-warmed MRC#OS medium (Maria Medical Foundation, Seoul, South Korea). Cumulus cells were removed with hyaluronidase (300 IU/mL) treatment for 3–5 min. Each oocyte was individually cultured in a 20 μL microdroplet of MRC#OS medium covered with mineral oil.

Images of the denuded oocytes were acquired using a Nikon microscope (Nikon, Japan) immediately after transfer.

Following image acquisition, IVF was performed in the same microdroplets using sperm isolated from the cauda epididymides of mature B6D2F1 male mice. Sperm were capacitated in MRC#OS fertilisation medium for 2 h at 37°C under 5% CO₂ prior to insemination.

After 5 h of insemination, presumptive zygotes were identified based on extrusion of the second polar body during the washing process and transferred to fresh 10 μL droplets of MRC#OS medium (one embryo per droplet) under mineral oil. Presumptive fertilisation was assessed based on second polar body extrusion and subsequent cleavage to the 2-cell stage; pronuclear formation was not examined.

Embryos were incubated at 37°C in a humidified atmosphere containing 6.0% CO₂, and morphological development was assessed daily until day 4, when the blastocyst stage was determined. Oocytes that did not develop to the blastocyst stage by day 4 were classified as non-blastocysts.

Oocyte imaging

Bright-field images of denuded oocytes were acquired using a Nikon inverted microscope equipped with a Nikon DS-Fi3 colour digital camera (Nikon, Japan). Images were captured at 200× magnification and a resolution of 2880 × 2048 pixels. To ensure consistency throughout the study, all images were acquired using an identical imaging protocol, including the same microscope system, camera, magnification, exposure time (30 ms), and analogue gain (2.2×). For each oocyte, a single bright-field image was obtained at the focal plane corresponding to the maximal oocyte diameter. Z-stack imaging was not performed. The acquired images were subsequently used for cytoplasmic ROI segmentation and radiomic feature extraction.

Image processing and region of interest (ROI) segmentation

The cytoplasmic ROIs were manually annotated by trained embryologists using binary segmentation masks. To ensure accuracy, all binary masks were converted to binary ROIs and processed using the PyRadiomics mask-correction function (correctMask = True) before feature computation. This procedure corrects potential mask inconsistencies and ensures that the segmented cytoplasmic region is represented as a valid connected ROI suitable for radiomic feature extraction. All images and masks were resampled to an isotropic pixel spacing of 1 × 1 to ensure uniformity in shape-based measurements. Additional details regarding image preprocessing, radiomic feature extraction settings, feature normalisation procedures, and clustering parameters are summarised in Supplementary Table 1 (see section on Supplementary materials given at the end of the article) to facilitate reproduction of the analytical workflow.

Radiomic feature extraction and normalisation

Radiomic feature extraction was performed using PyRadiomics (v3.1.0) (Van Griethuysen et al. 2017) and custom Python (v3.8.0) scripts. Eight two-dimensional morphological shape features were extracted, including the major axis length, elongation, sphericity, and perimeter (Supplementary Table 2). In addition, 16 first-order intensity features were computed, including mean, entropy, variance, and skewness (Supplementary Table 3).

Image intensities were normalised before feature extraction by centring the pixel intensity distribution to zero mean and unit variance, followed by scaling with a factor of 100. This preprocessing step was applied uniformly to all images to standardise radiomic feature extraction and reduce minor baseline intensity variation among images. All extracted feature values were subsequently standardised using Z-score normalisation before statistical analysis. All feature classes in PyRadiomics were disabled by default to avoid redundancy and enhance interpretability. Only two classes of morphological features (shape2D) and intensity features (first order) were selectively analysed.

Clustering analysis

The study comprised 19 experimental replicates. Because blastocyst formation may be influenced by culture conditions, the clustering framework was established using radiomic features from 695 Control oocytes collected during the first 12 baseline replicates, thereby defining the clusters independently of culture-condition effects. The fitted feature scaler and Gaussian mixture model (GMM) were subsequently applied, without refitting, to 1,366 oocytes collected during the remaining seven experimental replicates, which included six culture conditions: Control, three concentrations of resveratrol, and two concentrations of vitamin B12. These data were used to evaluate associations between culture condition and blastocyst formation within the predefined clusters.

Unsupervised clustering was performed using a GMM (Zhang et al. 2025a) with model selection based on the Akaike information criterion (AIC) and Bayesian information criterion (BIC) (Hajihosseinlou et al. 2024, Kim et al. 2024). GMM was selected because it provides a probabilistic clustering framework that models each cluster as a Gaussian distribution and estimates the probability of cluster membership for each oocyte. This approach is particularly suitable for radiomic data, where biological phenotypes may overlap and cluster boundaries may not be sharply separated. The optimal number of clusters was determined to be three. For each oocyte, the GMM estimated the probability of belonging to each cluster based on the multidimensional radiomic feature space, and the oocyte was assigned to the cluster with the highest membership probability. Clusters were labelled based on blastocyst formation rates. Clustering quality was assessed using the silhouette coefficient (Adiba & Siswanto 2024), and the results were visualised using principal component analysis (PCA) (Salloum et al. 2024).

Culture conditions and outcome assessment

Each oocyte was assigned to one of six culture conditions: a control group or one of five experimental groups supplemented with resveratrol (R5010, Sigma-Aldrich, USA) at 0.1, 0.5, or 1.0 μM, or vitamin B12 (V6629, Sigma-Aldrich, USA) at 50 or 200 pM.

These culture conditions were maintained throughout embryo development. Blastocyst formation was recorded as a binary outcome. All analyses were conducted at the individual oocyte level, linking image-derived features to blastocyst formation status, culture conditions, and age group. A schematic summary of the overall workflow, including mouse treatment, oocyte retrieval, imaging, feature extraction, clustering, and outcome evaluation, is shown in Fig. 1.

Figure 1.

Overview of study workflow.

Statistical analysis

Blastocyst formation rates were compared between the groups using either the chi-square test or Fisher’s exact test depending on the distribution of expected cell counts. Specifically, the chi-square test was applied when the expected cell frequency in all cells exceeded 5, and Fisher’s exact test was used when any expected frequency was ≤5 (Smith et al. 2024, Lee et al. 2025). Differences in radiomic features among clusters were evaluated using independent two-sample t-tests. As each comparison group included more than 30 samples, normality was assumed based on the central limit theorem (Ghasemi & Zahediasl 2012, Fukuda 2024). Statistical significance was defined as a two-sided P-value of <0.05. All statistical analyses were conducted using the scipy.stats (v1.10.1) (Chandra et al. 2022) module in Python (v3.8.0), and results were visualised using the matplotlib library.

Results

Cluster definition and developmental characteristics

GMM-based clustering of radiomic features identified three distinct oocyte clusters. The GMM-derived clusters showed clear separation when visualised using PCA, supporting the presence of three phenotypically distinct clusters (Fig. 2). The latent feature distribution and the estimated Gaussian mixture components underlying the clustering model are further illustrated in Supplementary Fig. 1. Cluster structure was further supported by a silhouette coefficient of 0.351. To characterise the developmental potential of each cluster, blastocyst formation rates were compared across the three groups using oocytes cultured in the control medium only to avoid the confounding effects of culture media. The results are summarised in Table 1. Cluster 1 exhibited a significantly lower blastocyst formation rate (28.2%, 44/156), whereas Cluster 2 exhibited the highest rate (42.4%, 156/368). Cluster 3 had an intermediate rate of 34.9% (105/301). Statistical comparisons confirmed that Cluster 2 had significantly higher blastocyst formation rates than Clusters 1 (P = 0.002) and 3 (P = 0.048). The difference between Clusters 1 and 3 was not statistically significant (P = 0.149). Additional subgroup analysis by age revealed that in all three clusters, young oocytes had significantly higher blastocyst formation rates than old oocytes, as shown in Supplementary Table 4 (Cluster 1: P = 0.011; Cluster 2: P = 0.040; Cluster 3: P < 0.001). However, among young oocytes, the blastocyst formation rates were similar across clusters, suggesting an equivalent developmental potential. By contrast, among old oocytes, Cluster 2 maintained significantly better outcomes than Clusters 1 and 3 (P < 0.05).

Figure 2.

Gaussian mixture model-based cluster assignment results. (A) Akaike information criterion (AIC) and Bayesian information criterion (BIC) values for Gaussian mixture models with different numbers of components. The lowest BIC value was used to determine the optimal number of clusters. (B) Three-dimensional principal component analysis (PCA) visualisation of the radiomic feature space. Each point represents an individual oocyte, coloured according to its assigned cluster. The axes correspond to the first three principal components, illustrating the distribution and separation of the three GMM-derived clusters. AIC, Akaike information criterion; BIC, Bayesian information criterion; PCA, principal component analysis; GMM, Gaussian mixture model.

Table 1.

Blastocyst formation rates across radiomics-based clusters.

	Cluster 1	Cluster 2	Cluster 3
Blastocysts, n	44	156	105
Non-blastocysts, n	112	212	196
Total, n	156	368	301
Blastocyst formation rate, %	28.2	42.4	34.9
P value
vs Cluster 1		0.002	0.149
vs Cluster 2	0.002		0.048
vs Cluster 3	0.149	0.048

Radiomic feature distributions across clusters

Quantitative analysis of radiomic features revealed distinct morphological and intensity-based phenotypes across the three clusters (Supplementary Table 5). Morphological features, such as major axis length, pixel surface, and perimeter, followed the pattern: Cluster 3 < Cluster 1 < Cluster 2, indicating that Cluster 2 oocytes had the largest overall size. The elongation feature, which decreased with increasing elongation, was the lowest in Cluster 2 and highest in Cluster 1 (2 < 3 = 1), confirming that Cluster 2 oocytes exhibited the most elongated morphology. Sphericity was highest in Cluster 2, whereas Clusters 1 and 3 exhibited similar but lower values (1 = 3 < 2). The perimeter surface ratio was the highest in Cluster 3 (2 < 1 < 3), indicating greater boundary irregularity despite its smaller size.

First-order intensity features also differed significantly across the clusters. Cluster 3 exhibited the highest mean, median, and minimum intensities, reflecting overall brightness. However, Cluster 1 showed the greatest intensity range and total energy, suggesting a higher internal contrast and signal strength. Notably, entropy, variance, and mean absolute deviation followed the pattern, Cluster 3 < Cluster 1 < Cluster 2, indicating that Cluster 2 oocytes had the most heterogeneous cytoplasmic texture, whereas Cluster 3 oocytes were the most uniform. These observations were supported by the distribution visualisations in boxplots (Supplementary Figs 2 and 3) and were tested using independent two-sample t-tests (Supplementary Table 5). Representative bright-field images of oocytes from each radiomics-defined cluster, together with the corresponding cytoplasmic ROI overlays, are presented in Supplementary Fig. 4. Representative examples illustrating differences in cytoplasmic texture characteristics, including high-and low-entropy oocytes, are shown in Supplementary Fig. 5. These visual examples are consistent with the radiomics-derived differences in both morphological and texture-related features observed among the clusters and provide qualitative illustrations of the features used for radiomics-based classification.

Effect of culture media within each cluster

To evaluate the interaction between morphological subtype and culture environment, blastocyst formation rates were compared among the five experimental culture media and control conditions within each cluster (Supplementary Table 6, Fig. 3). In Cluster 1, which showed the lowest baseline development under control conditions (28.2%), the experimental media significantly improved the developmental outcomes. Vitamin B12 increased the blastocyst formation rate to 46.5% at 50 and 200 pM (P = 0.001) and to 48.7% at 200 pM (P < 0.001). Resveratrol also significantly increased the blastocyst formation rate to 45.1% at 1.0 μM (P = 0.004), while 0.5 μM resulted in a modest but statistically significant improvement (39.8%, P = 0.048).

Figure 3.

Cluster-specific effects of culture media on blastocyst formation compared with control. Blastocyst formation rates are shown for each radiomics-defined oocyte cluster under control, resveratrol, and vitamin B12 supplementation conditions. Bars represent the percentage of oocytes that developed to the blastocyst stage within each treatment group. Asterisks indicate statistically significant differences compared with the corresponding control group within the same cluster.

In Cluster 3, which had an intermediate baseline blastocyst formation rate of 34.9% under control conditions, vitamin B12 produced the most pronounced effect at 50 pM, increasing the blastocyst formation rate to 60.4% (P = 0.001). Resveratrol also led to a significant improvement at 0.1 μM, raising the blastocyst formation rate to 50.0% (P = 0.010). The other media did not result in significant changes. By contrast, Cluster 2, which exhibited the highest baseline blastocyst formation rate under control conditions (42.4%), did not show statistically significant improvements with any of the five media. The highest observed rate was 57.9% with resveratrol 0.5 μM, which did not reach significance (P = 0.067).

Combined effect of culture medium and age

Blastocyst formation rates were compared across the five experimental media and a control medium within each age group and cluster (Table 2 and Fig. 4). In Cluster 1, young oocytes showed the most robust response to vitamin B12 200 pM, with a blastocyst formation rate of 57.3%, which was significantly higher than that of the control group (35.9%, P = 0.004). Resveratrol 0.1 μM improved the blastocyst rate to 55.0% (P = 0.020), while other treatments showed numerical increases that were not statistically significant: resveratrol 0.5 μM (48.0%, P = 0.176), resveratrol 1.0 μM (49.2%, P = 0.094), and vitamin B12 50 pM (48.3%, P = 0.093). In old oocytes, resveratrol 1.0 and 0.5 μM significantly improved blastocyst formation to 39.2% (P = 0.008) and 33.9% (P = 0.032), respectively, compared with 17.2% in the control. Vitamin B12 50 pM (43.9%) and 200 pM (36.9%) led to significant improvements (P = 0.001 and P = 0.012, respectively).

Table 2.

Combined effect of age and culture medium on blastocyst formation.

Cluster/age/culture medium	Blastocysts (n)	Non-blastocysts (n)	Total (n)	BL rate (%)	P value (vs control)
Cluster 1
Young
Control	33	59	92	35.9
Res 0.1	33	27	60	55.0	0.020
Res 0.5	22	24	46	47.8	0.176
Res 1.0	32	33	65	49.2	0.094
B12_50	42	45	87	48.3	0.093
B12_200	51	38	89	57.3	0.004
Old
Control	11	53	64	17.2
Res 0.1	10	39	49	20.4	0.663
Res 0.5	21	41	62	33.9	0.032
Res 1.0	19	29	48	39.6	0.008
B12_50	25	32	57	43.9	0.001
B12_200	24	41	65	36.9	0.012
Cluster 2
Young
Control	97	109	206	47.1
Res 0.1	9	7	16	56.3	0.480
Res 0.5	13	11	24	54.2	0.511
Res 1.0	7	8	15	46.7	0.975
B12_50	11	7	18	61.1	0.254
B12_200	12	5	17	70.6	0.063
Old
Control	59	103	162	36.4
Res 0.1	6	13	19	31.6	0.759
Res 0.5	9	5	14	64.3	0.040
Res 1.0	3	11	14	21.4	0.388*
B12_50	8	16	24	33.3	0.864
B12_200	4	10	14	28.6	0.774*
Cluster 3
Young
Control	72	89	161	44.7
Res 0.1	29	22	51	56.9	0.130
Res 0.5	46	40	86	53.5	0.189
Res 1.0	18	39	57	31.6	0.083
B12_50	17	11	28	60.7	0.118
B12_200	14	14	28	50.0	0.605
Old
Control	33	107	140	23.6
Res 0.1	16	23	39	41.0	0.031
Res 0.5	18	35	53	34.0	0.144
Res 1.0	15	33	48	31.3	0.292
B12_50	12	8	20	60.0	<0.001
B12_200	9	14	23	39.1	0.114

^*Fisher’s exact test.

Figure 4.

Age-stratified analysis of culture medium effects on blastocyst formation within each cluster. Blastocyst formation rates are shown for young and old oocytes within each radiomics-defined cluster under control, resveratrol, and vitamin B12 supplementation conditions. Bars represent the percentage of oocytes that developed to the blastocyst stage within each cluster–age–treatment subgroup. Asterisks indicate statistically significant differences compared with the corresponding control group within the same cluster–age subgroup.

In Cluster 2, young oocytes exhibited high baseline performance (47.1%) and did not show significant improvements across any treatment, although B12 200 pM yielded the highest increase of 70.6% (P = 0.063). In old oocytes, only resveratrol 0.5 μM produced a significant increase in the blastocyst formation rate to 64.3 versus 36.4% in control (P = 0.040).

In Cluster 3, old oocytes responded strongly to vitamin B12 50 pM, with the blastocyst formation rate increasing from 23.6 to 60.0% (P < 0.001). Resveratrol 0.1 μM also significantly increased the rate to 41.0% (P = 0.031), while the differences with other treatments did not reach statistical significance. Among young oocytes, none of the treatment groups showed significant differences compared with the control (44.7%), although B12 50 pM (60.7%, P = 0.118) and resveratrol 0.1 μM (56.9%, P = 0.130) exhibited favourable trends.

Effect of optimal culture medium across cluster age combinations

In a post hoc analysis of the existing dataset, we compared blastocyst formation rates under control conditions with those observed for the culture condition showing the highest blastocyst formation rate within each of the six cluster–age subgroups. The optimal supplementation condition identified for each cluster–age subgroup is summarised in Supplementary Table 7. For all the combinations, the optimised medium yielded higher blastocyst formation rates than the control medium. For example, in the case of old oocytes, the rate increased from 17.2 to 43.9% in Cluster 1 and from 23.6 to 60.0% in Cluster 3. When the data were aggregated across all combinations, the overall blastocyst formation rate increased from 36.97% (control) to 56.00% (best medium), representing a statistically significant improvement (P < 0.001, chi-square test; Table 3).

Table 3.

Overall blastocyst formation rate comparison.

	Control	Optimal supplementation
Blastocysts, n	305	126
Non-blastocysts, n	520	99
Total, n	825	225
Blastocyst formation rate, %	36.97	56.00

Discussion

In this study, we applied a radiomics-guided unsupervised clustering approach to classify oocytes into distinct phenotypic groups and demonstrated that these phenotypes exhibit differential developmental responses to culture medium supplementation. Using a dataset of 2,061 mouse oocytes, we identified three phenotypic clusters that differed in baseline blastocyst formation rates and in their responsiveness to resveratrol and vitamin B12 across maternal age groups. These findings extend oocyte phenotyping beyond descriptive classification towards a functional, clinically relevant framework for phenotype-guided embryo-culture optimisation.

The three clusters exhibited distinct morphological and cytoplasmic characteristics. The high-performing cluster was characterised by larger oocyte size, higher sphericity, and greater cytoplasmic textural heterogeneity, whereas the lower- and mid-performing clusters showed smaller size, less regular morphology, and more uniform intensity distributions (Supplementary Figs 2, 3, 6). These differences, which are supported by feature distributions and visualisations (Supplementary Figs 2 and 3), likely reflect intrinsic variation in cytoplasmic organisation, mitochondrial distribution, and metabolic capacity, all of which are known to influence oocyte developmental competence. Importantly, these features were quantified objectively, reducing the subjectivity inherent in conventional morphological assessment.

A key finding of this study is that responsiveness to culture supplementation was strongly phenotype-dependent. Oocytes in the lower-performing and mid-performing clusters demonstrated substantial improvements in blastocyst formation under specific supplementation conditions, whereas the high-performing cluster showed limited responsiveness. This pattern suggests that oocytes with intrinsically optimal characteristics already operate near a developmental ceiling, making additional supplementation less effective. By contrast, oocytes with suboptimal baseline phenotypes appear to retain greater functional plasticity and may benefit from targeted modulation of oxidative stress, mitochondrial activity, or metabolic pathways.

Maternal age further modified supplementation responsiveness. Across clusters, young oocytes exhibited higher baseline developmental competence, whereas old oocytes showed greater variability and, in certain phenotypes, pronounced responsiveness to supplementation. Notably, some phenotypes retained substantial developmental plasticity despite ageing, achieving marked improvements in blastocyst formation under specific culture conditions. These findings indicate that chronological age alone is insufficient to predict oocyte functional potential and that phenotype–age interactions are critical determinants of treatment response.

The phenotype-specific supplementation patterns observed here provide a plausible explanation for heterogeneous results reported in previous studies. Although resveratrol improves oocyte quality by reducing oxidative stress and enhancing mitochondrial function, the reported optimal concentrations vary widely (Takeo et al. 2013, Wang et al. 2014, Sovernigo et al. 2017). Similarly, vitamin B12 plays a key role in one-carbon metabolism and redox balance; however, its effects on embryo development are context -dependent (Zacchini et al. 2017, Rostami et al. 2020, Roy et al. 2020, Swelum et al. 2022). Our findings indicate that such variability may arise, at least in part, from underlying phenotypic heterogeneity among oocytes rather than from experimental inconsistency alone.

From a clinical perspective, these results challenge the long-standing ‘one-size-fits-all’ paradigm in embryo culture. Rather than uniformly applying identical media and supplementation protocols to all oocytes, our data support a stratified approach in which culture conditions are tailored to intrinsic oocyte phenotypes and maternal age. Importantly, this strategy provides guidance not only on which supplements may improve developmental outcomes but also on which conditions offer little or no benefit. Defining ineffective or unnecessary interventions is essential for avoiding overtreatment and minimising metabolic or oxidative stress on embryos.

This phenotype-guided, targeted approach has meaningful implications for IVF practice. By selectively improving outcomes in oocyte subgroups with lower baseline competence, personalised culture strategies may increase overall blastocyst yield without additional ovarian stimulation or invasive interventions. Concurrently, avoiding unnecessary supplementation in non-responsive or already high-performing oocytes may enhance safety and efficiency. Phenotypic classification can be derived from routine microscope-acquired images and implemented without major changes to existing laboratory workflows. When combined with phenotype-specific response profiles, this framework could be incorporated into AI-assisted IVF systems to support real-time culture medium selection, standardise practice across laboratories, and reduce operator-dependent variability.

Compared with our earlier radiomics-based clustering study (Kim et al. 2025), which demonstrated that cytoplasmic morphological heterogeneity could predict blastocyst formation, the present investigation extends the framework towards clinical applicability. By integrating a larger dataset and incorporating maternal age and culture media as functional variables, this study connects quantitative phenotype classification to outcome-driven optimisation.

Unlike the previous descriptive model, which characterised morphology without considering environmental modulation, the current analysis reveals phenotype-specific responsiveness to different culture conditions. This design more closely mirrors the decision-making context in IVF laboratories, where embryologists must determine whether certain oocytes might benefit from antioxidant or metabolic supplementation.

Some limitations of this study should be acknowledged. First, all data were derived from a single research centre and generated using a mouse model. Therefore, external validation and further studies in human oocytes will be required to determine the generalisability of the identified phenotypes and supplementation responses.

Second, the biological mechanisms underlying the radiomics-defined phenotypes were not directly evaluated. Although differences in cytoplasmic organisation, mitochondrial function, metabolic activity, and genetic background may contribute to the observed clustering patterns, these factors were not measured in the present study. Because embryos were generated from B6D2F1 × B6D2F1 crosses, genetic heterogeneity may also have contributed to differences in developmental competence and supplementation responsiveness. Future studies incorporating functional, molecular, and genetic assessments will be required to clarify the biological basis of the identified phenotypes.

Third, ROI segmentation was performed manually, introducing potential operator variability. The future adoption of automated segmentation approaches may further enhance scalability and reproducibility. Fourth, MRC#OS is a proprietary, bicarbonate-buffered medium formulated for single-step culture from fertilisation to the blastocyst stage. Although it is manufactured under standardised quality-control procedures, its detailed composition is commercially confidential and, therefore, could not be fully reported, which may limit independent replication of the culture conditions. Fifth, a matched vehicle control was not included; however, resveratrol was delivered in 0.0005% DMSO, a concentration that showed no detectable effect on embryonic development in preliminary testing. Nevertheless, the absence of a matched vehicle control limits our ability to completely exclude a potential contribution of DMSO to the observed outcomes. Sixth, fertilisation was inferred from second polar body extrusion and subsequent cleavage rather than confirmed by pronuclear assessment. Therefore, failed fertilisation, parthenogenetic activation, and post-fertilisation developmental arrest could not be definitively distinguished among oocytes that did not form blastocysts. Seventh, the radiomic analysis was based on a single two-dimensional bright-field image. Consequently, the extracted features may not fully represent the three-dimensional morphology or spatial heterogeneity of the entire oocyte. Future studies using multiple focal planes or Z-stack imaging are needed to determine whether volumetric information provides a more complete characterisation of oocyte phenotype. Eighth, the present findings are based on post hoc associations observed in the existing dataset and do not establish that cluster-guided supplementation improves developmental outcomes. Morphological features showed substantial overlap among clusters, whereas cluster separation was more apparent in the first-order intensity features, whose biological significance remains uncertain. Prospective studies in which oocytes are classified before predefined culture-condition assignment are required to validate this approach.

In summary, this study demonstrated that oocyte phenotype and maternal age jointly determine responsiveness to culture medium supplementation, supporting a shift from empirical, one-size-fits-all medium application towards a rational, stratified, and data-driven culture strategy. In practice, such information could guide the design of individualised embryo culture protocols, contributing to improved cumulative pregnancy rates through the production of more embryos, while simultaneously reducing unnecessary interventions. This targeted approach promises higher implantation rates, reduced overtreatment, and safer, more efficient IVF protocols.

In conclusion, radiomics-based oocyte phenotyping identified clusters with different baseline blastocyst formation rates and heterogeneous associations between culture condition and developmental outcome. These findings provide a basis for prospective studies evaluating whether cluster-guided supplementation can improve blastocyst formation.

Declaration of interest

H M Kim is an employee of Kai Health. H Lee is a shareholder of Kai Health. The other authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the work reported.

Funding

This work was supported by the Korea Medical Device Development Fund by the Korea government (Ministry of Science and ICT, Ministry of Trade, Industry and Energy, Ministry of Health & Welfare, Ministry of Food and Drug Safety) (grant number RS-2023-00254175).

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author. Data regarding any of the subjects in the study have not been previously published unless specified. Data will be made available to the editors of the journal for review or query upon request.

Code availability

The core custom Python code used for radiomic feature selection, standardisation, Gaussian mixture model fitting, and cluster assignment is publicly available at https://github.com/hikhm/mouse-oocyte-gmm-clustering. The repository includes a synthetic-data generator to demonstrate the expected input structure and verify code execution. The synthetic data are not the original study data and are not intended to reproduce the numerical results reported in the manuscript.