Introduction

The nucleus of mammalian cells is organized into functionally specialized compartments that are required for the control of gene expression and differentiation. In the interphase cell nucleus, heterochromatin domains are organized as highly condensed, transcriptionally repressive nuclear aggregates or chromocenters that are essential for accurate chromosome segregation and maintenance of genome stability1,2. The oocyte nucleus, also known as the germinal vesicle (GV), exhibits a unique chromatin configuration that is subject to dynamic modifications during the final stages of post-natal oocyte growth. Functional differentiation of chromatin structure in mammalian pre-ovulatory oocytes is characterized by striking changes in nuclear architecture in preparation for meiosis onset. This process takes place through a mechanism of large-scale chromatin remodeling and is a critical component of the epigenetic maturation that is required to confer fully-grown oocytes with both meiotic and developmental potential1,2. In both human and mouse oocytes, a decondensed chromatin configuration is initially present to support the high levels of global transcription required to synthesize and accumulate maternal mRNA stores essential for meiotic maturation and the first cleavage divisions of the embryo3,4,5. In these oocytes, heterochromatin domains or chromocenters lack any association with the nucleolus resulting in the formation of a non-surrounded nucleolus (NSN) configuration6. In situ hybridization and live cell imaging with major satellite DNA probes to visualize heterochromatin domains revealed that chromocenters become progressively associated with the nucleolus and subsequent large-scale chromatin remodeling leads to the formation of a peri-nucleolar rim resulting in the transition into a surrounded nucleolus (SN) configuration in preparation for meiosis onset6,7,8. This critical developmental milestone is also accompanied by global repression of nascent transcriptional activity in the oocyte genome3,4,5.

Both genetic as well as pharmacological manipulation of large-scale chromatin structure indicate that remodeling chromatin into the SN configuration may provide pericentric heterochromatin with a functional configuration required for timely meiotic progression and accurate chromosome segregation6. Consistent with this notion, NSN oocytes exhibit abnormal meiotic maturation and frequently fail to develop beyond the 2-cell stage following in vitro fertilization9,10,11. Thus, accumulating evidence indicates that remodeling chromatin into the SN configuration is of critical importance to confer the mammalian oocyte with full meiotic and developmental potential1,6,9,10,11.

Identification of cellular or molecular markers of oocyte quality and developmental potential is a pressing need and a step of critical importance to implement both clinical and research applications of select competent oocytes in the fields of human infertility as well as animal reproduction. However, until now, detection and/or quantification of any potential reliable cellular marker(s) of developmental potential requires invasive procedures, such as protein crosslinking and immuno-fluorescent analysis of individual oocytes limiting any subsequent downstream applications.

Recent advances in machine learning now make it possible to develop computational deep neural networks with the capacity to recognize critical information from unlabeled microscopic images of cells that are not immediately apparent to the human eye without prior invasive analysis12,13. Computational neural networks such as label-free imaging13 and similar modalities such as In Silico labeling12 represent a paradigm shift in the non-invasive analysis of cellular function as they allow for the acquisition of quantitative information on organelles and cellular substructure from phase contrast or transmitted light images in real-time, allowing for accurate multi-parametric analysis of live cells12,13. Here, we leveraged the analytical power of state-of-the-art methods in deep learning network development to predict the state of chromatin configuration and developmental potential in unlabeled mouse pre-ovulatory oocytes. Our strategy for the non-invasive selection of oocytes with high developmental potential is focused on fluorescent image prediction of highly resolved chromatin configuration at the germinal vesicle stage and the localization of histone tri-methylation at lysine 9 (H3K9me3), an epigenetic mark required for maintenance of chromosome structure during cell division14, two bona fide markers directly associated with oocyte meiotic and developmental potential.

Our approach combines a fluorescence prediction network with a classification step to inform chromatin state and genome integrity. Chromatin information was extracted following nuclear segmentation using a deep neural network trained to predict chromatin fluorescence maps from bright-field images. The fluorescence maps were then scanned by three classification networks to identify whether the oocyte exhibits an NSN or SN configuration. Our pipeline achieved 91.3% accuracy in the classification of fixed oocytes and 85.7% accuracy in the classification of live oocytes. Machine learning tools have been recently developed for the analyses of mammalian gametes15,16. However, to the best of our knowledge, this is the first neural network pipeline specifically designed for the non-invasive prediction of chromatin structure and developmental potential in live mammalian oocytes using label-free, bright field images. Our deep learning selection pipeline is fast and efficient in providing important information on oocyte chromatin state within minutes. As a proof of concept, we demonstrate its application for the non-invasive selection and downstream transcriptional profiling of oocytes with high developmental potential, generating a unique dataset of meiotic and developmentally competent oocytes without prior exposure to any fluorescent DNA-binding ligand stains that may exert a detrimental effect on chromatin structure and/or the oocyte transcriptome.

Results

Development of a deep learning pipeline for the non-invasive prediction of chromatin configuration in pre-ovulatory oocytes

Live cell imaging of mammalian oocytes provides information-rich, phase contrast or transmitted light micrographs in which the large size, positional, and structural information on both the oocyte nucleus as well as the nucleolus are readily apparent (Fig. 1A). However, the type of chromatin configuration present on each oocyte is not detectable without previous invasive immunofluorescence analysis of nuclear architecture (Fig. 1B). Thus, we developed a deep learning pipeline, that can accurately predict fluorescent images of chromatin configuration from bright-field, live oocyte images and then generate an NSN/SN classification. Our pipeline can efficiently ‘Cast’ a fluorescence prediction of large-scale chromatin structure, at high resolution on a bright-field image, allowing for the non-invasive selection of live oocytes. Importantly, based on the accurate fluorescence prediction and identification of SN oocytes, our pipeline not only discerns chromatin structure but also precisely selects for live oocytes with high meiotic and developmental potential.

Fig. 1: Major chromatin remodeling during the transition from the NSN to SN configuration in GV stage oocytes.
figure 1

A Bright-field image of a GV stage oocyte under confocal microscopy. The region of interest (square) enclosing the nucleus (blue dashed line) is segmented and utilized for the analysis of chromatin configuration. The position of the nucleolus is indicated (asterisk). B During the transition from NSN to SN configuration, oocytes acquire a distinct chromatin rim around the nucleolus (asterisks), which is associated with increased developmental and meiotic potential and concomitant with decreased transcriptional activity. Invasive fluorescent labeling is currently standard practice for the determination of chromatin configuration, as shown using H3K9me3 immunochemistry (red) and DAPI chromatin labeling (blue). NSN oocytes exhibit a decondensed chromatin configuration, while SN oocytes show a compact chromatin structure and a bright chromatin rim surrounding the nucleolus.

Full size image

The structure of our Fluorescence prediction and Classification on Bright-field (Fluo-Cast-Bright) pipeline17 is illustrated in Fig. 2. Our pipeline was initially trained using fixed oocytes following protein crosslinking and immunofluorescence analysis of nuclear architecture. Stacks of bright field confocal images were used to generate a 3-D reconstruction of the entire oocyte nucleus or germinal vesicle for accurate nuclear segmentation (Fig. 2A). To train our fluorescence prediction network, we used the corresponding, 3-D pixel registered images (all the channels represent the same spatial localization) within our region of interest of two target labels, DAPI and H3K9me3, to determine the type of chromatin configuration present on each oocyte (Fig. 2A).

Fig. 2: Fluorescence prediction and Classification on Bright-field (Fluo-Cast-Bright) pipeline.
figure 2

A The fluorescence prediction network was constructed based on the UNet++ architecture and was trained using 3-D bright-field images as input and the combined fluorescence maps of DAPI (DNA) and H3K9me3 as target labels. B The classification networks were constructed based on the ResNeXt101 architecture and were trained using predicted fluorescence maps from the best fluorescence prediction network as input and the correct NSN/SN label as the target label. C The workflow of the Fluo-Cast-Bright pipeline using a 3-D bright-field image stack from an oocyte with an unknown chromatin configuration.

Full size image

To devise a fluorescence prediction model, we used a convolutional neural network (CNN) with a three-dimensional, U-Net++18 architecture (Fig. 2A and Supplementary Fig. 1A) for image segmentation and translation tasks, which is capable of extracting information from high-level image features, such as nuclear morphology as well as low-level features, such as lines and edges and generate a comprehensive interpretative system that uses relative position and image features to identify and segment objects and discover hidden features for image translation18. This architecture is further enhanced by skip pathways, which allow linking low-level features to higher-level understanding of the image to improve performance18 (Supplementary Fig. 1A). To achieve fluorescence prediction, the network is trained with randomly sampled 3-D images and with a loss function, which guides the network to produce a predicted fluorescent map with minimal mean-square error compared to the ground truth or actually combined fluorescence maps of DAPI and H3K9me3 localization in the germinal vesicle (Fig. 2B). Our classification network is based on the ResNeXt-10119 architecture (Fig. 2B; Supplementary Fig. 1B), which contains identity shortcut connections that enable complex computational analysis to improve image classification19. The classification network was then trained to distinguish between NSN and SN oocytes using 2-D images of the 3-D predicted fluorescence maps generated from the fluorescence prediction model (Fig. 2B). The 2-D predicted fluorescence maps and the manually annotated NSN/SN classification for each image were used for the training of the classification networks (Fig. 2B).

Selection of these two architectures was based following a direct comparison of eight different model architectures and specific parameters for their applicability across many image analytical tasks, as well as their optimal efficacy in predicting chromatin state, specifically. The UNet++ architecture demonstrated superior performance in our region of interest prediction with the lowest mean squared error when compared to other commonly used architectures (Supplementary Fig. 2A). Moreover, the ResNeXt-101 architecture exhibited a greater classification accuracy compared to other commonly used architectures in our fluorescence map classification task (Supplementary Fig. 2B). Importantly, hyper-parameter testing also demonstrated the optimal set of parameters used in our model (Supplementary Fig. 2C–F).

To train our fluorescence prediction and classification model pipeline (Fig. 2A and B), we used the 3-D bright-field and fluorescence image stacks of 138 fixed oocytes, out of which 96 were used for training and 42 were used for validation. The model with the best performance on the images used for validation was designated as the best model, and the predicted fluorescence map of 42 validation images produced from the best model was then used to train the three classification models using three different train-valid splits. We used 3-D images to train the fluorescence prediction model and 2-D images to train the classification models (Fig. 2C). While the relationship between different 2-D cross-sections contributes to fluorescence prediction, a single 2-D cross-section contains sufficient information to conduct the classification of an oocyte as NSN or SN. Thus, we used 16, 2-D fluorescence maps from each 3-D predicted fluorescence map stacks (672, 2-D fluorescence maps in total), to train the three classification models (Fig. 2B and C). The best classification models were chosen according to their validation performance.

To test the efficacy of our non-invasive classification approach, we tested our Fluo-Cast-Bright pipeline on unlabeled bright-field images of 23 fixed oocytes. Single oocytes were transferred into individually labeled wells of a glass bottom 96-well plate and photographed one at a time to carefully track each oocyte. These bright-field oocyte images were analyzed using our pipeline to assign an NSN- or SN-predicted label to each oocyte (Fig. 2C). Following the initial computational classification, DAPI staining was used for individual analysis of the same fixed oocytes to reveal the actual chromatin configuration NSN or SN of each oocyte using confocal microscopy. Comparison between the predicted chromatin state and the actual chromatin state (or ground truth) revealed that the chromatin configuration of 21 out of 23 oocytes was predicted correctly (accuracy = 91.3%; Fig. 3A). We then conducted a second test on fixed oocytes. This time, we imaged 23 fixed and unlabeled oocytes using glass-bottom Petri dishes, and fixed oocytes were transferred to micro drops of MEM medium (50 μl) under mineral oil to simulate the conditions for live oocyte imaging. Comparison between the computationally predicted chromatin state and actual chromatin state determined by DAPI staining and confocal microscopy revealed that the chromatin configuration of 21 out of 23 oocytes was predicted correctly (accuracy = 91.3%). Importantly, analysis of three additional predictive metrics revealed similarly high values (precision = 91.4%, recall = 96.4%, F1 score = 93.8%), demonstrating that the performance of our neural network is transferable across different imaging conditions in fixed oocytes and provides proof of concept for the accurate prediction of chromatin state using only bright field images.

Fig. 3: Accuracy and application procedure of the Fluo-Cast-Bright pipeline.
figure 3

A Bar plot showing the classification performance of the Fluo-Cast-Bright pipeline on fixed oocytes (n = 2 independent tests). B Pool of live GV stage oocytes prior to imaging, segmentation, and non-invasive classification. C Bar plots showing the classification performance of Fluo-Cast-Bright pipeline on live oocytes (n = 2 independent tests). D Segmented nuclear regions of bright-field images and classification of the chromatin configuration as either NSN or SN using the Fluo-Cast-Bright pipeline. E The classified oocytes can be used for downstream applications or fixed and labeled to determine the accuracy of the pipeline. Corresponding DAPI (Cyan) label of the bright-field images shown in (D) for ground truth validation of chromatin state. All oocytes except those demarcated (red X) were accurately predicted by the Fluo-Cast-Bright pipeline.

Full size image

Non-invasive prediction of chromatin configuration in live pre-ovulatory oocytes

Next, we tested the prediction accuracy of our Fluo-Cast-Bright pipeline using live oocytes (Fig. 3B), as this is critical for the widespread research and clinical applications of our pipeline. Live cell imaging revealed the presence of subtle, albeit detectable movements of the oocyte nucleus. Thus, the prediction accuracy of our Fluo-Cast-Bright model decreased due to intracellular movement of subcellular components in live oocytes, which were detectable in the <15 min that were required to collect the bright-field image stacks of the entire nucleus. Although these movements may affect both the fluorescence prediction as well as the classification steps, we found that re-training of the classification network alone with live oocytes was sufficient to improve the accuracy of our pipeline as the networks were able to identify the effect of movement in predicted fluorescent maps and eliminate it during the classification step. We completed this re-training using bright-field image stacks generated from 63 live oocytes and used fourteen 2-D fluorescence maps from each 3-D predicted fluorescence map stack (882, 2-D fluorescence maps in total) to re-train the three classification models. The best classification models were again chosen according to the results of the validation performance.

After re-training the classification networks, we conducted a test using 42 new live oocyte images, which did not participate in any of the training or the re-training processes. Unlabeled bright-field images were analyzed by the pipeline to generate an initial NSN or SN classification, based on a non-invasive computational analysis of live oocyte images. The actual chromatin state or ground truth was revealed by DAPI staining and confocal microscopy after the predictions were generated. This test was performed in two independent experimental replicates with 21 live oocytes per replicate (Fig. 3B) collected over the course of 2 days. Comparison between the predicted chromatin state and actual chromatin ground truth revealed that the chromatin configuration of 18 out of 21 oocytes was predicted correctly (accuracy = 85.7%, precision = 87.1%, recall = 87.1%, F1 score = 87.0%) in each group (Fig. 3C). A workflow showing the processes by which live oocyte accuracy testing was conducted and how the Fluo-Cast-Bright pipeline can be used to select both NSN or SN oocytes for potential research and or clinical applications is illustrated in Fig. 3D and E.

Identification of imaging features critical for deep learning pipeline performance

Comparison between the predicted fluorescence maps and the bright-field images in Fig. 4A, revealed that the fluorescence prediction network learnt to identify chromatin structure by detecting readily apparent information on the number and position of the nucleolus in the bright field images (Fig. 4A asterisks), as well as hidden features from the adjacent euchromatin and heterochromatin regions. Importantly, comparison of the predicted fluorescent maps with the ground truth fluorescent maps of both fixed and live oocytes revealed remarkable similarities allowing to distinguish differences in the extent of euchromatin decondensation in the NSN oocytes and faithfully representing the presence of two nucleoli in some oocytes (Fig. 4A, asterisks). Predicted fluorescence maps from fixed SN oocytes reveal that the network was also able to detect the presence of prominent heterochromatin foci or chromocenters (arrowheads) and large-scale chromatin configuration patterns of the image correctly, while also placing a brighter and more distinct chromatin ring surrounding the nucleolus. Importantly, predicted fluorescence maps from fixed NSN oocytes were also accurate in showing a more decondensed euchromatin structure that occupies the entire germinal vesicle. The live SN oocytes, on the other hand, produced dimmer chromocenters (Fig. 4A; dashed rectangles) in the resulting predicted fluorescence maps, compared to the predicted fluorescence maps obtained from fixed SN oocytes (arrowheads), primarily due to the increased intracellular movement in live oocytes.

Fig. 4: Imaging features critical for deep learning pipeline performance.
figure 4

A Comparison between true and predicted fluorescence maps demonstrates that the fluorescence prediction network is capable of accurately identifying both nucleolar structure (asterisks) and critical heterochromatin domains (arrowheads and dashed boxes). B Class activation maps illustrate the importance ranking (color scale) of regions within images used by the classification networks.

Full size image

Some oocytes exhibit a transient, intermediate chromatin configuration in which a perinucleolar rim begins to form to a different extent, even when euchromatin in the GV is still decondensed and occupies most of the volume within the germinal vesicle (Supplementary Fig. 3). Because these oocytes have not completed the transition into an SN configuration and may thus have a lower developmental potential, our classification network was trained to designate this transient stage as NSN as these oocytes will still need to undergo significant large-scale chromatin remodeling to reach the SN configuration and acquire full meiotic and developmental potential. Notably, the predicted fluorescence maps produced from live NSN oocytes exhibited decondensed euchromatin, allowing for their accurate identification in the classification step (Fig. 4A).

To determine which features are the most important for the classification networks, we used the ablation class activation map method20,21, which shows regions in the fluorescent images that are critical for accurate classification. To locate regions in the input image that the model utilized for classification we used the Ablation-CAM/grad-cam method20,21. This strategy determines image features critical for accurate classification by stochastic removal of features and comparisons on the resulting effect on output accuracy20,21. The activation maps were extracted from all three of our classification models, and the average class activation map was drawn onto the predicted fluorescence map used for classification. Activation heat maps showing the regions of high importance for the classification of chromatin configuration in fixed and live oocytes are shown in Fig. 4B. The activation maps reveal that the classification networks focused on different nuclear regions in different oocytes to produce the classification score, predominantly focusing on both euchromatin (Euc) as well as prominent heterochromatin domains (Fig. 4B; arrowhead), which are most indicative of the chromatin configuration of each oocyte. Importantly, the most critical regions always included all or part of the chromatin ring or karyosphere on each oocyte, showing that the brightness of the chromatin ring around the nucleolus played an important role in the classification.

Effect of DNA minor groove-binding ligands on nucleolar structure and oocyte transcriptional profile

Exposure of live oocytes to cell-permeable, DNA-binding ligand stains, such as Hoechst 33342 and microinjection of nucleolar fluorescent markers, such as fibrillarin-GFP have been recently used to identify meiotically competent mouse oocytes that exhibit the SN chromatin configuration22,23. However, our results indicate that even after short-term exposure (25 min) of live oocytes to 0.5 μg/ml Hoechst 33342, the nucleolar structure is severely compromised (Fig. 5A and B). Signs of nucleolar disruption detected under bright-field imaging of live oocytes included a significant increase (p = 0.0463) in the proportion (56.6 ± 11.5) of oocytes that exhibit severe nucleolar shrinking and deformation as well as nucleolar fragmentation (43.3 ± 15.3; p = 0.0369) compared with controls (Fig. 5A and B). Notably, the corresponding fluorescent staining of chromatin structure revealed diffuse Hoechst staining inside the nucleolus and the formation of chromatin fragments (Fig. 5B; arrowheads). These results indicate that even a short exposure to a DNA-binding ligand (Hoechst 33342) can induce changes that are suggestive of nucleolar stress and fragmented chromatin formation in live oocytes.

Fig. 5: Effect of Hoechst treatment on nucleolar integrity and oocyte meiosis.
figure 5

A Percentage of oocytes with nucleolar abnormalities were detected by bright-field analysis in non-treated controls and after 25 min Hoechst treatment. Error bars represent the sample standard deviation (n = 3 replicates of 10 oocytes for control and treatment). B Representative bright-field images of the oocyte nucleolus (*) and their corresponding DAPI staining showing nucleolar shrinkage and fragmentation in the Hoechst treatment group. C Percentage of oocytes showing chromosome misalignment or spindle abnormalities after in vitro maturation (IVM) in controls and following Hoechst treatment for the first hour of maturation (1 h) or throughout IVM (17 h). Error bars represent sample standard deviation (n = 3 replicates of 10 oocytes for control and every treatment). D Representative fluorescence micrographs of control oocytes as well as oocytes with chromosome misalignment or spindle and chromosome abnormalities following Hoechst exposure. Chromosomes are shown in gray, spindle microtubules (acetylated α-tubulin) are labeled green and spindle pole aMTOCs (pericentrin) are depicted in red. Asterisks indicate statistically significant differences (P < 0.05).

Full size image

To determine whether exposure of live oocytes to a DNA-binding ligand affects chromosome segregation during in vitro maturation (IVM), we exposed oocytes to 0.5 μg/ml Hoechst 33342 for 1 h at the germinal vesicle stage or during the entire meiotic maturation period (17 h). Exposure to Hoechst 33342 at both time intervals induced substantial abnormalities in chromosome segregation patterns (Fig. 5C and D). Statistical analyses showed that the percentage of oocytes with chromosome misalignment defects was significantly higher in both Hoechst treatment groups (control vs. Hoechst 1 h (51.3 ± 31.3); p = 0.034, control vs. Hoechst 17 h (90 ± 10); p = 0.037). Moreover, exposure to Hoechst for 17 hours during meiotic maturation resulted in a significant increase (p = 0.046) in the proportion of oocytes (76.6 ± 11.5) that exhibited spindle structural abnormalities including unfocused and multipolar spindles, while chromosome structural abnormalities such as clustered chromosomes or scattered chromosomes also showed a significant increase (p = 0.046) after long term Hoechst exposure. These results indicate that even short-term exposure of live oocytes to Hoechst 33342 at the germinal vesicle stage may induce chromosome alignment defects and that long-term exposure during oocyte maturation exerts a highly detrimental effect on chromosome segregation and meiotic spindle stability.

To determine whether exposure to Hoechst 33342 or microinjection of GFP-labeled nucleolar markers affects the patterns of gene expression in mouse oocytes, we compared two publicly available transcriptome data sets22,23 that used a 15-min exposure to Hoechst 33342 or live cell imaging of Fibrillarin-GFP to select for NSN or SN oocytes22,23. We found significant differences in the transcriptional profile of NSN oocytes selected by these two approaches (Fig. 6A). For example, although 40% of transcripts exhibited similar levels of expression and read counts did not change between the Hoechst-treated group and the GFP-Fibrillarin microinjected group, 26% of transcripts exhibited a lower (absolute fold change ≥ 2) expression level and 34% of transcripts exhibited a higher (absolute fold change ≥ 2) expression level in the Hoechst treated group (Fig. 6A). Differences in the transcriptional profile of SN oocytes selected by these two methods were also detected with 47% of transcripts showing higher expression levels and 20% of transcripts exhibiting lower expression in the Hoechst treated group (Fig. 6B). Notably, the principal component analysis indicated that oocytes tend to cluster by method of selection rather than the type of chromatin configuration (Fig. 6C). Gene ontology (GO) analysis revealed that the top 15 enriched terms for genes that showed significantly different read counts in the NSN oocytes correspond to factors involved in translation factor activity and RNA binding, histone binding, mRNA binding and histone deacetylase binding (Fig. 6D). Moreover, the top 15 enriched GO terms for genes that show significantly different read counts in SN oocytes between the two approaches correspond to factors involved in ribonucleoprotein complex biogenesis and assembly, cell cycle, DNA replication, nuclear division, and ribosome biogenesis (Fig. 6E). Examples of the log2 transformed gene expression levels of transcripts that showed significantly different read counts are represented in a heat map with their corresponding GO terms (Fig. 6F). Genes related to nuclear functions (histone binding, nuclear division, DNA replication) and chromatin remodeling (deacetylase and kinase functions) show significantly different read counts. Importantly, we found significant changes in transcripts involved in ribosome metabolism, the chromatin reader LRWD1, and the Polycomb protein MBTD1 in Hoechst-treated oocytes as well as a significant 41.1-fold increase in transcript levels of the heat shock protein HSP90AA1 after Fibrillarin-GFP labeling (Fig. 6F). These results indicate that even short-term exposure to Hoechst 33342 may significantly affect the patterns of expression of critical chromatin remodeling and ribosome biogenesis factors at the GV stage. Importantly, the increased levels of expression of heat shock protein HSP90AA1 after live cell imaging may suggest a potential response to genotoxic stress following exposure to ultraviolet light during prolonged fluorescence microscopy. Thus, these methods of separation between NSN and SN oocytes have a significant effect on the transcriptional profile of oocytes that may limit subsequent downstream applications.

Fig. 6: In silico analysis and comparison of transcriptional profiles of mouse oocytes exposed to Hoechst 33342 or Fibrillarin-GFP imaging.
figure 6

A and B Comparison of publicly available transcriptome profiles22,23. The percentage of genes that showed no significant change or significantly increased, or decreased expression (absolute fold change ≥ 2.0) in Hoechst-treated NSN or SN oocytes. C Low-dimensional representation of sample clustering generated from principal component analysis of the read counts. D Top 15 enriched GO terms for genes, which showed significantly different read counts in the NSN oocytes between the two studies (unadjusted P-values were used). E Top 15 enriched GO terms for genes, which showed significantly different read counts in the SN oocytes between the two studies (unadjusted P-values were used). F Log2-transformed gene expression levels of genes, which showed significantly different read counts represented in a heat map with their corresponding GO terms.

Full size image

Transcriptome analysis following non-invasive selection of live oocytes with high developmental potential

Using our Fluo-Cast-Bright pipeline for the non-invasive selection of live NSN and SN oocytes, we analyzed and compared the transcriptome profile of these two types of oocytes using RNA sequencing (Fig. 7). Fully-grown oocytes were obtained from the large antral follicles of non-superovulated female mice on day 20 of post-natal development. Single, denuded oocytes at the germinal vesicle stage were transferred to individual wells containing 100 μl MEM medium supplemented with 10 μM Milrinone and imaged under bright-field confocal microscopy (Fig. 7A). The 3-D images of all collected oocytes were then processed immediately for nuclear segmentation and computational prediction of chromatin configuration. Live, GV-stage oocytes were then pooled based on the prediction of the Fluo-Cast-Bright pipeline into NSN (n = 15) and SN (n = 15) oocytes and processed for RNA sequencing in two replicates (Fig. 7A).

Fig. 7: Transcriptome profiles of live NSN and SN oocytes following non-invasive selection using Fluo-Cast-Bright.
figure 7

A Experimental design: Live GV stage oocytes were individually imaged by bright-field microscopy, and 3-D image stacks were subjected to fluorescence prediction followed by classification using the Fluo-Cast-Bright pipeline as either NSN or SN stage oocytes. Pools of live oocytes (n = 15) of each chromatin configuration were then subjected to RNAseq in two replicates. B Pearson correlation plot. Correlation (r) values between the read counts of transcripts in each sample. Color scale represents the range of the correlation coefficients (r) displayed. C Heatmap of differentially expressed transcripts. Color scale indicates log2 transformed gene expression levels revealing clustering of samples by chromatin state. D Low-dimensional representation of sample clustering generated from principal component analysis based on the expression levels of 48574 mouse transcripts with not-null expression. E Volcano plot indicating differentially expressed transcripts. Significantly up- (red) or down- (green) regulated transcripts in SN oocytes are shown (q value < 0.05, absolute fold change > 2).

Full size image

Analysis of quality control parameters from our RNA-seq libraries such as total clean reads and total mapping ratio revealed a high reproducibility between experimental replicates (Supplementary Fig. 4). The Pearson’s correlation coefficient for global gene expression levels between every two samples (Fig. 7B) indicates that overall, the transcriptomes of NSN oocytes and SN oocytes are similar. However, analysis of differentially expressed transcripts detected a significant difference in expression levels for a subset of 707 transcripts between NSN and SN oocytes (Fig. 7C). Importantly, dendrogram analysis of transcriptional profiles indicated that oocytes cluster by the state of chromatin configuration suggesting that significant differences can be detected in the transcriptional profile between NSN and SN oocytes (Fig. 7C). Principal component analysis (PCA) using the “compressed variance” of differentially expressed transcripts clearly separated oocytes by chromatin state based on PC1 (Fig. 7D). Notably, analysis of absolute loading values (Supplementary Fig. 5A) revealed that PC1 variance is accounted for by a cluster of highly abundant maternal transcripts, such as S-phase kinase-I (SKIP-1) a meiosis-specific subunit of the SCF (Skp1-Cullin1-F-box) complex required for meiotic competence24 as well as members of the maternal subcortical complex such as NLRP14, ZBED3, OOEP and PADI6, which are required for female fertility25,26,27,28. Gene ontology analysis revealed that these transcripts are involved in the regulation of protein metabolism and chromosome condensation (Supplementary Fig. 5B).

Comparison of differentially expressed transcripts between NSN and SN oocytes revealed that meiotically competent SN oocytes exhibit a significant (q-value < 0.05), up regulation (absolute fold change > 2) of 365 transcripts and a significant down regulation of 342 transcripts (Fig. 7E). Notably, the top-most up-regulated transcript in SN oocytes, Xpo1 (exportin 1) exhibited a log2-fold overexpression of 24.5 (q-value = 6.37E−5). XPO1 is a nuclear transport receptor, also known as CRM1, that is involved in protein and RNA export from the nucleus, required for the transport of hundreds of proteins in both somatic cells and oocytes, and directly involved in oocyte GV breakdown and meiotic resumption29. Additional transcripts with similarly high fold up-regulation in SN oocytes were Furin (log2-fold change 22.8, q-value = 3.12E−4), Kif23 (kinesin family member 23, also known as MKlp1; log2-fold change 22.8, q-value = 2.94E−4) and Tbl1xr1 (log2-fold change 23.1, q-value = 2.36E−04). Furin is a critical serine protease that cleaves immature pro-proteins into their corresponding active forms and regulates oocyte meiosis via Akt phosphorylation30. KIF23 is a microtubule kinesin involved in spindle assembly and remodeling31, while TBL1XR1 is a histone-binding component of the N-CoR and HDAC3 complex, which, exhibits chromatin-remodeling activity32. Other significantly up-regulated transcripts include the actin cytoskeleton regulatory factor33 CAP1 (cyclase-associated actin cytoskeleton regulatory protein 1; log2-fold change = 10.1, q-value = 2.03E−7), which is involved in spindle migration, cortical actin cap formation and cytokinesis34, and transcripts for Cnot6l (CCR4-NOT transcription complex subunit 6 like; log2-fold change = 9.88, q-value = 7.80E−6), which is involved in the degradation of maternal transcripts during oocyte maturation35,36. Notably, the transcript with the most highly significant q-value in SN oocytes was Gpr1 (G protein-coupled receptor 1; q-value = 2.39E−40), a gene that exhibits dynamic maternal imprinting in pre-implantation embryos37.

Amongst down-regulated transcripts in SN oocytes, we detected Nek1 (NIMA-related kinase 1), Bcat1 (branched-chain amino acid transaminase 1), Cnot3 (CCR4-NOT transcription complex subunit 3), and Gtse1 (G2 and S-phase expressed 1). NEK1 (log2-fold change −23.7, q-value = 1.36E−4) is a hub signaling kinase that modulates DNA repair capacity and is involved in sensing and repair of DNA double-strand breaks at the G2-M transition in somatic cells38. The metabolic enzyme BCAT1 exhibited a log2-fold reduction of −24.5 (q-value = 6.95E−5). Bcat1 redox function is critical for the regulation of protein oxidation during mitosis and to maintain the centromeric localization of Aurora Kinase B39, while CNOT3 (log2-fold change = −23.3, q-value = 1.90E−4) is involved in the regulation of cell cycle progression40. GTSE1, the transcript that exhibited the most significant q value (log2-fold change = −1.40, q-value = 2.22E−21), is expressed exclusively during the late G2/M transition where it is known to play a role in the maintenance of G2 arrest and as a regulator of microtubule dynamics during mitosis41,42.

Reactome43 enrichment analysis revealed that transcripts that are significantly up-regulated in SN oocytes are enriched for functional terms such as the formation of WDR5-containing histone-modifying complexes and epigenetic regulation of gene expression, the RHO GTPase, RAC1 GTPase, RHOA GTPase, and CDC42 GTPase cycles, which are essential for accurate completion of the metaphase-I to metaphase-II transition and cytokinesis44 as well as deadenylation-dependent mRNA decay (Fig. 8A). Whereas gene ontology (GO)45 molecular function (MF) analysis revealed enrichment for factors associated with nucleoside-triphosphatase regulator activity, GTPase and transcription co-regulator activity (Fig. 8B). Functional enrichment analysis of biological process (BP) revealed association with peptidyl-lysine modification and GTPase signaling, methylation and histone modification as well as biological processes related to chromosome segregation and chromatin remodeling (Fig. 8C). Cellular component (CC) term enrichment analysis also revealed significant enrichment for factors related to the actin cytoskeleton, ribonucleoprotein granules and microtubules as well as histone deacetylase complex and the histone methyl transferase complex (Fig. 8D). In contrast, we observed a down-regulation of factors associated with SUMOylation of DNA damage response and repair proteins in SN oocytes (Supplementary Fig. 6). These results provide critical evidence indicating that cellular processes such as maternal mRNA deadenylation, histone post-translational modifications, epigenetic regulation of chromatin remodeling and signaling pathways for the control microtubule dynamics during the metaphase-I to metaphase-II transition are important components of the SN oocyte’s program for the acquisition of meiotic potential.

Fig. 8: Gene ontology enrichment analyses of significantly up-regulated transcripts in SN oocytes.
figure 8

A Enrichment of biological pathways based on Reactome database. B Overrepresented GO-Terms molecular function (GO_MF), C (biological process, GO_BP), and D cellular component (GO_CC). Shading indicates the level of significance, and the size of the bubble correlates with the number of transcripts per category (unadjusted P-values were used).

Full size image

Discussion

Non-invasive selection of mammalian oocytes with high meiotic and developmental potential is a pressing need in the fields of human infertility as well as animal reproduction. However, until now, the identification of reliable cellular and or molecular markers of oocyte quality and developmental potential that are compatible with cell viability, as well as downstream research or clinical applications has remained elusive. Here, we developed Fluorescence prediction and Classification on Bright-field (Fluo-Cast-Bright), a novel pipeline for the non-invasive identification of chromatin structure and developmental potential in live mouse oocytes. Our convolutional neural network is capable of extracting information on the type of chromatin configuration present in fully-grown oocytes from non-invasive bright-field images obtained from live oocytes, generating a predicted fluorescent map of DAPI staining as well as H3K9me3 localization in the germinal vesicle. Our classification network was capable of distinguishing between oocytes that exhibit an NSN or SN configuration and thereby efficiently identified SN oocytes with high developmental potential. Fluo-Cast-Bright is thus capable to non-invasively and efficiently ‘cast’ a fluorescence prediction of large-scale chromatin structure on a bright-field image of unlabeled oocytes and achieved 91.3% accuracy in the classification of chromatin state in fixed oocytes and 85.7% accuracy in live oocytes. Importantly, our approach does not require any special imaging conditions or prolonged live cell imaging that may compromise cell viability, making it widely applicable in any conventional clinical setting or research laboratory.

As a proof of concept, we conducted the first transcriptome analysis following the non-invasive selection of live NSN and SN oocytes. Our results indicate that meiotically competent SN oocytes exhibit a higher expression of transcripts encoding for critical factors such as Exportin 1 (Xpo1), which is required for nuclear export of mRNA as well as hundreds of proteins during the G2/M transition, in addition to factors required for degradation of maternal mRNA stores, histone methylation as well as protein enzymatic activation. Importantly, our results indicate that many of these factors are required for the regulation of nuclear architecture, chromosome stability, the metaphase-I to metaphase-II transition, and the control of microtubule dynamics in preparation for the completion of meiosis. Our results reveal critical molecular markers for oocyte quality and provide additional evidence to dissect the molecular mechanisms regulating the oocyte’s program for nuclear and cytoplasmic maturation during the acquisition of meiotic and developmental potential. Importantly, our pipeline provides a fast and non-invasive selection of meiotically competent oocytes for downstream research and clinical applications.

Convolutional neural networks are powerful tools to derive quantitative subcellular information from bright-field or differential interference contrast microscopy images from somatic cells and tissues12,13. Ours is the first pipeline specifically designed for the non-invasive analysis of nuclear architecture in live mammalian oocytes. A large body of evidence indicates that mouse SN oocytes exhibit higher meiotic and developmental potential1,6,9,10,11. For example, SN oocytes exhibit higher rates of meiotic maturation in vitro and development to the blastocyst stage following in vitro fertilization1,6,9,10,11. Thus, the use of our Fluo-Cast-Bright pipeline for the non-invasive identification of SN oocytes also allows for the efficient selection and sorting of live SN oocytes with high developmental potential. Our pipeline is capable of generating a prediction of chromatin state from a single-time-point, bright-field 3-D image of a live mouse oocyte taken within an average of 10 min under a conventional confocal microscope plus an additional 5 min for nuclear segmentation and computational image analysis. It achieved high accuracy NSN/SN classification performance on both fixed and live cells and generated fluorescence prediction maps that were remarkably similar to the corresponding ground truth fluorescent maps observed in fixed oocytes. Importantly, this prediction accuracy was obtained using 138 fixed and 105 live oocytes for training and validation, with tests demonstrating a robust and high performance of our model in all metrics evaluated here (accuracy, precision, recall, and F1 score). Our current studies are aimed at establishing whether these parameters can be further improved with the use of larger live cell imaging training data sets. We have recently demonstrated that the chromocenters within the SN oocyte nucleus exhibit high mobility in live oocytes8. This is consistent with the reduced definition of large pericentric heterochromatin domains (chromocenters) in the predicted fluorescent maps obtained following live oocyte imaging. However, re-training the classification networks using live oocytes was sufficient to recover the performance of our network and generate predicted fluorescent maps based on both readily apparent as well as hidden information on bright-field images that faithfully reflected the position of chromocenters in live oocytes. Thus, retraining the classification model alone was sufficient to recover the prediction accuracy and enabled us to determine the chromatin configuration of each live oocyte label-free. It is also possible that different laboratories may have different microscopes and/or objectives with different numerical apertures that yield bright field images with a different level of resolution. In which case, high-accuracy predictions will require a simple retraining step of the classification models to adjust for their image resolution levels.

Importantly, our pipeline provides a reliable indication of the chromatin state of live oocytes without affecting cellular viability, so that researchers or clinicians can make informed decisions while selecting oocytes for experiments or clinical applications. Fluo-Cast-Bright is also fast in generating a final chromatin prediction after nuclear segmentation, as our results indicate that batch prediction alone for 3-D image stacks of 42 oocytes takes only about 1.5 min on an Nvidia Tesla P100 graphics processing unit.

Several deep learning tools have been recently generated for the analysis of male and female gametes, albeit with different objectives. For example, a deep convolutional neural network was developed and trained using bright-field images of human spermatozoa with known DNA integrity to predict DNA quality in test samples, achieving 86th percentile in the prediction of DNA fragmentation index on sperm cells15. However, this network was developed for the analysis of fixed spermatozoa, and the algorithm will likely need drastic modifications and optimization procedures to work well on live cells as spermatozoa exhibit high motility and vigorous multidirectional movement. In another study, time-lapse bright field microscopy was used in combination with particle image velocimetry and a feed-forward artificial neural network for quantitative analysis of cytoplasmic movement velocity as a predictor of developmental competence in mouse oocytes16. However, this method requires a specialized cell-screening system, and the oocytes need to be imaged for 15 h, which limits the applicability of the protocol. More recently, oocyte phenotyping based on analysis of the texture of the zona pellucida and cytoplasmic particle size has shown some predictive value on oocyte maturation potential and to distinguish knockout oocytes with different genetic backgrounds46. Consistent with previous reports47 this approach confirmed that small oocytes with a thin zona pellucida fail to undergo germinal vesicle breakdown and remain arrested at the GV stage. However, this approach does not distinguish between NSN oocytes, which are capable of meiotic maturation but have low developmental potential, and SN oocytes, which exhibit a higher potential for development to the blastocyst stage.

Both, DNA-binding ligands such as Hoechst 33342 or microinjection of a GFP-Fibrillarin marker have been recently used to separate mouse NSN and SN oocytes for experimental analysis22,23. However, a direct comparison of transcriptional profiles in NSN and SN oocytes obtained using these two approaches suggests the presence of cytotoxic effects and DNA damage on these oocytes. Our results also indicate that even a brief exposure (25 min) to 0.5 μg/ml Hoechst 33342 induces striking changes in the nucleolar structure and chromatin configuration of oocytes. The nucleolus is a critical nuclear compartment for ribosome biogenesis, its morphology is known to correlate with the levels of rRNA transcription, cell growth and metabolic rates48. It also acts as a key hub in the cellular stress response. Thus, it is conceivable that the abnormalities we observe here originated from a stress response to DNA damage, which involves the activation of signaling pathways, that affect cell-cycle progression and trigger cellular senescence48. Consistent with this notion, oocytes microinjected with GFP-Fibrillarin23 exhibit high levels of γ-H2AX phosphorylation, a bona fide marker for the presence of DNA damage and double-strand DNA breaks. The nucleolus is also an important platform for the organization of chromatin structure and regulation of genome integrity, thus the chromatin structural abnormalities we detected here after Hoechst exposure may have originated not only from DNA damage but also due to a disruption of the balance between ribosome biogenesis and chromatin organization49,50. Importantly, in silico analysis and principal component analysis also revealed that this short exposure to Hoechst resulted in dramatic changes in gene expression involved in ribosome biogenesis consistent with the changes in nucleolar structure and chromosome instability. Many genes related to meiotic progression were also found to be differentially expressed between oocytes labeled by Hoechst22 and oocytes labeled by fibrillarin-GFP23. Moreover, we found changes in the levels of expression for transcripts related to nuclear functions such as histone binding, cell division, and chromatin remodeling potentially accounting for the significant increase in chromosome and spindle abnormalities observed in Hoechst treated oocytes in this study.

To demonstrate downstream applications of our pipeline, we conducted a transcriptome analysis following a non-invasive selection of NSN and SN oocytes. This is the first time that an NSN and SN transcriptomics dataset has been generated without using any fluorescent labels and is, therefore a, faithful representation of transcriptomic changes in live oocytes during the NSN to SN transition. RNA-sequencing analysis revealed important transcriptional differences between these two types of oocytes. Both gene ontology (GO)45 and Reactome43 term analysis pointed to pathways related to chromatin, epigenetics, cell cycle regulation, and nuclear functions.

Notably, the most up-regulated transcript in SN oocytes encodes for the nuclear export receptor exportin 1 (Xpo1). XPO1 is also known as CRM1 (chromosome region maintenance 1) and plays a critical role in the transport of myriad proteins51. Proteome analysis revealed close to 1000 proteins are CRM1 cargo substrates in Xenopus oocytes, and their nuclear export is essential for critical cellular processes as diverse as centrosome function, cytoskeleton dynamics, ribosome maturation, and mRNA degradation52. Notably, chemical inhibition of XPO1 protein delays germinal vesicle breakdown, while over-expression accelerates this process in pig oocytes29. Thus, the accumulation of higher Xpo1 transcript and protein levels may be a critical step in preparation for meiosis onset and the acquisition of meiotic competence. This is consistent with previous studies, which revealed that NSN oocytes had a lower GVBD rate compared to SN oocytes23. Thus, our study provides evidence indicating that Xpo1 transcript levels may be a critical molecular marker for oocyte quality and developmental potential.

Furin is a pro-protein convertase that is required for the proteolytic cleavage of numerous proteins into their active isoforms. Conditional deletion during mouse oocyte growth induces ovarian follicle arrest at the late pre-antral stage and results in subsequent infertility53. At the germinal vesicle stage, Furin is localized to the plasma membrane30, where it subsequently activates the Insulin-like growth factor 1 (IGF-1) receptor to maintain microvillus organization and activate protein kinase B (PKB/Akt) to promote the resumption of meiosis30,54,55. Our results indicate that SN oocytes also accumulate higher levels of transcripts required for spindle and actin filament regulation in preparation for meiotic progression, including Kif23 and Cap1. KIF23/MKLP1 is recruited to the spindle mid zone by the inner centromere protein (INCENP) which is critical for mid-body formation56. In Drosophila oocytes, KIF23 exhibits a microtubule cross-linking kinesin activity that is essential for the regulation of meiotic spindle length and maintenance of microtubule stability during the metaphase-II stage31. During mouse oocyte meiosis, CAP1 co-localizes with the plasma membrane actin cap, where it is required for actin disassembly, spindle migration, asymmetric division, and polar body extrusion34,57. Actin filament reorganization is essential for accurate chromosome segregation during meiosis I58, thus accumulation of higher Cap1 transcripts in SN oocytes may also contribute to the accurate segregation of chromosomes in oocytes of high developmental potential.

TBL1XR1 is an F-box-like/WD40 repeat-containing protein, and a member of the N-Cor complex32, known to functionally interact with Sycp3-like Y-linked (SLY), a multi-copy gene required for post-meiotic sperm chromatin remodeling59. Notably, Tbl1xr1 is highly expressed as a maternal transcript in zebrafish embryos60. Maternal transcripts synthesized during oocyte growth undergo complex temporal regulation during meiotic maturation61. Selective degradation of maternal mRNA transcripts is essential for the acquisition of both meiotic and developmental competence. CNOT6L is a subunit of the CCR4–NOT complex, the eukaryotic mRNA deadenylase, that is preferentially expressed in mouse oocytes, its functional ablation interferes with mRNA degradation and results in meiotic spindle abnormalities, meiotic arrest, and severe subfertility36. Interestingly, the function of CNOT6L in maternal mRNA degradation is mediated through the RNA-binding protein ZFP36l2, which was recently shown to regulate global transcriptional silencing and chromatin modifications, including increased histone methylation (H3K9me3) associated with the transition into the SN configuration in mouse oocytes, providing a functional link between mRNA decay and global transcriptional quiescence during the acquisition of meiotic competence62. Thus, SN oocytes accumulate higher levels of key transcripts that are required not only for maternal mRNA degradation as well as large-scale chromatin remodeling but also higher levels of transcripts that are required for critical processes such as spindle assembly, migration, and remodeling to ensure accurate chromosome segregation during the metaphase-I to metaphase-II transition.

In contrast, we found that SN oocytes exhibit lower levels of transcripts for several critical factors associated with the response to DNA damage/repair compared with NSN oocytes. For example, NEK1 is critical for the detection of DNA double-strand breaks and its functional ablation in mitotic cells affects mitochondrial function, including a reduction in complex-I activity38. Notably, complex-I activity is known to be suppressed in a developmentally regulated manner in human and Xenopus early growing oocytes to prevent the formation of reactive oxygen species of mitochondrial origin63. BCAT1 is an essential metabolic enzyme that regulates branched-chain amino acid catabolism and is also required for the regulation of redox pathways39. Bcat1 transcripts are similarly regulated and exhibit a conserved function during the process of physiological aging in C. elegans, zebra fish and mouse somatic cells64. Importantly, its induced overexpression has been recently associated with enhanced oocyte quality and extension of reproductive lifespan in the C. elegans model65. Our results provide evidence that critical factors required for the repair of DNA damage and cellular aging are downregulated at the transcriptional level during the transition into the SN configuration, potentially reflecting a developmental switch in the DNA damage response and/or branched-chain amino acid catabolism associated with the increased chromatin condensation present in SN oocytes or perhaps as a result of the existence of specialized pathways to resolve DNA breaks during the G2/M transition in pre-ovulatory oocytes. Consistent with this notion, our recent studies revealed a striking difference in the chromatin response of NSN and SN oocytes to DNA double-strand breaks induced by acute γ-irradiation8. Our studies also provide molecular insight into the factors and potential mechanisms associated with the oocyte’s program for the acquisition of meiotic and developmental potential. Further studies will be required for the functional analysis of specific molecular markers associated with enhanced oocyte quality identified in this work.

In conclusion, our results provide evidence that deep learning networks can extract hidden information from bright-field images of germinal vesicle stage oocytes that normally only become visible with invasive labeling methods. Fluo-Cast Bright has potential for both research and clinical applications in human and animal reproduction. Similar strategies can be applied to other cell types, to facilitate a non-invasive analysis of cellular potency. For example, induced differentiation of cardiomyocyte or neuronal precursor cells from iPSC cells must retain high levels of DNA integrity in order to be considered for any potential therapeutic applications66. Thus, in silico prediction of DNA integrity and other cell quality attributes will be essential to expand the clinical applications of iPSC-derived cells with therapeutic potential.

Methods

Animals

All experiments were conducted in accordance with guidelines and following approval of protocols by the ‘Institutional Animal Care and Use Committee’ (IACUC) of the University of Georgia. We have complied with all relevant ethical regulations for animal use. Female mice of the species Mus musculus and a hybrid genetic background of C57BL/6/DBA were housed at ambient temperature (24–26 °C) with a 12 h light/dark light cycle and food and water provided ad libitum and used for oocyte collection between postnatal days 18 and 22.

Oocyte collection and culture

GV stage oocytes were collected from non-primed female mice of a C57/Bl6/DBA2J F1 hybrid background between postnatal days 20–24 by follicular puncture. Cumulus oocyte complexes were then transferred to minimal essential medium (MEM), supplemented with 3 mg/ml bovine serum albumin (Sigma Aldrich, St. Louis, MO) and milrinone (Sigma; 1 μg/ml) to prevent meiotic resumption. Surrounding cumulus cells were removed by gentle pipetting and denuded oocytes were maintained at 37 °C in MEM/BSA plus milrinone under 5% CO2, 5% O2, and balanced N2 until subjected to the procedures outlined below.

Imaging of fixed oocytes following immunofluorescence staining

The denuded oocytes were fixed in 4% paraformaldehyde in phosphate buffered saline (PBS) solution, supplemented with 0.05% Triton X-100 for 30 min at 37 °C and subsequently blocked in buffer containing 5% fetal bovine serum, antibiotic–antimycotic and 0.01% Triton X-100 in PBS at 4 °C. To reveal the state of chromatin in the nucleus, oocytes were immunolabeled with specific polyclonal antibodies directed against histone H3 trimethylated at lysine 9 (H3K9me3, 1:200, Cat#ab5819, Abcam, Cambridge, MA, USA) in blocking buffer overnight at 4 °C. Following repeated washes in blocking buffer, the cells were then incubated with a specific anti-rabbit Alexa Fluor 555-conjugated secondary antibody (1:1000, Cat#A21430, Life Technologies, Eugene, OR) for 1 h at 37 °C. After a series of final wash steps in blocking buffer, the oocytes in solution were exposed to DAPI (4’,6-diamidino-2-phenylindole) to counterstain the DNA and transferred to individual wells of optical bottom 96-well plates (Greiner Bio-One, Frickenhausen, Germany). Imaging was performed on a Nikon Eclipse Ti-U/D-Eclipse C1 laser-scanning confocal microscope equipped with a ×40 objective lens following sequential (frame lambda) excitation with 401 and 561 nm Coherent Sapphire lasers. Image acquisition was conducted using EZC1 3.91 software (Nikon) with a step size of 0.8 μm and a Z-stack range of ~12.8 μm to acquire 3-D image stacks of each oocyte with a bright-field channel, in addition to the fluorescence channels. The image stacks were taken at a resolution of 16*1024*1024 px3, with a pixel size of 800*148*148 nm3 in the final 3D image stacks. The H3K9me3 and DAPI fluorescence signals were merged into a single chromatin fluorescence map. The cuboid volume in the image stack containing the nucleus was then segmented with the aid of labeling software labelme67 and resized to 16*128*128 px3. The bright-field image and fluorescence map in these segmented image stacks were used as the training set. Each oocyte image stack was labeled as SN or NSN based on the combined fluorescence map of DAPI and H3K9me3.

Imaging of live mouse oocytes

For live oocyte imaging, denuded oocytes were placed one by one in 180 µl micro-drops of MEM/BSA/milrinone on glass-bottom Petri dishes (Mattek, Ashland, MA) overlayed with mineral oil. The glass bottom dish was placed in an environmental chamber with an atmosphere of 5% CO2, 5% O2, and 90% N2 at 37 °C. Imaging and segmentation were conducted exactly as described above. After imaging, the oocytes were then fixed and stained with DAPI to determine their true chromatin configuration.

Hoechst treatment and in vitro maturation of GV oocytes

To observe the effect of DNA dyes on live GV oocytes, oocyte culture media was supplemented with 0.5 μg/ml of the minor groove DNA binding ligand Hoechst 33342 (Life Technologies, Eugene, OR) and oocytes were exposed for 20–25 min before imaging. The experiment was conducted in three biological replicates and compared to equal numbers of non-treated control oocytes (10 oocytes per group and replicate). Imaging and segmentation were conducted exactly as described above using a confocal microscope to obtain image stacks for fluorescence prediction as well as the detection of abnormalities caused by Hoechst treatment.

Control oocytes, as well as oocytes following short-term exposure (1 h) or continuous exposure (17 h) to 0.5 μg/ml Hoechst 33342 were subjected to in vitro maturation (IVM) for 17 h in DMEM supplemented with 5% FBS. For each control and treatment group, 3 replicates of 10 oocytes were used.

In vitro matured oocytes were then fixed, blocked, and immunostained as described above using a primary antibody cocktail consisting of a mouse anti-acetylated α-tubulin antibody (1:1000, T6793, Sigma Aldrich) and a rabbit anti-pericentrin antibody (1:1000, PRB432C, Covance, Princeton, NJ) in blocking buffer for 1 h at 37 °C. AlexaFluor-conjugated secondary antibodies included AF555-anti-rabbit and AF488-anti-mouse antibodies (1:1000, Life Technologies, Eugene, OR) After immunofluorescence labeling, the oocytes were mounted on glass slides using Vectashield mounting medium containing DAPI (VectaShield, Vector Laboratories, Burlingame, CA). The slides were then analyzed and imaged using a Leica DMRE epifluorescence microscope. Only oocytes with spindle orientations allowing unequivocal analysis were included in the data collection.

Comparison between two published oocyte transcriptome datasets

The transcriptome data analysis was carried out using the publicly available data sets comparing the transcriptomes of NSN and SN mouse oocytes, including Ma et al.22 and Wang et al.23. The average read counts of 757 genes common in both published datasets were used to carry out our analyses of the effect of Hoechst 33342 and Fibrillarin-GFP on transcriptome profiles, and genes with absolute fold change >2 were considered genes with a significantly different expression.

First, we used the prcomp method in R68 to perform a principal component analysis on the read counts of 757 genes to determine the origin of variance in the two datasets (difference between SN vs. NSN or Hoechst vs. fibrillarin-GFP). Then, we compared the expression values of each gene in SN and NSN oocytes between the two studies. Read counts of the two datasets showing an absolute fold change ≥ 2 were considered differentially expressed in response to Hoechst exposure in SN as well as NSN oocytes. We performed functional annotation for each group by enrichment analysis based on the GO database45.

Fluorescence prediction and Classification on Bri g ht-field (Fluo-Cast-Bright) pipeline

Our Fluo-Cast-Bright pipeline involves a fluorescence prediction model followed by a classification step and is a software developed based on the pytorch framework69. The fluorescence prediction model is a CNN network with a 3D U-Net++18 architecture (Supplementary Fig. 1A). To train for fluorescence prediction, we applied modifications inspired by the pytorch-fnet of the Allen Institute for Cell Science13, such as training on batches of randomly sampled 3D patches of images. The network was also trained with a loss function, which guides the network to produce a fluorescence map with minimal mean-squared error to the actual combined fluorescence map of DAPI and H3K9me3.

The training took place in the typical forward–backward fashion of CNNs for 300 epochs, with 50 steps in each epoch. The images fed into the network were 3-D image patches with a size of 16 × 64 × 64 px3 (ZXY) randomly sampled from the training set and stored in a patch buffer. The patch buffer started with 16 randomly sampled patches and replaced one of its patches with a newly sampled patch at an interval of about 31 steps. Training was performed with Adam optimizer and a learning rate of 0.001. The deep supervision function of the U-Net++ architecture was not applied.

The classification network has the ResNeXt-10119 architecture (Supplementary Fig. 1B) and was trained to distinguish between SN and NSN oocytes using 2-D slices of 3D predicted fluorescence maps produced by the fluorescence prediction model. We used the predicted fluorescence map generated from validation images of the fluorescence prediction network to train the classification models, because of the resemblance of these predicted fluorescence maps to the fluorescence prediction network output from new oocyte images.

For training, the 3-D predicted fluorescence maps were split into individual 2-D fluorescence maps, these 2-D fluorescence maps were shuffled and fed into the network in random batches with 32 images in each batch. Training was performed with a learning rate of 0.0005 on the Adam optimizer to maximize the accuracy of the NSN/SN classification. The correct NSN/SN label of each 2-D fluorescence map was determined by the true NSN/SN chromatin state of each oocyte. The training was performed using three different train/valid splits to produce three different classification models. Training of a fluorescent prediction model takes ~65 min on the NVIDIA A100 Tensor Core graphic processing unit (300 epochs), while training of a classification model takes ~4 min on the NVIDIA A100 Tensor Core graphic processing unit (120 epochs).

Training of the fluorescence prediction model was based on bright-field images of 96 3D fixed oocyte images as a training set and 42 3D fixed oocyte images as a validation set and was subsequently used for the fluorescence prediction of both fixed and live oocytes. In contrast, the training of the classification model for fixed oocytes was based on the 2D sections of 42 3D predicted fluorescence maps of fixed oocyte images, while training of the classification model for live oocytes was based on the 2D sections of 63 3D predicted fluorescence maps of live oocyte images. This difference in training between fixed and live oocyte images was deemed necessary to avoid decreased classification accuracy due to movements of subcellular structures and organelles in live but not in fixed oocytes, and the different classification models were subsequently applied according to oocytes that had not participated in training or validation to determine the chromatin configuration.

For the classification of the NSN or SN configuration using only bright-field images of oocytes that never participated in training or validation, a 3-D predicted fluorescence map was first produced with the fluorescence prediction network. Then the three classification networks scanned through each 2-D cross-section of the 3-D predicted fluorescence map to determine whether the fluorescence map reveals NSN or SN configuration. For each 2-D fluorescence map, each classification network provides a percentage score for NSN and SN. The average of the NSN and SN percentage scores predicted by the three models was then used to determine the label of the 2-D fluorescence map. The new live oocyte is then designated as NSN and SN based on the majority prediction of 2-D fluorescence maps.

We selected our model architectures based on the accuracy to predict chromatin configuration. The different parameters and optimal performance of our selected model architectures were examined by training and comparing eight commonly used architectures, as well as after conducting a parameter sweep (Supplementary Fig. 2A–D). Three tests were carried out on each model architecture and each set of parameters. In each test, training was carried out using the same training and validation set images, and the best validation performance achieved by the model was recorded. We also tested the classification accuracy of eight different classification models by carrying out a prediction on the same set of new live oocyte images (Supplementary Fig. 2B).

The publicly accessible ‘grad-cam’ tool20 was used to carry out the ablation class activation map method21 on our classification models to interpret how our pipeline reached its classification from predicted fluorescence maps. This method determines image features critical for accurate classification by stochastic removal of features and the resulting effect on output accuracy21 (Fig. 4B).

RNA-Seq on NSN and SN oocytes separated by the Fluo-Cast-Bright pipeline

We used our Fluo-Cast-Bright pipeline for the non-invasive selection of live oocytes with different chromatin state in two replicates. Oocytes were classified as either NSN or SN according to our pipeline. Oocytes with the NSN chromatin configuration were selected and pooled (n = 15), while the same number of oocytes with the SN configuration (n = 15) were selected and pooled. Oocytes of each chromatin state were processed for sample preparation, including removal of the zona pellucida using Tyrode’s solution and lysis for RNA-Seq70.

Transcriptome analysis was carried out on a DNA nanoball sequencing platform, also known as the DNBSEQTM platform. The genome reference we used for mapping is Genome Assembly GRCm38 (mm10). The data size generated from sequencing was about ~6 gb per sample and had an average genome mapping ratio of ~94.3% using HISAT2 (Version v2.0.4)71. The genome mapping ratios were very uniform, suggesting that the samples are comparable. Novel coding transcripts obtained from sequencing were first merged with reference transcripts to obtain the complete reference. Bowtie272 (Version v2.2.5) was then used to map clean reads to this reference to calculate the gene expression level of each sample and achieved an average mapping ratio of ~82.9%.

Principal component analysis on the read counts of all 48,574 mouse transcripts with not-null expression was carried out using the PCA tool of Scikit-learn73. The feature importance of each transcript in PC1 was calculated as loadings, which represent the degree of association of features with the PC axis74. Loadings are calculated by the eigenvector coefficient of the feature in the PC multiplied by the square root of the eigenvalue (amount of variance summarized by PC)74. Transcripts with absolute loading greater than 1 were selected as transcripts of interest (Supplementary Fig. 5A) and GO term analysis was carried out on these transcripts (Supplementary Fig. 5B).

Differential expression between NSN and SN samples was analyzed with DESeq275, and transcripts with an absolute fold change > 2 and q-value < 0.05 were used to identify differentially expressed genes (DEGs). Functional analyses were performed for these DEGs by enrichment analysis based on the GO database45 and Reactome database43.

Statistics and reproducibility

RNA-seq was conducted as a duplicate experiment, while the effect of Hoechst on oocyte morphology observations was carried out on three independent biological replicates. To ensure the stringency of comparison, the Pairwise Mann–Whitney U test (Wilcoxon signed-rank test) in R68 was used to assess statistical significance. Data are presented as means, and variation between individual replicates is indicated as the standard deviations. The asterisk in bar plots represents a significance of p-value < 0.05. Error bars on bar plots represent the sample standard deviation. GO term and Reactome enrichment analyses were carried out using the clusterProfiler76 package in R.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.