Integrative dual-track transcriptomics reveals stage-specific coordination, regulatory divergence, and HSP90AA1-associated remodeling in human folliculogenesis

Abstract

Human folliculogenesis depends on coordinated yet non-identical developmental remodeling in the oocyte and its surrounding granulosa cells. When these two compartments remain synchronized and when they diverge into lineage-specific regulatory states, however, remains incompletely resolved. Here we performed an integrative dual-track re-analysis of the human RNA-seq dataset GSE107746, modeling oocytes and granulosa cells as distinct but developmentally linked compartments across follicular progression. Analysis of 148 sequencing libraries showed that compartment identity was the dominant source of transcriptomic variation, supporting compartment-aware downstream interpretation. Within this framework, oocytes followed a relatively continuous developmental trajectory, with substantial transcriptional remodeling already evident across adjacent stages, whereas granulosa cells showed weaker early-stage contrasts but markedly stronger late-stage reorganization, particularly around the antral and preovulatory transitions. Functional enrichment indicated that oocyte maturation was associated with RNA-processing and broader genome-regulatory remodeling, whereas granulosa maturation was dominated by progressive mitochondrial and bioenergetic activation. Co-expression analysis showed that both compartments contained strong late-stage programmes together with inverse early-state modules, indicating a shared systems-level architecture of maturation, although the hub-gene composition and biological content of these programmes were largely compartment-specific. Machine-learning validation reinforced this asymmetry: oocyte stage classification was best recovered from a compact eigengene-based representation, whereas granulosa stage discrimination was better resolved by a broader differential-expression-derived feature set. At the gene level, HSP90AA1 emerged as a stage-associated marker with compartment-specific behavior, showing progressive attenuation across oocyte development, assignment to the selected oocyte blue module, and sharper transitional dynamics in granulosa cells. Together, these findings support a model in which human folliculogenesis proceeds through coordinated but non-equivalent transcriptomic remodeling, with shared developmental logic at the systems level but distinct molecular execution in germline and somatic compartments.

Graphical abstract

1. Introduction

Human folliculogenesis is a prolonged developmental process in which the oocyte and its surrounding granulosa cells mature through continuous bidirectional communication, reciprocal metabolic support, and tightly regulated signaling exchange rather than through isolated cell-autonomous programs.6, 9 The functional outcome of this partnership extends beyond oocyte growth itself, because follicle survival, endocrine competence, and acquisition of developmental potential depend on the integrity of the surrounding somatic microenvironment as well as the germline compartment.5, 9 Recent single-cell and transcriptomic studies of the human ovary have reinforced this view by showing that follicular development is accompanied by substantial cellular heterogeneity, lineage diversification, and context-dependent remodeling across granulosa, theca, stromal, and germ-cell populations.3, 7, 8 Together, these observations place folliculogenesis within a systems framework in which developmental competence emerges from coordinated multicellular regulation rather than from oocyte-intrinsic maturation alone.

This broader biological perspective also sharpens the central unresolved question for ovarian transcriptomics. Stage-associated gene catalogues are informative, but they do not by themselves explain how oocyte and granulosa-cell programs remain aligned during some phases of follicle growth while becoming increasingly specialized during others.5, 7 The biology of early folliculogenesis suggests a period of tighter inter-compartment coordination, because granulosa cells support oocyte survival, nutrient exchange, and signaling competence during the earliest growth transitions. By contrast, later development is more likely to involve progressive functional specialization as meiotic competence, steroidogenic activity, extracellular communication, and somatic support functions become redistributed across the follicle.2, 3, 17 Recent studies of ovarian cortical follicles and broader follicular cell-state landscapes have further shown that developmentally adjacent follicles can still harbor meaningful transcriptional heterogeneity, supporting the idea that coordination across compartments is dynamic rather than uniform.⁷ These observations raise a mechanistic question that has not yet been fully resolved: at which developmental transitions do oocyte and granulosa-cell programs remain synchronized, and at which points do they begin to diverge in a way that reflects a regulatory handoff between germline and somatic compartments?

Several studies have already reused paired folliculogenesis transcriptomic resources, but most have pursued narrower analytical objectives than this coordination-focused question. Wang and colleagues reanalyzed human folliculogenesis transcriptomes through a cell-specific network framework and highlighted marked network rewiring in antral-stage oocytes, shifting emphasis toward hub structure and subnetwork dynamics.¹⁰ Liu and colleagues used the same developmental setting to prioritize biomarkers of ovarian reserve, particularly in primordial follicles, thereby focusing on candidate marker genes rather than on cross-compartment developmental synchronization.⁴ Zhang and colleagues then applied a lncRNA-centered sub-pathway strategy to primordial follicle activation and concluded that lncRNA-associated signals display stronger cell-type specificity than protein-coding genes in early follicular development.¹² More recently, Zhao and colleagues restricted their analysis to granulosa-cell lncRNA dynamics and linked SNHG18 to ovarian aging, glycolytic disruption, and apoptosis (X.⁴¹). Disease-oriented overlays have also been prominent, including integrated bioinformatics studies of polycystic ovary syndrome that mapped pathological granulosa-cell findings onto normal follicular developmental expression patterns.¹¹ The broader interpretive landscape has additionally intersected with pathway-focused work on human ovarian steroidogenesis, granulosa-cell glycolysis, chromatin remodeling, and ROCK1-mediated follicle maintenance, reinforcing how often follicular transcriptomes are mined for pathway validation or gene-centered inference.1, 13, 14, 17 Collectively, the literature shows that stage-resolved folliculogenesis datasets have already been highly productive for biomarker discovery, lncRNA prioritization, disease overlays, and compartment-specific hub or pathway analysis, but they have been used less often to ask when developmental regulation is shared versus compartment-specific across the oocyte-granulosa axis.

In this context, the paired human RNA-seq study of Zhang and colleagues provides an especially powerful framework because it profiled oocytes and their corresponding granulosa cells across five follicular stages and showed that the two compartments display distinct yet developmentally linked transcriptional programs, including differences in transcriptional machinery, transcription factor networks, and predicted intercellular signaling interactions.¹⁵ The associated dataset, GSE107746, therefore offers more than a stage-wise expression atlas, because it preserves developmental order and compartmental identity in the same experimental framework. Our core research question is how oocyte and granulosa-cell transcriptomic programs are coordinated and progressively decoupled across human folliculogenesis, and which developmental transitions mark the principal regulatory handoff points between germline and somatic compartments. The objective of this study is to perform an integrative dual-track transcriptomic analysis of oocytes and granulosa cells across human folliculogenesis to identify stage-specific transcriptional reprogramming, shared versus compartment-specific biological processes, and candidate regulatory modules associated with developmental coordination or divergence. We hypothesized that human folliculogenesis is characterized not by uniform parallel maturation of oocytes and granulosa cells, but by stage-specific shifts in inter-compartment coordination, such that key developmental transitions are marked by regulatory divergence and compartment-dominant molecular programs. Unlike previous studies that primarily mined GSE107746 for biomarkers, lncRNA signatures, disease overlays, or compartment-specific hub genes (4, 10, 11, 12; X.⁴¹), the present study frames human folliculogenesis as a stage-resolved two-compartment coordination system and investigates when developmental regulation is shared versus compartment-specific.

2. Materials and methods

2.1. Study design and dual-track analytical framework

The study was designed as a dual-track, stage-resolved transcriptomic analysis of human folliculogenesis in which oocytes and granulosa cells were modeled as distinct but developmentally interacting compartments. To preserve this biological structure, all major analytical steps were organized in a compartment-aware manner, allowing each track to be processed independently before cross-compartment integration. This design enabled both within-compartment characterization of developmental transitions and explicit comparison of shared versus compartment-specific transcriptional programs.

The framework was motivated by the hypothesis that folliculogenesis involves stage-specific shifts in germline-somatic coordination rather than uniform parallel maturation. Accordingly, the analytical workflow was structured to identify the developmental intervals at which oocyte and granulosa-cell transcriptomes appear most coordinated, the transitions at which molecular divergence emerges, and the regulatory modules associated with these handoff points.

To implement this design, compartment-specific pipelines were used for transcript quantification, filtering, normalization, exploratory structure analysis, differential expression, enrichment analysis, and co-expression network construction. These compartment-resolved outputs were then integrated to assess coordinated versus asynchronous biological processes and to construct a stage-informed model of coordination-to-divergence across folliculogenesis.

2.2. Dataset retrieval, metadata harmonization, and sample-manifest locking

Raw sequencing data associated with GSE107746 were staged in a dedicated Google Cloud Storage bucket and processed on a high-memory Google Compute Engine instance (c4-highmem-48; 48 vCPUs; 372 GB RAM; 500 GB boot disk) configured for RNA-seq analysis. At project initiation, a complete recursive inventory of bucket contents was generated and frozen to establish a reproducible record of all available sequencing objects. This inventory confirmed the presence of 148 sequencing libraries represented by 296 compressed FASTQ files, consistent with a paired-end sequencing design.

To convert the storage-level inventory into an analysis-ready metadata framework, bucket object paths were parsed to reconstruct run-wise FASTQ pairing. For each library, the SRA run accession, inferred sample identifier, and corresponding read 1 and read 2 file paths were captured. This manifest served as the primary file-level reference for downstream quality control and quantification.

To harmonize raw sequencing files with public repository metadata, the SRA run table associated with the study was imported into the project workspace and joined to the FASTQ manifest using SRA run accessions as the primary key. This integration produced a project-level sample sheet containing, for each library, the run accession (SRR), inferred sample identifier, paired FASTQ paths, SRA sample accession (SRS), BioSample accession (SAMN), and GEO sample accession (GSM). Internal consistency checks confirmed complete accession linkage across all 148 libraries, with no missing values in the joined SRS, SAMN, or GSM fields.

These metadata tables constituted the foundational sample-tracking layer for all subsequent analyses. By separating raw object inventory, file-pair reconstruction, and accession harmonization into explicit manifest-generation steps, the workflow ensured that downstream quality control, transcript quantification, and cross-compartment analyses were anchored to a stable and auditable sample definition framework.

2.3. Final cohort definition and stage-track composition

Following accession harmonization and manifest locking, the finalized analytical cohort was defined at the level of paired-end sequencing libraries. Each retained library was required to have complete identifier linkage across the principal repository accessions used in the workflow, including SRR, SRS, SAMN, and GSM, together with project-level sample naming and paired FASTQ assignment. GEO-derived biological annotations were then standardized so that every library could be assigned consistently to one of the two study compartments, oocyte or granulosa, and to a follicular developmental stage.

To stabilize the biological design before downstream processing, stage and compartment annotations were reviewed jointly rather than independently. This step was used to ensure that the final cohort reflected a coherent developmental framework for later within-compartment and cross-compartment analyses. Libraries lacking an interpretable stage-track assignment were not retained in the finalized analytical set. The resulting metadata table contained no missing track or stage values among the retained libraries.

After curation, the finalized study set was summarized as a fixed stage-by-track count table, which was preserved as a locked design reference for downstream analysis. This table was used to define the shared developmental stages available for cross-compartment comparison and to distinguish them from stages represented in only one compartment. The locked cohort structure was then propagated to all subsequent steps, including quality control, transcript quantification, differential expression analysis, enrichment analysis, and co-expression network construction.

A key outcome of this cohort-definition step was the formal distinction between shared and compartment-specific developmental stages. Primordial, primary, secondary, and antral stages were treated as the common stage framework across the two compartments. By contrast, any stage represented in only one compartment was treated as lineage-specific and was excluded from direct cross-compartment stage matching. This rule was adopted before read-level processing and was maintained throughout downstream interpretation to preserve biological comparability.

2.4. Raw-read quality control, adapter trimming, and post-trimming quality reassessment

Paired-end RNA-seq FASTQ files from the GSE107746 dataset were processed within a structured dual-track re-analysis framework designed for ovarian folliculogenesis, with downstream interpretation planned separately for the oocyte and granulosa-cell compartments. Initial read quality was assessed at the raw FASTQ level using FastQC on each library, followed by MultiQC aggregation to provide a cohort-wide overview of sequence quality, adapter contamination, base composition, duplication-related metrics, GC-content behavior, and overrepresented sequence profiles across all samples.

The raw-read quality assessment identified substantial 3′ adapter contamination and notable terminal quality decay in a large proportion of libraries, supporting the need for preprocessing before transcript quantification. Adapter trimming and quality filtering were then performed for all paired-end libraries using fastp in a manifest-driven workflow. Trimmed FASTQ files were generated for both read mates of each sample, and per-sample fastp HTML and JSON reports were retained as part of the reproducible project record. Post-trimming read quality was reassessed with FastQC and summarized again with MultiQC to evaluate the technical effect of preprocessing before downstream quantification.

2.5. Reference preparation and decoy-aware Salmon indexing

Transcript quantification was performed against a human reference based on the GRCh38 assembly. Protein-coding cDNA transcripts and annotated non-coding RNA transcripts were combined into a unified transcriptome reference. To support decoy-aware transcript quantification, a gentrome was generated by concatenating the combined transcriptome with the GRCh38 primary assembly genome sequence. A decoy list was extracted from the genome FASTA headers and used during Salmon index construction to reduce spurious transcript assignments arising from genomic sequence similarity. The final decoy-aware index was built with Salmon v1.10.2 and retained as the locked reference index for all downstream abundance estimation.

2.6. Transcript quantification from trimmed paired-end libraries

Transcript abundance was estimated from the trimmed paired-end FASTQ files using Salmon v1.10.2 in automatic library-type detection mode. Quantification was performed with selective alignment enabled through –validateMappings, together with sequence-specific and GC-bias correction using –seqBias and –gcBias. A manifest-driven execution model was used throughout, such that each sample identifier was linked to its corresponding trimmed read pairs and quantification output directory. All sample-level runs were performed within a reproducible batch framework using controlled parallelization, stable directory structure, and structured logging. One Salmon output directory and one sample-specific execution log were generated per library.

2.7. Gene-level filtering, normalization, and quality control

Gene-level count, abundance, and effective-length matrices generated by tximport were used as the starting point for downstream preprocessing. Because the study was designed as a dual-track re-analysis of ovarian folliculogenesis, sample-level track labels and developmental stage annotations were retained throughout preprocessing to support subsequent track-aware exploratory, differential, and network-based analyses. The gene-level count matrix was used for filtering and normalization, whereas the abundance and effective-length matrices were retained as companion outputs for downstream interpretation.

Genes with no signal across the cohort were removed prior to expression-aware filtering. Low-expression filtering was then performed on the nonzero gene-level count matrix using an edgeR-based workflow, with expression retention evaluated in the context of the biological group structure defined by the interaction of cellular track and developmental stage. This approach was used to remove genes lacking sufficient expression support for downstream statistical modeling while preserving genes informative within the experimental structure of the dataset.

After filtering, library composition bias was adjusted using trimmed mean of M values (TMM) normalization as implemented in edgeR. Normalized expression values were then represented as log-counts per million (log-CPM) for downstream sample-level quality assessment and exploratory structure analysis. Raw and filtered DGEList objects, normalized log-CPM matrices, gene-filtering summaries, and sample-level normalization tables were retained as locked outputs within the project structure to ensure reproducibility and direct traceability to downstream analyses.

Filtering and normalization were performed as a dedicated preprocessing stage operating on the locked tximport-derived gene-level expression resource. This stage was executed through a script-based workflow with explicit input/output paths, retained tabular summaries, and serialized R objects for downstream reuse. Sample identifiers were preserved throughout all preprocessing objects, allowing direct continuity between transcript quantification, gene-level summarization, exploratory analysis, and later dual-track biological interpretation.

2.8. Principal component analysis and exploratory structure assessment

To examine the global and within-compartment organization of the filtered expression matrix, principal component analysis was performed on the TMM-normalized log-CPM matrix derived from the 24,684 genes retained after expression filtering. PCA was computed across samples using the normalized gene-expression matrix, and the resulting sample scores were integrated with locked metadata containing cellular track and developmental stage annotations.

An initial PCA was performed across the full cohort of 148 libraries to identify the dominant axes of transcriptional variation after filtering and normalization. Because the study was designed as a dual-track re-analysis, samples in the combined PCA were visualized by cellular track and developmental stage. To summarize group-level dispersion in principal component space, 95% normal-data ellipses were drawn for the two cellular compartments.

As the combined analysis showed strong compartment-level separation, PCA was then repeated independently within the oocyte and granulosa subsets to resolve developmental-stage structure after removing the dominant between-track effect. Separate oocyte-only and granulosa-only PCA plots were generated, and 95% normal-data ellipses were drawn by developmental stage. Ellipse generation was restricted to stages with sufficient sample representation to support stable visualization.

To support exploratory quality assessment, a sample-to-sample Pearson correlation matrix was calculated from the normalized log-CPM matrix. For each sample, the median correlation to all other samples was summarized and combined with z score-based screening on the first three principal component axes. Samples showing unusually low median correlation or extreme principal component coordinates according to heuristic thresholds were flagged as candidate outliers for follow-up inspection. This screening procedure was used as an exploratory assessment step and not as an automatic exclusion rule.

2.9. Track-aware differential expression analysis

Because principal component analysis showed that cellular track was the dominant source of global transcriptional variation, differential expression analysis was performed separately within the oocyte and granulosa compartments. This track-aware design was used to prevent the strong between-compartment signal from obscuring developmental-stage contrasts within each lineage.

Differential expression analysis was carried out on the filtered TMM-normalized dataset comprising 24,684 retained genes. The filtered edgeR DGEList object generated during preprocessing was used as input, and stage-wise modeling was performed using the limma-voom framework. For each track, developmental stage was modeled as the sole explanatory factor using a no-intercept design matrix, enabling direct construction of biologically ordered pairwise contrasts. After voom transformation, linear models were fitted and moderated using empirical Bayes shrinkage in limma.

Within the oocyte compartment, contrasts were defined across Primordial, Primary, Secondary, and Antral stages. The primary stepwise contrasts were Primary versus Primordial, Secondary versus Primary, and Antral versus Secondary, together with broader developmental comparisons including Secondary versus Primordial, Antral versus Primary, and Antral versus Primordial. Within the granulosa compartment, contrasts were defined across Primordial, Primary, Secondary, Antral, and Preovulatory stages. Stepwise contrasts included Primary versus Primordial, Secondary versus Primary, Antral versus Secondary, and Preovulatory versus Antral, together with broader comparisons spanning earlier and later granulosa states.

For each contrast, all retained genes were tested and full ranked tables were generated. Genes with Benjamini-Hochberg false discovery rate (FDR) below 0.05 were considered statistically significant. For focused downstream interpretation and figure annotation, more stringent highlighted subsets were additionally defined using FDR < 0.05 together with an absolute log2 fold-change threshold of at least 1. In parallel, an omnibus stage-effect test was performed independently within each track to identify genes varying significantly across the full developmental series.

Differential expression results were summarized using three complementary figure types. First, a contrast-level summary bar plot was generated to compare the number of significant genes identified across all track-specific contrasts at FDR < 0.05. Second, multi-panel volcano plots were produced for key contrasts selected to represent major developmental transitions in each lineage. These included oocyte Antral versus Primordial and Secondary versus Primary, and granulosa Preovulatory versus Primordial, Preovulatory versus Antral, and Antral versus Secondary. In the volcano plots, vertical dashed lines indicated log₂ fold-change thresholds of ±1 and the horizontal dashed line indicated the FDR threshold of 0.05. Third, matched multi-panel MA plots were generated for the same key contrasts to display log₂ fold-change as a function of average expression.

2.10. Functional enrichment analysis

To interpret the biological programs associated with the strongest developmental transitions, enrichment analysis was focused on a small set of contrasts selected from the track-specific differential expression results. Rather than surveying all pairwise comparisons, we concentrated on transitions that were most informative for lineage progression and state change within each compartment. In the oocyte track, these were Antral_vs_Primordial and Secondary_vs_Primary. In the granulosa-cell track, they were Preovulatory_vs_Primordial, Preovulatory_vs_Antral, and Antral_vs_Secondary. This design allowed enrichment to be anchored to biologically meaningful steps spanning broad developmental progression as well as narrower transitional intervals.

For each contrast, genes were separated by direction of change into upregulated and downregulated sets. Only genes meeting both statistical and effect-size criteria were retained, using FDR < 0.05 together with log₂FC ≥ 1 for positively shifted genes and log₂FC ≤ −1 for negatively shifted genes. This was done to ensure that downstream functional interpretation reflected changes that were not only statistically supported but also large enough to be biologically informative.

Functional enrichment was performed with the gprofiler2 R package using the gost() function and hsapiens as the reference organism. For each contrast, the background universe was defined as the full set of genes tested in that specific differential expression comparison, rather than a generic genome-wide background, so that enrichment remained tied to the effective testing space of each model. Annotation sources were restricted to Gene Ontology Biological Process, KEGG, and Reactome, and multiple-testing correction was handled within the g:Profiler framework using false discovery rate control.

To preserve reproducibility while keeping interpretation biologically focused, two complementary output layers were retained for every contrast and direction. The first preserved the original gost() results unchanged and served as the primary statistical record. The second was derived for interpretation and figure construction by excluding ontology root or ultra-broad labels and, for focused review, retaining a specific-term subset with term_size ≤ 500. This approach reduced the influence of very general parent categories without altering the underlying enrichment results.

Biological interpretation was based on recurring functional themes across the filtered outputs rather than on isolated term labels. Enrichment patterns were therefore read in the context of developmental stage progression, compartment identity, and agreement across annotation sources. Categories that appeared to reflect broad annotation behavior rather than central follicular biology, including sparse receptor-family and olfactory-style terms recovered in some weak contrast directions, were retained in the statistical output but were not emphasized in the mechanistic interpretation.

2.11. Weighted gene co-expression network analysis of compartment-specific developmental programs

2.11.1. Track-specific network design

To define coordinated transcriptional programs associated with follicular progression within each cellular compartment, weighted gene co-expression network analysis was performed separately for the oocyte and granulosa tracks. This design was chosen because exploratory structure analysis showed that compartment identity was the dominant source of global transcriptomic variation across the full cohort. A combined network would therefore have been driven primarily by oocyte-granulosa separation and would have obscured the within-lineage developmental structure that was central to this study. Separate networks were constructed instead to resolve stage-associated co-expression architecture within each compartment and to permit biologically grounded comparison of maturation programs across the germ-cell and somatic lineages.

2.11.2. Expression inputs and preparation of track-specific matrices

Network analysis was based on the filtered TMM-normalized logCPM expression matrix generated in the normalization workflow. The project-wide matrix initially contained 24,684 retained genes across 148 samples. Using the verified sample metadata, this expression resource was partitioned into track-specific matrices corresponding to the final locked study cohort. The oocyte matrix comprised 77 samples spanning primordial, primary, secondary, and antral stages. The granulosa matrix comprised 71 samples spanning primordial, primary, secondary, antral, and preovulatory stages.

Because co-expression analysis depends on within-track variability rather than global project-wide expression alone, each compartment-specific matrix was re-evaluated for admissibility after splitting. This step retained all samples in both compartments but excluded genes with zero variance or related non-informative behavior within the relevant track. The final matrices entering network construction therefore contained 24,653 genes for oocytes and 24,403 genes for granulosa cells.

2.11.3. Sample-level review before network construction

Within each compartment, sample relationships were examined before network construction to determine whether the subsequent co-expression structure reflected coherent developmental variation rather than isolated sample-level instability. Hierarchical clustering was used to visualize sample relationships, and sample-level connectivity was calculated to identify poorly integrated samples. Although lower-connectivity samples were observed, they were concentrated predominantly in early developmental states, particularly among primordial oocytes and primordial or secondary granulosa samples, and were interpreted as biologically plausible heterogeneity rather than technical outliers. No samples were excluded at this stage, and all retained samples were carried forward into network construction.

2.11.4. Soft-threshold selection and module detection

Soft-thresholding power was selected independently for each compartment using a signed network framework. Candidate powers were evaluated across a standard range, and the first value achieving satisfactory scale-free topology fit was retained for each track. This resulted in a soft-thresholding power of 9 for the oocyte network and 14 for the granulosa network. Separate powers were retained because the two compartments were not assumed to share identical correlation structure or network density.

Signed adjacency and signed topological overlap were then used to construct the compartment-specific networks. Modules were identified by dynamic tree cutting using conservative parameters chosen to balance biological resolution and module stability. Specifically, minModuleSize = 30, deepSplit = 2, pamRespectsDendro = TRUE, and mergeCutHeight = 0.25 were used. Genes that did not consolidate into stable co-expression structure were assigned to the grey set. Topological overlap matrices were not written to disk at this stage in order to maintain computational efficiency and storage stability.

2.11.5. Module eigengenes and developmental trait relationships

To identify developmental programs associated with follicular progression, module eigengenes were related to ordered stage progression within each compartment. Developmental order was encoded numerically according to the biologically established sequence for each track. For oocytes, this order was primordial, primary, secondary, and antral. For granulosa cells, the order was primordial, primary, secondary, antral, and preovulatory. Stage-specific variables were also retained to allow inspection of individual stage relationships.

Pearson correlation was used to quantify module eigengene–trait associations, and significance was estimated using standard Student asymptotics. False discovery rate correction was applied across module–trait tests. Modules showing strong positive association with ordered progression were interpreted as late-stage maturation programs, whereas modules showing strong negative association were interpreted as early-state programs that declined as maturation proceeded.

2.11.6. Selection of biologically informative modules

Because the purpose of the network analysis was biological synthesis rather than exhaustive cataloguing of all modules, a focused subset of stage-informative modules was prioritized in each compartment. Selection was based on the strength and direction of association with ordered developmental progression. In the oocyte network, turquoise was retained as the dominant late-stage progression module, while royalblue and blue were retained as the principal inverse modules. In the granulosa network, turquoise was retained as the dominant late-stage progression module, while magenta and orangered4 were retained as the principal inverse modules. These selected modules were carried forward for hub-gene prioritization, cross-track comparison, and downstream integration with differential expression and enrichment results.

2.11.7. Hub-gene prioritization within selected modules

To identify candidate drivers within the selected developmental modules, hub genes were ranked using a combined evidence framework rather than by module membership alone. For each selected module, three complementary quantities were calculated: module membership to the relevant eigengene, intramodular connectivity, and gene significance with respect to ordered stage progression. Module membership captured the closeness of each gene to the eigengene-defined module profile, intramodular connectivity quantified local embedding within the module, and gene significance measured the relationship between gene expression and developmental order.

These quantities were standardized and combined into a composite ranking score so that genes were prioritized when they were both structurally central to the module and strongly aligned with the developmental axis represented by that module. Full ranked hub-metric tables were retained for each selected module, and top-ranked hub sets were extracted for biological interpretation and downstream integration.

2.11.8. Gene annotation and cross-compartment comparison

To enable biological interpretation of prioritized hubs, Ensembl gene identifiers were annotated to gene symbols and gene biotypes using the local GRCh38 reference GTF employed elsewhere in the project. This ensured consistency between quantification, gene summarization, and downstream interpretation. Annotated hub tables were generated both for the combined selected-module hub sets and for the module-specific ranked outputs.

Cross-compartment integration focused on the selected stage-informative modules from the oocyte and granulosa networks. For each module, module size, eigengene association with ordered progression, false-discovery-adjusted significance, annotation rate of the prioritized hub set, and representative hub genes were summarized. Overlap between compartments was then assessed using annotated symbols from the top-ranked hub sets, with protein-coding overlap examined separately. This comparison was used to distinguish shared developmental architecture at the network level from compartment-specific molecular driver composition.

2.11.9. Integration with differential expression and enrichment results

To place the selected co-expression modules into the broader biological framework of the study, the prioritized hub sets were overlaid onto the key differential expression contrasts already used for enrichment analysis. For each selected module, the strongest matching developmental contrast was defined by the greatest overlap between top-ranked hubs and direction-consistent differentially expressed genes. These module–contrast relationships were then interpreted alongside the corresponding enrichment themes to determine whether the network-defined developmental programs were concordant with the functional remodeling already identified by stage-wise differential expression analysis.

2.12. Gene-level integration within compartment-specific developmental programmes

To place representative stage-associated genes within the broader compartment-specific regulatory landscape identified in this study, we integrated normalized expression profiles, track-specific differential expression results, WGCNA module context, and the finalized machine-learning framework. Expression was examined separately in oocytes and granulosa cells using the filtered TMM-normalized logCPM matrix. Gene-level developmental association was evaluated within the existing track-specific differential expression framework, including both the omnibus stage-effect model and the pairwise limma-voom contrasts. Network-level contextualization was performed by mapping genes to the annotated WGCNA outputs to determine module assignment, module membership, and stage-association statistics relative to the selected modules prioritized for manuscript interpretation. Where appropriate, module eigengene behavior was examined across follicular stages to assess concordance between gene-level and module-level developmental patterns. Genes were also compared against the finalized machine-learning feature sets to determine whether their contribution was represented directly as a compact predictor or indirectly through selected eigengene structure. This integrative step was intended to relate individual gene behavior to the broader developmental programmes operating within each compartment.

3. Results

3.1. Final cohort structure and stage distribution across compartments

Metadata curation and cohort locking yielded a finalized analytical set of 148 paired-end sequencing libraries. All retained libraries were supported by complete accession linkage across SRR, SRS, SAMN, and GSM identifiers, together with stable project-level sample naming and paired FASTQ assignments. In addition, all retained libraries received complete biological annotation for both cellular compartment and developmental stage, leaving no missing track or stage values in the finalized cohort.

The resulting cohort showed a structured but asymmetrical stage distribution across the two compartments. Within the granulosa track, the finalized set comprised 8 primordial, 15 primary, 6 secondary, 24 antral, and 18 preovulatory libraries. Within the oocyte track, the finalized set comprised 17 primordial, 25 primary, 12 secondary, and 23 antral libraries. These counts established the fixed biological sample structure for all subsequent analyses.

This distribution clarified an important feature of the study design. Four developmental stages, primordial, primary, secondary, and antral, were represented in both compartments and therefore defined the shared framework for stage-matched cross-compartment comparisons. In contrast, the preovulatory stage was present only in the granulosa track and therefore represented a granulosa-specific extension of the developmental series rather than a shared follicular stage across both compartments. Accordingly, downstream dual-track comparisons were restricted to the shared stages, whereas preovulatory samples were interpreted only within the granulosa lineage.

The finalized cohort therefore supported two complementary analytical directions. First, it enabled within-compartment reconstruction of developmental progression in both oocytes and granulosa cells. Second, it enabled explicit comparison of shared-stage transcriptional coordination across compartments while preserving the biologically distinct terminal extension represented by preovulatory granulosa samples.

3.2. Raw-read quality control and trimming outcomes

Initial quality assessment of the 296 raw FASTQ files showed that the dataset retained generally acceptable core sequencing structure but also exhibited widespread technical issues concentrated at read termini. Basic Statistics, Per Base N Content, and Sequence Length Distribution passed uniformly across all files, indicating that the libraries were intact and that read files were structurally consistent. Per Sequence Quality Scores were also broadly acceptable at the raw-read stage, with 265 files passing and 31 flagged as warnings. However, Per Base Sequence Quality showed a markedly less favorable pattern, with only 116 files passing, 72 receiving warnings, and 108 failing. This indicated substantial heterogeneity in terminal base-quality behavior across the dataset.

The strongest technical signal prompting preprocessing was adapter contamination. In the raw-read MultiQC summary, Adapter Content passed in only 1 of 296 files, while 34 files were flagged as warnings and 261 failed. Visual inspection of the adapter-content profiles showed a strong increase in adapter-associated signal toward the 3′ ends of many reads. Representative libraries displayed substantial terminal adapter burden, with end-of-read adapter content reaching approximately 22.4% in 10G_R1, 27.9% in 10G_R2, 14.8% in 20G_R1, 28.1% in 28G_R2, 49.0% in A10O_R1, and 51.8% in A10O_R2. Per-base quality profiles further showed progressive quality decay toward late sequencing cycles, particularly in read 2 and in a subset of read 1 files. In several libraries, end-cycle quality values declined into the low- to mid-20 Phred range, and in some cases dropped below Q20. These combined patterns indicated that direct quantification of the raw FASTQ files would not be optimal.

Trimming with fastp was therefore applied to all 148 paired libraries. The preprocessing workflow completed successfully across the full dataset and produced 296 trimmed FASTQ files together with 148 HTML reports and 148 JSON reports. A representative validation run on sample A10O illustrated the scale of improvement achieved by preprocessing. For this library, fastp identified canonical Illumina TruSeq adapter sequences in both reads and reported 28,198,040 reads with adapter trimming, with 3,306,656,610 bases removed due to adapter contamination. Read-quality metrics improved substantially after filtering. In read 1, the proportion of Q30 bases increased from 69.13% before filtering to 89.24% after filtering. In read 2, the corresponding increase was from 64.35% to 84.13%. These results were consistent with the broader dataset-wide QC patterns and supported the decision to use trimmed rather than raw reads in downstream analysis.

Post-trimming FastQC and MultiQC reassessment confirmed that preprocessing substantially improved the most technically relevant QC modules. For Per Base Sequence Quality, the number of passing files increased from 116 in the raw dataset to 244 after trimming, while failures decreased from 108 to 17. Per sequence quality scores improved from 265 passes and 31 warnings in the raw reads to 296 passes and no warnings after trimming. Adapter Content also improved markedly, with passing files increasing from 1 to 119 and failures decreasing from 261 to 126. More modest improvements were observed for sequence duplication levels and overrepresented sequences.

Not all FastQC warning classes were eliminated after trimming. Per base sequence content remained largely unchanged, with only 2 files passing in both raw and trimmed datasets. Per sequence GC Content also remained problematic in many libraries and showed no systematic improvement after trimming. These residual warnings were interpreted as consistent with transcriptome-sequencing libraries, in which biological composition bias, transcript abundance heterogeneity, and non-random sequence structure commonly produce persistent FastQC alerts even when data remain suitable for expression analysis. Likewise, sequence length distribution changed from universal pass status in the raw reads to universal warning status after trimming, which was expected because adapter and quality trimming generated variable read lengths across the dataset.

Overall, the combined before-and-after QC evidence showed that preprocessing substantially improved terminal base quality and reduced adapter-associated contamination while preserving complete library representation across all samples. On this basis, the trimmed FASTQ files were selected as the finalized inputs for transcript quantification.

3.3. Preprocessing improved technical read quality and justified use of trimmed reads for downstream quantification

The preprocessing stage was completed for all 148 paired-end libraries, yielding 296 trimmed FASTQ files together with 148 fastp HTML reports and 148 fastp JSON reports. Comparison of raw and post-trimming MultiQC summaries showed clear improvement in the major technical quality modules most relevant to downstream quantification. For per-base sequence quality, the number of FASTQ files classified as FastQC PASS increased from 116 before trimming to 244 after trimming, while FAIL classifications decreased from 108 to 17. For per-sequence quality scores, all 296 FASTQ files passed after trimming, compared with 265 PASS and 31 WARN at the raw-read stage. Adapter-content profiles also improved substantially, with PASS classifications increasing from 1 to 119 and FAIL classifications decreasing from 261 to 126 after trimming.

These results indicated that trimming successfully reduced the dominant technical burdens present in the raw data, particularly 3′ adapter carryover and terminal quality deterioration, Table 1. Residual warnings in modules such as per-base sequence content, per-sequence GC content, duplication-related measures, and overrepresented sequences were retained as part of the dataset profile but were interpreted as compatible with transcriptomic library behavior rather than as evidence of dataset failure. On this basis, the trimmed FASTQ files were accepted as the locked downstream inputs for transcript abundance estimation.

Table 1.

FastQC module status before and after trimming across 296 FASTQ files derived from 148 paired-end libraries.

FastQC module	Raw reads			Trimmed reads			Main change
	Pass	Warn	Fail	Pass	Warn	Fail
Basic Statistics	296	0	0	296	0	0	Unchanged
Per Base Sequence Quality	116	72	108	244	35	17	Strong improvement
Per Tile Sequence Quality	189	102	5	163	126	7	Mixed
Per Sequence Quality Scores	265	31	0	296	0	0	Strong improvement
Per Base Sequence Content	2	38	256	2	40	254	Largely unchanged
Per Sequence GC Content	26	18	252	18	9	269	Biologically driven
Per Base N Content	296	0	0	296	0	0	Unchanged
Sequence Length Distribution	296	0	0	0	296	0	Expected post-trim shift
Sequence Duplication Levels	8	32	256	13	47	236	Moderate improvement
Overrepresented Sequences	0	10	286	1	22	273	Minor improvement
Adapter Content	1	34	261	119	51	126	Strong improvement

3.4. Salmon quantification completed successfully across all libraries

Following completion of preprocessing, transcript abundance estimation was performed for all 148 paired-end libraries against the decoy-aware GRCh38 Salmon index. Quantification completed successfully for all samples, and a valid quant.sf abundance file was produced for each library. No sample-level failures were recorded in the batch execution logs. The final output therefore comprised a complete transcript-level quantification layer for all 148 libraries under a single reference configuration and a single locked quantification workflow.

Each completed sample directory contained the expected Salmon abundance table together with run metadata, providing a consistent and analysis-ready resource for downstream transcript-to-gene summarization. Because all libraries were quantified from the accepted trimmed-read manifest and against the same decoy-aware index, the resulting abundance estimates were generated under a uniform computational framework. This completed the transcript-level quantification stage and established the input required for the next phase of the analysis, namely gene-level summarization with tximport followed by filtering, normalization, exploratory structure analysis, differential expression modeling, and downstream dual-track biological interpretation.

3.5. Expression filtering reduced the imported gene set to an analysis-ready matrix

The tximport stage yielded a gene-level expression resource comprising 81,412 genes across 148 RNA-seq libraries. Removal of genes with zero counts across all samples reduced this set to 63,607 genes with measurable expression signal in at least one library. Subsequent expression-aware filtering further reduced the dataset to 24,684 genes retained for downstream analysis. Thus, the filtering stage removed a substantial proportion of low-information features while preserving a focused analysis matrix supported by observable expression across the cohort.

Following expression filtering, TMM normalization was successfully applied to the retained 24,684-gene matrix across all 148 samples. The resulting filtered and normalized expression object had final dimensions of 24,684 genes by 148 libraries. Sample-level summaries confirmed preservation of all libraries through preprocessing and recorded both filtered library sizes and normalization factors for each sample. These outputs established the normalized gene-expression matrix that will be used for downstream quality assessment, exploratory structure analysis, differential expression modeling, and network-based analyses, Fig. 1A.

Fig. 1.

Principal component analysis of the filtered TMM-normalized log-CPM matrix. (A) Combined PCA across all 148 samples. Points represent individual libraries, colored by cellular track and shaped by developmental stage. Ellipses indicate 95% normal-data ellipses for the oocyte and granulosa compartments. PC1 explained 28.25% and PC2 explained 7.33% of the variance. (B) Oocyte-only PCA across 77 samples, colored by developmental stage. Ellipses indicate 95% normal-data ellipses by stage. PC1 explained 18.01% and PC2 explained 9.60% of the variance. (C) Granulosa-only PCA across 71 samples, colored by developmental stage. Ellipses indicate 95% normal-data ellipses by stage. PC1 explained 15.18% and PC2 explained 10.11% of the variance.

The filtering and normalization stage generated a complete set of locked downstream objects, including raw and filtered DGEList containers, normalized log-CPM matrices, a gene-filtering summary table, and sample-level normalization metadata. Because sample identifiers were preserved unchanged across all preprocessing outputs, the normalized matrix remains directly linkable to the original study metadata, track assignments, developmental stages, and prior transcript-level quantification outputs. This completed the core preprocessing transition from imported gene-level estimates to a compact analysis-ready expression resource for the subsequent stages of the dual-track re-analysis.

3.6. Oocyte-only PCA revealed progressive stage-related restructuring with partial continuity between adjacent stages

To resolve developmental-stage structure within the oocyte compartment, PCA was repeated using only the 77 oocyte samples. In this analysis, PC1 explained 18.01% of the variance and PC2 explained 9.60%, with PC3 contributing a further 6.38%. Stage-aware visualization showed that primordial and primary oocyte samples clustered predominantly on the negative side of PC1 and displayed substantial overlap, indicating continuity between these early developmental states. Secondary oocyte samples shifted toward more positive PC1 values and were concentrated largely in the upper-right region of the PCA space. Antral oocyte samples occupied the most positive PC1 range and showed broader dispersion along PC2, suggesting increased heterogeneity at this stage relative to earlier oocyte states. Overall, the oocyte-only PCA supported a stage-associated progression from early negative-PC1 states toward later positive-PC1 states, while also indicating that transitions between adjacent stages remained gradual rather than sharply discrete, Fig. 1B.

3.7. Granulosa-only PCA showed stronger stage resolution, with preovulatory samples forming the most distinct cluster

Track-stratified PCA was also performed on the 71 granulosa samples. In the granulosa-only analysis, PC1 explained 15.18% of the variance and PC2 explained 10.11%, with PC3 explaining 5.47%. Stage-level structure was more clearly resolved in granulosa than in oocytes. Primordial granulosa samples occupied predominantly negative PC1 and positive PC2 space and showed the broadest within-stage dispersion. Primary granulosa samples clustered nearer the center of the PCA space, whereas antral granulosa samples formed a comparatively compact cluster at mildly positive PC1 and negative PC2 values. Preovulatory granulosa samples formed the most distinct and tightest cluster, occupying strongly positive PC1 and positive PC2 space with minimal overlap with the other stages. Secondary granulosa samples were more broadly distributed and overlapped partially with neighboring stages, consistent with a more transitional or heterogeneous transcriptional state. These patterns indicate that developmental progression is especially well resolved within the granulosa lineage once the dominant oocyte-versus-granulosa effect is removed.

3.7.1. Exploratory structure analysis supports a track-aware downstream analytical framework

The exploratory analyses indicate that the normalized dataset is organized hierarchically. At the global level, the strongest source of transcriptional variation reflects cellular compartment, with clear separation between oocyte and granulosa expression profiles. Within each compartment, developmental stage contributes additional structure, with especially sharp stage resolution in granulosa cells and more continuous progression in oocytes. These findings support the dual-track design of the study and justify downstream differential and systems-level analyses in a track-aware framework.

3.7.2. Exploratory screening identified a small granulosa-enriched subset of candidate outliers

Heuristic outlier screening identified five candidate samples. All five belonged to the granulosa compartment, with four assigned to the primordial granulosa stage and one to the secondary granulosa stage, Fig. 1C. These samples were flagged primarily because of reduced median inter-sample correlation rather than because of extreme displacement along the major principal component axis. Despite this localized signal, the overall cohort structure remained coherent, and the candidate outliers were retained for downstream analysis pending targeted review rather than automatic exclusion.

3.8. Stage-wise differential expression was extensive in both oocytes and granulosa cells

Track-aware differential expression analysis was performed on 24,684 retained genes using separate limma-voom models for oocyte and granulosa samples. In both compartments, developmental progression was associated with widespread transcriptional change. Across all contrasts combined, 21,726 unique genes in oocytes and 21,738 unique genes in granulosa cells were significant at FDR < 0.05 in at least one comparison. When the more stringent criterion of FDR < 0.05 together with |log2 fold-change| ≥ 1 was applied, 21,373 unique oocyte genes and 21,695 unique granulosa genes remained represented across the union of track-specific contrasts, Fig. 2. These values indicate that developmental remodeling affected the majority of the filtered gene-expression space in both lineages.

Fig. 2.

Summary of significant genes across track-specific developmental contrasts. Bar plot showing the number of genes significant at FDR < 0.05 for each oocyte and granulosa differential expression contrast. Differential expression was modeled separately within each cellular track using limma-voom on the filtered TMM-normalized dataset.

3.8.1. Oocyte contrasts showed large transcriptional shifts across both adjacent and broad developmental transitions

Within the oocyte compartment, all tested contrasts yielded large numbers of significant genes. Primary versus Primordial identified 10,551 genes at FDR < 0.05, Secondary versus Primary identified 10,553 genes, and Antral versus Secondary identified 9498 genes. These results indicate that substantial transcriptional change occurred at each successive developmental transition rather than being confined to broad early-versus-late comparisons.

Broader oocyte contrasts were also highly significant. Secondary versus Primordial identified 5955 genes, Antral versus Primary identified 11,770 genes, and Antral versus Primordial identified 15,589 genes, representing the strongest oocyte contrast tested. Directional summaries showed marked asymmetry across transitions. For example, Primary versus Primordial was dominated by downregulated genes at the stricter FDR and fold-change threshold, whereas Secondary versus Primary showed a stronger upregulated component, with 7653 upregulated versus 2661 downregulated genes. The largest figure-level oocyte contrast, Antral versus Primordial, contained 5820 upregulated and 9483 downregulated genes under the stringent threshold and reached a maximum absolute log2 fold-change of 12.90. Together, these findings support progressive but nonuniform transcriptional remodeling across oocyte development.

3.8.2. Granulosa cells showed especially strong late-stage transcriptional remodeling

Within the granulosa compartment, the extent of differential expression varied more sharply across contrasts. The weakest granulosa contrast was Secondary versus Primordial, which identified only 578 genes at FDR < 0.05, while Secondary versus Primary yielded 1403 significant genes. These relatively modest early-stage contrasts suggest that some intermediate granulosa states remain transcriptionally closer to one another than comparable oocyte stages.

In contrast, later granulosa transitions showed very strong differential expression. Antral versus Secondary identified 10,102 genes, Preovulatory versus Antral identified 11,199 genes, Antral versus Primordial identified 14,171 genes, Preovulatory versus Secondary identified 16,469 genes, and Preovulatory versus Primordial identified 17,488 genes, the strongest granulosa contrast in the study. Under the stringent FDR and fold-change threshold, these later contrasts were dominated by large downregulated gene sets. For example, Preovulatory versus Primordial contained 3937 upregulated and 13,548 downregulated genes, whereas Preovulatory versus Antral contained 2707 upregulated and 8139 downregulated genes. The maximum absolute log2 fold-changes among significant genes were also large in these contrasts, reaching 11.90 for Preovulatory versus Primordial and 11.22 for Preovulatory versus Antral. These results indicate that granulosa transcriptional remodeling becomes particularly pronounced in later follicular development and is strongest at the transition into the preovulatory state.

3.8.3. Differential expression patterns were consistent with the PCA-defined structure

The differential expression results were highly consistent with the exploratory PCA findings. In oocytes, large numbers of significant genes were already evident between adjacent stages, supporting the interpretation of continuous but progressive transcriptional restructuring across development. In granulosa cells, early-stage contrasts were comparatively weaker, whereas contrasts involving antral and preovulatory stages were much stronger, mirroring the sharper stage resolution observed in the granulosa-specific PCA. The DE summary figure, volcano plots, and MA plots Fig. 3, Fig. 4; collectively reinforced this pattern by showing that broad late-stage granulosa contrasts and broad early-to-late oocyte contrasts produced the strongest signals.

Fig. 3.

Volcano plots for key developmental contrasts in oocytes and granulosa cells.

Multi-panel volcano plots showing log₂ fold-change versus -log10 FDR for selected track-specific contrasts. Dashed vertical lines indicate log₂ fold-change thresholds of ±1, and the dashed horizontal line indicates the FDR threshold of 0.05. Highlighted signals correspond to genes meeting FDR < 0.05 and |log₂ fold-change| ≥ 1.

Fig. 4.

MA plots for key developmental contrasts in oocytes and granulosa cells.

Multi-panel MA plots showing log₂ fold-change as a function of average expression (AveExpr) for selected track-specific contrasts. Dashed horizontal lines indicate log₂ fold-change thresholds of ±1. The plots summarize the abundance-dependent structure of differential expression within each lineage.

3.8.4. Omnibus stage-effect tests identified strongly stage-associated genes in both lineages

The omnibus stage-effect tests further confirmed strong developmental structure within each track. In oocytes, the highest-ranking stage-associated genes included UCHL1, ZFAND2A, H3—3B, BOD1, EIF1, YWHAB, NLRP9, TPT1, MT-CO1, and one highly significant but currently unannotated lncRNA transcript (ENSG00000290739). In granulosa cells, the top-ranked stage-associated genes included ACTB, RPLP0, EEF1A1, TPT1, RPS8, RACK1, RPL13A, RPL15, RPL3, and RBP1. These genes showed exceptionally large F statistics and extremely small adjusted p values, indicating robust stage-associated expression dynamics across the developmental series. The annotated omnibus results therefore provide a more biologically interpretable foundation for downstream pathway enrichment and network-level analysis than Ensembl identifiers alone.

3.8.5. Track-aware DE analysis supports lineage-specific downstream interpretation

Taken together, the stage-wise differential expression results indicate that both lineages undergo extensive transcriptional remodeling, but with distinct temporal patterns. Oocytes show substantial differential expression even across adjacent stage transitions, consistent with a progressive developmental continuum. Granulosa cells, by contrast, show relatively modest early-stage contrasts but much larger late-stage shifts, especially in comparisons involving the preovulatory state. These results reinforce the rationale for track-aware downstream interpretation and identify a focused set of contrasts for enrichment analysis and systems-level follow-up.

3.9. HSP90AA1 showed distinct stage-dependent expression dynamics in oocytes and granulosa cells

Among the stage-associated genes highlighted by the differential expression framework, HSP90AA1 displayed a clear lineage-specific pattern. In oocytes, HSP90AA1 was strongly significant in the omnibus stage-effect model and showed a gradual decline across development, with average normalized expression decreasing from 11.53 in primordial oocytes and 11.50 in primary oocytes to 11.25 in secondary oocytes and 10.59 in antral oocytes. This trend was reflected in modest but significant late-stage contrasts, including antral versus primary (log₂ fold-change = −0.90, adjusted P = 4.61 × 10^-4) and antral versus primordial (log₂ fold-change = −0.94, adjusted P = 4.39 × 10^-3), indicating progressive attenuation rather than an abrupt transcriptional switch. In granulosa cells, by contrast, HSP90AA1 showed a much sharper developmental pattern. Expression was comparable in primordial and primary granulosa cells (11.46 and 11.37, respectively), rose markedly in secondary granulosa cells (14.18), then declined in antral cells (11.23) and reached its lowest level in preovulatory granulosa cells (9.34). Consistent with this profile, HSP90AA1 was significantly upregulated in secondary versus primary granulosa cells (log₂ fold-change = 2.81, adjusted P = 1.17 × 10^-4) and secondary versus primordial granulosa cells (log₂ fold-change = 2.71, adjusted P = 8.81 × 10^-3), but was strongly downregulated in antral versus secondary, preovulatory versus antral, preovulatory versus primordial, and preovulatory versus secondary contrasts. Together, these results indicate that HSP90AA1 behaves as a broadly stage-associated gene in both compartments, but its developmental remodeling is far more pronounced in granulosa cells than in oocytes, Fig. 5.

Fig. 5.

HSP90AA1 expression across follicular stages in oocyte and granulosa tracks. Filtered TMM-normalized logCPM values for HSP90AA1 are shown across developmental stages in the oocyte and granulosa compartments. Points represent individual samples, boxes indicate the interquartile range, the center line marks the median, and whiskers extend to 1.5 × the interquartile range. The figure highlights a gradual decrease in oocytes across development and a sharper secondary-stage peak followed by marked reduction in antral and preovulatory granulosa cells.

3.10. Functional enrichment resolves distinct developmental programs in oocytes and granulosa cells

To define the biological programs underlying the strongest transcriptional transitions, enrichment analysis was focused on five contrasts selected to capture both broad developmental progression and narrower stage-to-stage shifts within each cellular compartment. The resulting patterns showed that oocyte and granulosa maturation are not simply different in magnitude. They are organized around distinct functional trajectories.

In oocytes, the broad antral-versus-primordial contrast was dominated by genome-regulatory programs associated with late developmental remodeling. Genes elevated in antral oocytes were strongly enriched for DNA recombination, chromosome segregation, regulation of cell-cycle phase transition, DNA replication, double-strand break repair, and related recombination-associated processes (Fig. 6A). These terms converged on chromosome dynamics and genome-integrity control rather than dispersing across unrelated pathways, indicating that progression to the antral state is accompanied by coordinated activation of regulatory systems involved in DNA maintenance, chromosomal organization, and cell-cycle-linked remodeling. By contrast, the reciprocal primordial-enriched direction yielded a less coherent set of terms and was not prioritized as the main mechanistic axis of interpretation.

Fig. 6.

Oocyte enrichment programmes across major developmental transitions. (A). Antral versus primordial oocytes: enrichment for DNA recombination, chromosome segregation, DNA replication and repair-associated functions, consistent with late-stage genome-regulatory remodeling. (B). Primary versus secondary oocytes: enrichment for RNA splicing, mRNA processing, ribonucleoprotein biogenesis and translational control, consistent with intermediate-stage reorganization of post-transcriptional regulatory architecture. Representative enriched terms are shown from the filtered specific enrichment layer (term_size ≤ 500). Dot size indicates overlap size; the x axis shows -log10(adjusted p).

A different pattern emerged in the secondary-versus-primary oocyte comparison. Here, the strongest functional signal was carried by genes elevated in primary oocytes, which were enriched for RNA splicing, mRNA processing, spliceosome-associated reactions, ribonucleoprotein complex biogenesis, protein-RNA complex assembly, translational initiation, and related post-transcriptional regulatory functions (Fig. 6B). This profile suggests that the primary stage is supported by a comparatively strong RNA-processing and translational-control architecture that becomes reorganized as oocytes move into the secondary state. Notably, genes elevated in secondary oocytes did not yield a comparably consolidated enrichment signature under the same framework. Together, these results indicate that oocyte maturation combines broad late-stage activation of genome-integrity and chromosome-regulatory programs with an intermediate reconfiguration of post-transcriptional control.

Granulosa cells followed a clearly different developmental logic. In the secondary-to-antral transition, genes elevated in antral granulosa were already enriched for mitochondrial translation, mitochondrial gene expression, cellular respiration, aerobic respiration, ATP synthesis coupled electron transport, oxidative phosphorylation, and related energetic functions (Fig. 7A). This indicates that metabolic activation is established by the antral stage rather than appearing only at the terminal preovulatory endpoint. The reciprocal secondary-enriched direction yielded only a sparse set of terms dominated by receptor-family and olfactory-style annotations, which were retained in the statistical output but not emphasized in biological interpretation.

Fig. 7.

Granulosa enrichment programmes define progressive metabolic activation and late-state specialization. (A) Antral versus secondary granulosa cells: onset of mitochondrial and respiratory activation. (B) Antral versus preovulatory granulosa cells: enrichment for replication- and chromosome-associated programmes. (C) Preovulatory versus antral granulosa cells: energetic intensification and mitochondrial specialization. (D) Preovulatory versus primordial granulosa cells: broad late-stage bioenergetic remodeling. Representative enriched terms are shown from the filtered specific enrichment layer (term_size ≤ 500). Dot size indicates overlap size; the x axis shows -log10(adjusted p).

This energetic programme intensified further in the antral-to-preovulatory transition. Genes elevated in preovulatory granulosa were strongly enriched for oxidative phosphorylation, aerobic respiration, ATP-generating metabolism, proton motive force-driven ATP synthesis, electron transport, mitochondrial organization, and respiratory chain assembly (Fig. 7C). In parallel, genes relatively higher in antral granulosa were enriched for DNA replication, chromosome segregation, nuclear division, replication fork processing, double-strand break repair, and checkpoint-associated regulatory processes (Fig. 7B). This contrast is especially informative because it suggests that the final transition into the preovulatory state is not marked by stronger replication-associated activity. Instead, granulosa cells appear to shift away from a more proliferative or chromosome-dynamics-oriented antral programme and toward a metabolically intensified preovulatory identity.

The same late-stage logic remained evident when preovulatory granulosa was compared with primordial granulosa across the broader developmental span. Genes elevated in preovulatory granulosa were strongly enriched for generation of precursor metabolites and energy, aerobic and cellular respiration, oxidative phosphorylation, mitochondrial translation, ATP synthesis coupled electron transport, mitochondrial gene expression, and respiratory chain complex assembly (Fig. 7D). In this broader comparison, the energetic and organellar signature was even more pronounced, supporting the view that late granulosa maturation is accompanied by extensive mitochondrial and bioenergetic reprogramming. By contrast, genes relatively higher in primordial granulosa again yielded only a limited set of receptor-family and olfactory-associated annotation categories, consistent with annotation-driven overrepresentation rather than a dominant developmental mechanism.

These enrichment results support a dual-track model of follicular maturation in which oocytes and granulosa cells undergo distinct but complementary forms of remodeling. Oocyte progression is characterized by broad genome-regulatory restructuring across the primordial-to-antral span and by intermediate reorganization of RNA-processing and translational-control architecture between the primary and secondary stages. Granulosa maturation, by contrast, is marked by progressive mitochondrial and energetic activation beginning at the antral stage and intensifying further in preovulatory cells, alongside a relative decline in replication- and chromosome-associated functions. This divergence reinforces the value of treating the oocyte and granulosa compartments as biologically linked, yet transcriptionally non-equivalent, developmental systems.

3.11. Co-expression programs underlying follicular progression in the oocyte and granulosa compartments

3.11.1. Oocyte maturation is organized by a dominant late-stage program together with declining early-state modules

Oocyte maturation was dominated by a large late-stage co-expression program that strengthened progressively across the primordial-to-antral trajectory. The clearest representation of this pattern was the turquoise module, which showed a very strong positive association with ordered stage progression and emerged as the largest non-grey module in the oocyte network. This indicates that later oocyte development is not defined by isolated marker changes, but by broad coordination across a substantial fraction of the transcriptome. In biological terms, the antral oocyte state appears to reflect consolidation of a mature transcriptional architecture rather than a simple extension of earlier expression patterns.

This late-stage program was accompanied by two strongly inverse modules, royalblue and blue, both of which declined as follicular progression advanced. Their inverse association with developmental order indicates that early oocyte identity is itself structured, rather than merely representing an undifferentiated baseline. Thus, the oocyte trajectory appears to involve two coordinated processes operating in parallel: emergence of a dominant late maturation program and attenuation of earlier-state transcriptional organization.

The biological content of these inverse modules suggests that they capture distinct layers of immature oocyte architecture. The royalblue module was enriched in hubs such as COX6A1, RPS14, RPL19, RPL11, RPS8, and PFDN2, pointing to an early biosynthetic and translational program. By contrast, the blue module included hubs such as GPM6B, RELN, DMD, PTPRE, LOXL2, SATB1, and FGF13, suggesting a broader regulatory and structural program. These patterns imply that early oocyte states are maintained by multiple coordinated systems, including translational infrastructure and wider regulatory architecture, both of which are progressively reduced as maturation proceeds.

The late-stage oocyte module also aligned closely with the strongest differential expression axis in the compartment. The turquoise module was most strongly anchored to the antral-versus-primordial comparison, and its leading hubs showed consistent agreement with the direction of stage progression. This convergence between co-expression structure and differential expression strengthens the interpretation that the late oocyte state is defined by broad remodeling rather than by a limited subset of stage markers. When considered alongside the enrichment results, this late-stage program is most consistent with the genome-regulatory and chromosome-associated remodeling already observed across the primordial-to-antral span, including DNA replication, chromosome segregation, recombination, and repair-related biology.

A more focused transition was evident between the primary and secondary stages. In that interval, enrichment had highlighted RNA splicing, mRNA processing, ribonucleoprotein biogenesis, and translational control on the primary-high side, while the secondary-high side did not yield a similarly consolidated functional signal. This asymmetry fits the network-level pattern, suggesting that part of the early oocyte program fades gradually rather than being replaced by a sharply defined intermediate module. Taken together, the oocyte network supports a developmental model in which early translational and regulatory states are progressively dismantled while a broad late-stage maturation program becomes increasingly dominant, Fig. 8.

Fig. 8.

Oocyte module–trait relationships identify a dominant late-stage maturation program and inverse early-state modules. Heatmap showing Pearson correlations between selected oocyte module eigengenes and developmental traits across the primordial, primary, secondary, and antral stages. Tile colour reflects the direction and magnitude of association, with red indicating positive correlation and blue indicating negative correlation. Values within tiles show the correlation coefficient and false-discovery-adjusted significance. The turquoise module marks the dominant late-stage oocyte program, whereas royalblue and blue represent inverse early-state programs that decline with progression across follicular development. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.11.2. Granulosa maturation is marked by late bioenergetic consolidation and loss of earlier somatic-state architecture

Granulosa maturation was likewise organized by a dominant late-stage co-expression program, but the biological character of that program was distinct from the oocyte trajectory. The strongest signal was again a turquoise module that increased markedly with ordered developmental progression and occupied the largest non-grey fraction of the granulosa network. This indicates that later granulosa maturation is also accompanied by strong transcriptome-wide coordination. However, in contrast to the oocyte compartment, the associated biology was much more clearly aligned with late metabolic and bioenergetic specialization.

This interpretation is strongly supported by the earlier enrichment analysis. Genes increased in antral relative to secondary granulosa cells were enriched for mitochondrial translation, mitochondrial gene expression, oxidative phosphorylation, respiration, and ATP-generating metabolism, and these energetic themes intensified further in the preovulatory contrasts. The granulosa turquoise module therefore appears to capture the systems-level organization of this late somatic maturation state. Rather than representing scattered changes in metabolic genes, it defines a coordinated late-stage program that becomes increasingly prominent as granulosa cells approach the preovulatory condition.

Set against this late-stage module were two inverse programs, magenta and orangered4, both of which declined with developmental progression. These modules indicate that immature granulosa states are supported by structured transcriptional programs that are progressively lost as follicles advance. The magenta module behaved as the stronger general inverse program and included hubs such as HSPB8, COL14A1, IER3, NKX3–1, TAC1, KRT18, HLA-A, PTGDS, and ETS1. This constellation suggests an earlier granulosa architecture involving signaling, stress-associated biology, extracellular or structural features, and cell-state identity. The orangered4 module, with hubs such as DSP, CDH3, PRDM1, MYLIP, JAG1, SOX4, and KCNK1, points to an additional early somatic layer with stronger regulatory and adhesion-related character.

Although both inverse granulosa modules mapped against late maturation, they were not identical in their developmental behavior. Magenta appeared to represent a broader immature-state program across the granulosa trajectory, whereas orangered4 showed a more selective relationship to later remodeling, especially around the antral-to-preovulatory boundary. This distinction suggests that granulosa maturation is not simply the loss of one early program followed by gain of one late program. Rather, it involves attenuation of multiple layers of immature somatic architecture, alongside emergence of a late energetic state that becomes especially prominent in preovulatory cells.

The link between the late granulosa program and the broad preovulatory-versus-primordial contrast was particularly strong. The turquoise module aligned closely with this span, and its leading hubs were directionally concordant with that contrast, reinforcing the view that late granulosa maturation reflects large-scale somatic reorganization. Overall, the granulosa network supports a model in which follicular progression culminates in a coordinated bioenergetic program, while earlier signaling, stress-related, regulatory, and adhesion-associated states are progressively down-weighted, Fig. 9.

Fig. 9.

Granulosa module–trait relationships resolve late bioenergetic maturation together with inverse early-state somatic programs. Heatmap showing Pearson correlations between selected granulosa module eigengenes and developmental traits across the primordial, primary, secondary, antral, and preovulatory stages. Tile colour reflects the direction and magnitude of association, with red indicating positive correlation and blue indicating negative correlation. Values within tiles show the correlation coefficient and false-discovery-adjusted significance. The turquoise module captures the dominant late-stage granulosa maturation program, whereas magenta and orangered4 represent inverse early-state programs that are progressively attenuated as granulosa cells approach the preovulatory state. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.11.3. Parallel developmental logic across compartments conceals distinct molecular driver architectures

When the two compartments were considered together, a shared architectural principle became clear. Both oocytes and granulosa cells showed one dominant module that increased with developmental progression and one or more inverse modules that declined as follicles matured. At this level, the logic of follicular progression is parallel across compartments: maturation involves consolidation of a coordinated late-stage transcriptional state together with attenuation of earlier-state co-expression structure.

Despite this common systems-level pattern, the molecular composition of the selected modules was largely compartment-specific. The dominant late-stage modules in the two compartments shared very little hub-level overlap, and the inverse modules were likewise built from mostly distinct gene sets. This indicates that oocytes and granulosa cells do not arrive at maturity through a shared set of central regulators. Instead, each lineage follows its own molecular route while conforming to the same broader developmental architecture.

The biological interpretation of these compartment-specific programs also differed in a coherent way. In oocytes, late maturation was linked most clearly to genome-regulatory and chromosome-associated remodeling, whereas the inverse modules pointed to loss of earlier translational and regulatory architecture. In granulosa cells, the late program was much more tightly associated with mitochondrial and energetic activation, while the inverse modules reflected earlier signaling, stress-state, regulatory, and adhesion-associated biology. Thus, the two compartments display parallel developmental organization but clearly distinct biological content.

The WGCNA results extend the earlier differential expression and enrichment analyses by showing that follicular progression is not simply a collection of independently changing genes. Instead, developmental change is organized into coherent co-expression programs whose timing, direction, and leading hubs reinforce the compartment-specific maturation logic of the oocyte and granulosa lineages. The overall picture is therefore one of coordinated but non-identical development: both compartments progress through ordered transcriptional consolidation, but each does so through its own lineage-specific molecular architecture, Table 2.

Table 2.

Cross-compartment summary of selected stage-associated WGCNA modules.

Compartment	Module	Module size	Association with stage progression (r)	FDR	Direction	Biological interpretation	Representative hub genes
Oocyte	turquoise	5294	0.909	5.57E-28	Positive (late-stage)	Late-stage oocyte maturation program	GEMIN5, BOD1, TUBAL3, BTG4, GNAT2, ZNF840P
Oocyte	royalblue	159	−0.755	6.75E-14	Negative (early-state)	Early oocyte translational/biosynthetic program	COX6A1, RPS14, RPL19, RPL11, RPS8, PFDN2
Oocyte	blue	2982	−0.745	2.26E-13	Negative (early-state)	Early oocyte regulatory/structural program	GPM6B, RELN, DMD, PTPRE, LOXL2, SATB1
Granulosa	turquoise	4602	0.918	7.54E-27	Positive (late-stage)	Late-stage granulosa maturation program	MRO, GRAMD1B, CNPY2, MRAP, ZDHHC12, SHKBP1
Granulosa	magenta	386	−0.718	9.95E-11	Negative (early-state)	Early granulosa signaling/stress-state program	HSPB8, COL14A1, IER3, NKX3–1, TAC1, KRT18
Granulosa	orangered4	80	−0.578	4.32E-06	Negative (early-state)	Early granulosa regulatory/adhesion program	DSP, CDH3, PRDM1, MYLIP, JAG1, KCNK1

3.12. Machine-learning validation of compartment-specific developmental programs

To test whether developmental stage could be recovered from biologically informed transcriptomic representations within each compartment, we performed track-specific multiclass classification using repeated stratified 5-fold cross-validation and Elastic Net multinomial logistic regression. This analysis was intended as an internal validation layer for the developmental structure already established by PCA, differential expression, enrichment analysis, WGCNA, and cross-layer integration, rather than as an external predictive benchmark. The full performance comparison across feature spaces is summarized in Table 3 and visualized in Fig. 10.

Table 3.

Cross-track benchmark summary for machine-learning validation of compartment-specific developmental programs.

Compartment	Rank	Feature_space	Balanced_accuracy	Macro_F1	Accuracy
Oocyte	1	Selected eigengenes	0.914 ± 0.005	0.921 ± 0.004	0.921 ± 0.003
Oocyte	2	Integrated panel	0.892 ± 0.034	0.901 ± 0.030	0.906 ± 0.026
Oocyte	3	All eigengenes	0.808 ± 0.034	0.817 ± 0.034	0.838 ± 0.027
Oocyte	4	Selected hubs	0.795 ± 0.034	0.806 ± 0.031	0.815 ± 0.026
Oocyte	5	DE panel	0.775 ± 0.025	0.770 ± 0.031	0.792 ± 0.024
Granulosa	1	DE panel	0.839 ± 0.036	0.858 ± 0.033	0.911 ± 0.021
Granulosa	2	Selected eigengenes	0.695 ± 0.031	0.774 ± 0.058	0.837 ± 0.017
Granulosa	3	Integrated panel	0.661 ± 0.026	0.776 ± 0.040	0.820 ± 0.014
Granulosa	4	Selected hubs	0.659 ± 0.025	0.730 ± 0.047	0.813 ± 0.014
Granulosa	5	All eigengenes	0.581 ± 0.025	0.660 ± 0.076	0.753 ± 0.015

Fig. 10.

Feature-space performance comparison across compartments. Balanced accuracy across repeated stratified 5-fold cross-validation for Elastic Net multinomial logistic regression models trained on biologically informed feature spaces within each compartment. Panel A shows the oocyte models and panel B shows the granulosa models. In oocytes, the selected eigengene representation gave the strongest overall performance, whereas in granulosa cells the DE-derived panel performed best. These results indicate that oocyte stage structure is most effectively captured by a compact module-level summary, whereas granulosa stage discrimination is better recovered from a broader transcript-level feature set.

In the oocyte compartment, the strongest performance was obtained from the compact selected-eigengene feature space defined by the three biologically prioritized WGCNA modules. This representation achieved a mean balanced accuracy of 0.914, a mean macro F1 of 0.921, and a mean overall accuracy of 0.921, making it the best-performing oocyte feature space in Table 3 and the highest-ranked model in the oocyte panel of Fig. 10. The integrated panel remained strong but was modestly weaker overall, with a mean balanced accuracy of 0.892, whereas broader or more explicitly gene-level representations showed a clearer decline, including all eigengenes (0.808), selected hubs (0.795), and the DE-derived panel (0.775). These results indicate that oocyte stage structure is most effectively captured by a compact systems-level summary rather than by expanded marker sets.

The class-level behavior of the oocyte models was also biologically coherent. Under the best-performing selected-eigengene model, the antral stage was recovered perfectly, and the remaining errors were concentrated primarily between neighboring developmental states, particularly around the earlier part of the trajectory, as shown in the oocyte confusion matrix in Fig. 11. The per-class recall profile in Fig. 12 similarly showed strong recovery across stages, with the highest recall for antral oocytes and comparatively lower recall for primordial oocytes. Together, these patterns are consistent with the continuous progression already observed in the oocyte PCA and network analyses, and support the interpretation that oocyte maturation is organized as coordinated transcriptomic remodeling across a small number of dominant developmental programs.

Fig. 11.

Confusion matrices for the best-performing model in each compartment. Aggregated held-out confusion matrices across repeated stratified 5-fold cross-validation for the best-performing Elastic Net multinomial logistic regression model in each compartment. Panel A shows oocyte classification using the selected eigengene feature space, and panel B shows granulosa classification using the DE-derived panel. In oocytes, residual misclassification was concentrated primarily between neighboring developmental stages, consistent with a continuous maturation trajectory. In granulosa cells, late stages were strongly resolved, whereas earlier and intermediate stages showed greater overlap, indicating that the dominant late maturation program is more sharply separated than the immature and transitional states.

Fig. 12.

Stage-wise recall for the best-performing model in each compartment. Per-class recall across repeated stratified 5-fold cross-validation for the best-performing Elastic Net multinomial logistic regression model in each compartment. Panel A shows oocyte classification using the selected eigengene feature space, and panel B shows granulosa classification using the DE-derived panel. In oocytes, recall remained high across stages, with the strongest recovery of the antral state and comparatively lower recall for primordial oocytes. In granulosa cells, preovulatory and antral stages were recovered most strongly, whereas primordial and especially secondary stages remained more difficult, indicating that early and intermediate granulosa states are less sharply resolved than the dominant late maturation program.

In the granulosa compartment, the optimal representation differed from that observed in oocytes. Here, the DE-derived panel gave the highest performance, with a mean balanced accuracy of 0.839, a mean macro F1 of 0.858, and a mean overall accuracy of 0.911, making it the top-ranked granulosa model in Table 3 and the strongest-performing feature space in the granulosa panel of Fig. 10. This substantially exceeded the compact selected-eigengene model, which reached a mean balanced accuracy of 0.695, as well as the integrated panel (0.661), selected hubs (0.659), and all eigengenes (0.581). Thus, granulosa stage-discriminative information appears to be distributed across a broader transcript-level feature landscape and is not fully compressible into a small number of dominant eigengene summaries.

Performance of Elastic Net multinomial logistic regression models across biologically informed feature spaces within each compartment. Values are mean ± SD across repeated stratified 5-fold cross-validation. In oocytes, the selected eigengene representation gave the strongest overall performance, whereas in granulosa cells the highest performance was obtained from the DE-derived panel, indicating that the transcriptomic structure most informative for stage discrimination differs between the two compartments.

3.13. HSP90AA1 links an early oocyte WGCNA program to broader integrative evidence

WGCNA further contextualized the stage-dependent behavior of HSP90AA1 by placing it within the selected oocyte blue module, one of the early-state modules that showed a negative association with stage progression. Although HSP90AA1 was not among the highest-ranked compact hubs of this module, its module membership and negative gene significance for stage were consistent with the expression pattern observed in the differential expression analyses, in which transcript abundance was highest in early oocyte stages and declined toward the antral state. This agreement between gene-level expression dynamics and module-level organization supports the interpretation that HSP90AA1 forms part of an early oocyte program that is progressively attenuated during maturation rather than a late-stage activation signature.

This network placement is informative in the broader integrative context of the study. The oocyte blue module was one of the biologically prioritized modules retained in the cross-layer integration framework and contributed to the selected eigengene representation that gave the best performance in oocyte stage classification. Although HSP90AA1 itself was not retained as a direct compact feature in the finalized machine-learning benchmark panels, it remains indirectly connected to the ML findings through this selected module structure. In this sense, HSP90AA1 provides a gene-level example of how biologically meaningful developmental signal can be embedded within a broader co-expression program that is more informative at the module level than as an isolated predictive feature. Together with its distinct differential-expression behavior across the oocyte and granulosa compartments, these findings position HSP90AA1 as a useful marker of compartment-specific developmental remodeling, Fig. 13; while also illustrating how gene-level and module-level evidence converge within the integrative analysis. Fig. 14 indicates the overall layout of the entire study.

Fig. 13.

HSP90AA1 maps to the selected oocyte blue module and follows its early-state developmental pattern. (A) HSP90AA1 highlighted within the oocyte blue module according to module membership and stage association. (B) Distribution of the oocyte blue-module eigengene across follicular stages, showing progressive attenuation toward the antral stage. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 14.

Integrated overview of the dual-track analytical framework and major findings across human folliculogenesis.

(A) RNA-seq data from GSE107746, comprising 148 paired-end libraries, were processed through a unified workflow including quality control and trimming with fastp, transcript quantification with Salmon against GRCh38, and gene-level filtering and normalization with tximport and TMM, yielding 24,684 genes for downstream analysis. The final cohort was then split by compartment into 77 oocyte (OO) and 71 granulosa-cell (GC) libraries, establishing parallel germline and somatic analytical tracks. Oocyte track (germline).

(B) Stage structure. The oocyte track included primordial (n = 17), primary (n = 25), secondary (n = 12), and antral (n = 23) stages. (C) Exploratory structure (PCA). Oocyte transcriptomes formed a continuous developmental gradient, with PC1 explaining 18% and PC2 explaining 9.6% of the variance; adjacent stages overlapped, indicating progressive rather than sharply discrete transitions. (D) Differential expression. Oocytes showed 21,726 differentially expressed genes across contrasts, with large transcriptional shifts at each stage transition; the strongest contrast was antral versus primordial, involving 15,589 genes with a maximum log2 fold change of 12.90.

(E) Functional enrichment. Oocyte maturation was dominated by genome-regulatory programmes. Antral-high signatures included DNA recombination, chromosome segregation, and double-strand break repair, whereas primary-high signals included RNA splicing and mRNA processing, consistent with post-transcriptional and genome-regulatory remodeling across development.

(F) Co-expression (WGCNA). Oocyte networks showed a late-stage module coupled to inverse early modules. The turquoise module represented dominant late maturation at antral stage, while royalblue marked early biosynthetic/translational activity and blue captured early regulatory/structural states. Representative hub genes included COX6A1, RPS14, GPM6B, FGF13, and SATB1. (G) Machine-learning validation. Oocyte developmental structure was best recovered by eigengene-based representation, indicating that compact WGCNA eigengenes captured stage progression more effectively than broader feature sets. (H) Gene-level integration (HSP90AA1). HSP90AA1 showed progressive attenuation across oocyte development, decreasing from logCPM 11.53 to 10.59 from primordial to antral stages, and was assigned to the oocyte blue module. Granulosa track (somatic). (B) Stage structure. The granulosa track included primordial (n = 8), primary (n = 15), secondary (n = 6), antral (n = 24), and preovulatory (Preov; n = 18) stages. (C) Exploratory structure (PCA). Granulosa transcriptomes showed sharper late-stage resolution than oocytes, with PC1 explaining 15.2% and PC2 explaining 10.1% of the variance; antral and preovulatory samples were clearly separated from early stages. (D) Differential expression. Granulosa cells showed 21,738 differentially expressed genes across contrasts, with relatively weak early contrasts but strong late-stage shifts; the strongest contrast was preovulatory versus primordial, involving 17,488 genes with a maximum log2 fold change of 11.90. (E) Functional enrichment. Granulosa maturation was characterized by progressive bioenergetic activation. Antral-high signatures included mitochondrial translation, oxidative phosphorylation, and ATP synthesis, whereas preovulatory-high signatures included respiratory-chain assembly and late metabolic specialization. (F) Co-expression (WGCNA). Granulosa networks also displayed a late-stage module with inverse early-stage modules. The turquoise module represented dominant late maturation at preovulatory stage, while magenta and orangered4 marked early-stage decline. In contrast to oocytes, hub-gene composition was distinct, supporting shared systems-level organization but different molecular drivers. (G) Machine-learning validation. Granulosa stage discrimination was best resolved by the differential-expression feature set, indicating that broader DE-derived representations outperformed eigengene summaries for this compartment. (H) Gene-level integration (HSP90AA1). HSP90AA1 showed a sharp secondary-stage peak in granulosa cells, reaching logCPM 14.18, followed by marked decline to 9.34 in the preovulatory stage. (I) Synthesis. Together, the dual-track analysis supports coordinated but non-equivalent transcriptomic remodeling across human folliculogenesis, with a shared systems-level architecture consisting of late-stage activation and inverse early modules, but compartment-specific molecular execution: genome-regulatory remodeling in oocytes versus bioenergetic activation in granulosa cells. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

4. Discussion

Our re-analysis supports a hierarchical view of human folliculogenesis in which cellular compartment identity precedes developmental-stage resolution. The strong global separation between oocyte and granulosa samples is consistent with the original human transcriptome study of folliculogenesis, which showed that oocytes and their paired granulosa cells follow cell-type-specific, stage-dependent transcriptional programs and engage in reciprocal signaling across follicle growth. More recent single-cell and single-follicle studies likewise emphasize that the adult ovary contains substantial germline and somatic heterogeneity, and that this complexity must be preserved rather than collapsed when interpreting follicular development.8, 15

Within that compartmental framework, however, the two lineages did not remodel in equivalent ways. The oocyte track in our data is most consistent with a progressive developmental continuum, with adjacent transitions already showing substantial differential expression but with the overall configuration remaining relatively smooth rather than sharply discontinuous. This interpretation aligns well with recent syntheses of human oogenesis/folliculogenesis transcriptomics, which describe continuous transcriptome dynamics from primordial to antral stages, alongside a more limited subset of genes peaking around the secondary stage and enriched for kinase-regulatory and signaling functions. Together, these observations suggest that oocyte maturation is not simply a series of binary switches, but a cumulative reweighting of transcriptional priorities across successive follicular stages.18, 31

By contrast, the granulosa compartment in our analysis shows a more uneven and stage-concentrated pattern of change, with relatively modest early contrasts but much stronger remodeling in later stages, especially around the transition toward antral and preovulatory states. This is biologically plausible because granulosa cells must progressively acquire specialized endocrine, structural, and signaling functions as follicles mature. Recent ovarian single-cell and spatial studies support this interpretation by showing that later peri-ovulatory somatic states are associated with increasingly distinct functional programs, including extracellular matrix remodeling, angiogenic signaling, hormone production, and cell-state diversification. Although species differences must be kept in mind, these data are consistent with our finding that granulosa transcriptional identity becomes especially dynamic in late folliculogenesis.8, 21, 28

The behavior of HSP90AA1 provides a useful gene-level example of this broader compartment-specific logic. In our dataset, HSP90AA1 was strongly stage-associated in both tracks, but its developmental profile differed markedly between them. In oocytes, expression declined gradually from early to late stages, consistent with a broad stage-associated trend rather than a sharp pairwise switch. In granulosa cells, however, HSP90AA1 rose strongly in the secondary stage and then fell across antral and preovulatory stages, indicating much more abrupt remodeling. Direct human functional data for HSP90AA1 in folliculogenesis remain limited, but recent experimental work supports the biological plausibility of this pattern. In bovine oocytes, HSP90AA1 inhibition impairs nuclear and cytoplasmic maturation, spindle integrity, epigenetic regulation, and early embryo development, indicating that HSP90AA1 can contribute to oocyte developmental competence. In ovarian granulosa-cell systems, HSP90α-linked regulation has also been connected to steroidogenic balance through androgen receptor–dependent mechanisms, suggesting that HSP90-family biology may be particularly relevant to somatic follicular function. These observations fit well with our results, in which HSP90AA1 behaves as a shared stage-associated gene but shows its strongest dynamic regulation in the granulosa lineage (29, 39; B.¹⁶).

Our results argue against viewing folliculogenesis as a single linear transcriptomic trajectory. Instead, they support a dual-track model in which oocyte and granulosa compartments retain distinct regulatory identities while undergoing coordinated, but not identical, forms of developmental remodeling. In this framework, the oocyte lineage appears comparatively continuous, whereas the granulosa lineage exhibits stronger late-stage transcriptional reorganization. This distinction is important for downstream interpretation because signals that appear moderate in one compartment may be central in the other, and because candidate genes such as HSP90AA1 may have shared stage association but lineage-specific regulatory meaning. More broadly, our findings reinforce the need for pathway and network analyses that preserve compartment context when inferring the molecular architecture of follicular development.

Recent human ovarian transcriptomic atlases and syntheses increasingly argue that folliculogenesis is not a single linear programme shared equally across follicular compartments, but a layered developmental process in which oocytes and granulosa cells retain distinct molecular priorities as follicles mature.¹⁸ Our enrichment results fit that view closely. Rather than revealing one common maturation signature, they point to two coordinated but non-equivalent trajectories: an oocyte-centred programme dominated by genome-regulatory and RNA-processing functions, and a granulosa-centred programme dominated by progressive mitochondrial and energetic activation. This interpretation is broadly consistent with the foundational human folliculogenesis atlas and with newer integrative overviews of stage-dependent ovarian gene expression. Emerging studies demonstrate that therapeutic interventions can modulate apoptosis-related gene networks, including caspase-mediated pathways, highlighting the broader plasticity of gene regulatory systems in cellular stress and disease contexts.³³

In the oocyte lineage, the broad primordial-to-antral transition was enriched for DNA recombination, chromosome segregation, DNA replication and repair-associated functions, whereas the primary-to-secondary transition was more strongly linked to RNA splicing, mRNA processing and related post-transcriptional processes. Read together, these findings suggest that oocyte maturation is organized across at least two regulatory layers: an earlier phase in which RNA handling and translational preparedness are extensively reconfigured, followed by a broader late-stage reinforcement of genome maintenance and chromosome-regulatory capacity. That interpretation aligns well with recent work emphasizing the exceptional requirement for genome protection in mammalian oocytes and the growing recognition that RNA-based regulation is central to female germ-cell competence and follicular development.19, 38

The granulosa results followed a different logic. Metabolic activation was already evident at the antral stage, where enrichment was dominated by mitochondrial translation, mitochondrial gene expression, respiration, oxidative phosphorylation and ATP-generating processes, and this signal intensified further in preovulatory granulosa. This pattern strongly suggests that late granulosa maturation is coupled to increasing energetic demand and to progressive mitochondrial specialization. That interpretation is well supported by recent human and translational literature showing that granulosa-cell metabolism changes markedly around ovulation, is linked to oocyte competence, and is increasingly viewed as a central component of late follicular function rather than a secondary accompaniment to it.23, 26, 30

The reciprocal granulosa signal is equally informative. In the antral-to-preovulatory comparison, genes relatively higher in antral granulosa were enriched for DNA replication, chromosome segregation, nuclear division, checkpoint functions and replication-associated repair, whereas preovulatory granulosa was dominated by respiratory and mitochondrial pathways. This suggests that the final transition into preovulatory identity is accompanied by a functional reorientation away from a more replication-associated antral state and toward a metabolically specialized somatic programme. In biological terms, the data support a model in which granulosa cells do not simply accumulate more transcriptional change as ovulation approaches; instead, they shift their regulatory emphasis from proliferative or chromosome-dynamics-associated functions toward the energetic infrastructure needed for terminal follicular maturation. Recent views of the human ovulatory process are consistent with this view, describing extensive peri-ovulatory granulosa remodeling rather than a simple extension of earlier growth programmes.23, 30

A further strength of the enrichment step is that it clarifies which patterns deserve mechanistic emphasis and which do not. Several weak opposite-direction granulosa contrasts were dominated by receptor-family and olfactory-style annotation categories. These terms were retained in the statistical record, but they were not treated as central biological drivers because they were sparse, repetitive and disconnected from the much more coherent mitochondrial, respiratory, RNA-processing and chromosome-regulatory themes recovered across the stronger contrasts. That distinction matters because it shifts the discussion away from simply listing significant terms and toward identifying the developmental programmes that recur consistently enough to organize follicular maturation. On that basis, the present results support a dual-track model in which oocyte development is shaped by genome-regulatory and post-transcriptional remodeling, whereas granulosa maturation is defined by progressive bioenergetic activation and by relative withdrawal from replication-associated functions. This is precisely the kind of compartment-specific structure that now justifies the next analytical step, namely track-resolved co-expression analysis by WGCNA, to determine whether these enrichment-defined programmes are embedded in robust stage-associated transcriptional modules.¹⁸

The WGCNA results strengthen the central interpretation of this study by showing that follicular progression is organized not simply as a list of differentially expressed genes, but as a set of coordinated developmental programs that emerge, intensify, and recede across maturation. That broader view is consistent with recent transcriptomic work showing that oocytes and their surrounding somatic cells follow related but increasingly specialized developmental trajectories, while intercellular communication becomes more active as follicles grow. Recent profiling of human cortical follicles has also emphasized that early follicle states are heterogeneous and not fully captured by morphology alone, which makes network-level analysis particularly valuable for resolving biologically meaningful structure within stage-defined cohorts.5, 7

Within the oocyte compartment, the dominant turquoise module supports the idea that late maturation reflects broad transcriptome-wide consolidation rather than isolated marker switching. This interpretation fits well with recent literature emphasizing that developmental competence in oocytes depends heavily on post-transcriptional control, including poly(A)-tail regulation, translational timing, and orderly maternal mRNA clearance during maturation and the oocyte-to-embryo transition.22, 27 In that context, the appearance of candidates such as BTG4 in the late-stage module is notable, because recent work has again underscored the importance of BTG4/CNOT-centered deadenylation control for maternal RNA clearance and oocyte quality in both mouse and human systems. Our enrichment of chromosome-associated, DNA replication, recombination, and repair terms in late oocytes likewise fits recent evidence that weakened DNA damage responses in oocytes are linked to chromosome mis-segregation and aneuploidy, and with a recent human ovarian multi-omics atlas showing coordinated regulatory rewiring across ovarian cell types during reproductive aging.32, 37, 40

The inverse oocyte modules are just as informative biologically. Rather than treating the primordial and primary states as incompletely matured versions of the antral state, the royalblue and blue modules indicate that early oocyte identity is maintained by organized biosynthetic, translational, structural, and regulatory programs that are progressively attenuated as maturation proceeds. That reading is compatible with recent work arguing that maternal RNA storage, selective translation, and RNA surveillance are not peripheral features of oogenesis, but core determinants of developmental readiness.⁴⁰ In this light, our inverse modules likely capture distinct layers of early oocyte architecture that must be reduced or reconfigured as the cell moves toward a late-stage state characterized by stronger genome-regulatory and chromosome-management demands.

The granulosa network points to a related but clearly different maturation logic. Here, the dominant turquoise module aligns with antral and especially preovulatory progression, and its integration with enrichment strongly supports a late-stage program centered on oxidative phosphorylation, respiration, mitochondrial gene expression, and ATP-generating metabolism. That result is consistent with recent work positioning granulosa cells as the metabolic and signaling support system of the follicle, and with newer reviews linking granulosa dysfunction, oxidative imbalance, and mitochondrial impairment to declining oocyte support capacity and reproductive aging.25, 34 At the same time, recent human granulosa data suggest that increased free-radical generation under FSH stimulation does not automatically imply overt oxidative injury, which is important for interpretation: the late granulosa turquoise module is better viewed as a coordinated energetic specialization program than as a generic stress response.

The inverse granulosa modules add further depth. The magenta module appears to capture a broader early somatic-state program, whereas orangered4 likely reflects a more selective regulatory or adhesion-linked layer that is especially relevant around the antral-to-preovulatory transition. That distinction fits well with current views of human ovulation, which emphasize that periovulatory remodeling in women involves coordinated but highly context-specific changes in signaling mediators, transcription factors, steroids, prostaglandins, proteolytic systems, and other local regulators, and that human ovulatory mechanisms cannot be assumed to be identical to those of model organisms. In that sense, our granulosa inverse modules may represent somatic architectures that are progressively displaced as the follicle approaches a more specialized periovulatory state.³⁴

One of the most important conclusions from the WGCNA analysis is that parallel developmental logic does not imply shared molecular leadership. Both compartments contain one dominant late-stage module and one or more inverse early-state modules, but the hub-level overlap between oocyte and granulosa networks is minimal. That pattern agrees with recent follicle transcriptome studies showing that oocyte–somatic coordination is real and developmentally consequential, yet still built from lineage-specific expression programs and signaling relationships. Our results therefore argue that the follicle behaves as a coordinated developmental unit while preserving strong compartmental specialization in the actual genes anchoring maturation. This is precisely why the track-wise design of the present analysis is so important: a pooled network would have blurred the very distinction that appears to be biologically central.³⁴

These points also help explain why integration with differential expression and enrichment was so informative. In both compartments, the selected WGCNA modules showed strong direction-consistent overlap with the key developmental contrasts, indicating that the network results are not abstract correlation structures imposed on the data after the fact. Rather, they appear to compress the same underlying remodeling detected by gene-wise analyses into coherent developmental programs and to highlight candidate hub genes that may anchor the transition between immature and late-stage follicular states.

A few limitations remain important. These networks were inferred within a single cohort and therefore support a strong within-study developmental model, but they do not by themselves establish universal modules across all human ovaries or prove causality for any individual hub gene. In addition, some late and inverse programs, especially in granulosa cells, likely represent composite biology that spans more than one functional axis. Even with those caveats, the present results provide a coherent compartment-aware account of follicular maturation: oocyte development is dominated by broad genome-regulatory and post-transcriptional reorganization, whereas granulosa maturation is shaped by late bioenergetic specialization together with progressive loss of earlier somatic-state architecture. Seen together, these co-expression programs help explain how tightly coupled, yet molecularly distinct, lineages advance through folliculogenesis.

To place representative stage-associated genes within the broader developmental architecture identified in this study, we integrated track-specific expression patterns, differential expression results, WGCNA module context, and the finalized machine-learning feature spaces. This framework was motivated by recent human ovarian atlases showing that follicular development is organized across distinct yet interacting cellular compartments, with substantial spatial and transcriptional heterogeneity among oocytes and granulosa subtypes.24, 36 In oocytes, gene-level interpretation was guided by the broader biology of maturation, which depends on regulated use of stockpiled maternal transcripts and late chromatin reorganization as developmental competence is established.20, 35 In granulosa cells, interpretation was anchored to evidence that peri-ovulatory granulosa physiology is accompanied by dynamic metabolic remodeling that correlates with oocyte competence and embryo developmental potential.³⁰ Within this integrative setting, genes of interest were mapped to the annotated WGCNA outputs to determine module assignment, module membership, and stage association relative to the selected manuscript-prioritized modules, and were then cross-referenced against the finalized ML panels to determine whether their contribution was represented directly as an explicit predictor or indirectly through eigengene structure (B.¹⁶). This approach was especially informative for HSP90AA1, because experimental work has already linked HSP90AA1 activity to oocyte maturation and early embryonic development, supporting its interpretation as a biologically meaningful marker rather than an isolated statistical signal.

5. Conclusion

This integrative dual-track re-analysis of human folliculogenesis demonstrates that oocytes and granulosa cells progress through development in a coordinated overall framework, yet do so through clearly distinct transcriptomic and regulatory trajectories. Compartment identity was the dominant axis of variation across the dataset, confirming that follicular maturation cannot be interpreted as a single uniform process. Within this structure, oocytes exhibited broader and more continuous stage-linked transcriptional remodeling, whereas granulosa cells showed comparatively modest early-stage shifts followed by stronger late-stage reorganization. Functional enrichment, co-expression architecture, and predictive modeling converged on the same central pattern: although both compartments undergo orderly developmental progression, they engage different biological programmes and differ in how strongly stage information is encoded in their transcriptomes. Oocyte maturation was more strongly associated with regulatory and developmental remodeling, while granulosa maturation was more closely linked to metabolic and mitochondrial activation. The behavior of HSP90AA1 further illustrated this compartment-specific divergence, supporting its relevance within the broader developmental framework identified here. Collectively, these findings refine the transcriptomic map of human folliculogenesis by showing that developmental coordination and regulatory divergence coexist across follicular stages. This compartment-aware view provides a stronger basis for future mechanistic studies of ovarian biology and may help guide biomarker discovery and translational investigations related to female reproductive function.

Availability of data and materials

All the data is presented in the manuscript. The raw data obtained from publicly available databases, the GSE107746 is available on SRA website; and the respective information are provided in the materials and methods section. All the bioinformatics packages used are publicly available.

Clinical trial number

Not applicable.

CRediT authorship contribution statement

Jing Wang: Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Formal analysis, Data curation, Conceptualization. Zhimin Xu: Writing – original draft, Visualization, Resources, Project administration, Methodology, Investigation, Data curation, Conceptualization. Jun Ning: Writing – review & editing, Visualization, Validation, Supervision, Software, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization.

Consent for publication

N/A.

Ethics approval and consent to participate

N/A.

Funding

National Natural Science Foundation of China Youth Science Fund Project (Grants No. 62306124); and Project of Science and Technology Development Jilin Province (202302033063SF).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.