We selected three primary publicly available mouse cochlear transcriptomic resources spanning development and injury/repair: (i) a developmental FACS RNA-seq dataset (GSE60019) profiling Atoh1-EGFP+ hair cells and the surrounding EGFP− fraction at E16, P0, P4 and P7 (Scheffer et al., 2015); (ii) adult inner- and outer-hair-cell microarrays (GSE56866) capturing sensory-epithelial trophic signalling programmes (Liu et al., 2014); and (iii) an NIHL RNA-seq dataset following mesenchymal stromal cell (MSC) therapy (GSE187029) (Warnecke et al., 2021). To test generalisability beyond the datasets used for module construction, we additionally analysed: (iv) an independent purified spiral ganglion neuron (SGN) bulk RNA-seq dataset spanning development (GSE132925), and (v) two independent injury/degeneration resources—GeoMx spatial transcriptomics of spiral ganglion regions after noise exposure (GSE312646) and rat spiral ganglion RNA-seq after kanamycin-induced deafening (GSE194063). Because these datasets were generated on heterogeneous platforms (bulk RNA-seq, FACS RNA-seq, microarrays and GeoMx) and across species, we did not merge expression matrices across studies; instead, all analyses were performed within each dataset (and within GeoMx segments), and cross-study comparisons are presented as standardized scores and effect sizes rather than absolute quantitative equivalence. Figure 1 summarizes the study design.

Processed expression tables were downloaded from GEO and analysed separately for the EGFP+ (hair-cell) and EGFP− (non-hair-cell) fractions at E16, P0, P4 and P7. For descriptive gene-level plots and fold-change summaries, we used library-size-normalized expression (counts-per-million, CPM) and log2-transformed values. For module scoring, expression values were standardized within the developmental dataset (z-scored across stages within each fraction) to emphasize temporal trends independent of platform-specific scaling.

Normalized microarray expression values were downloaded from GEO. Probe sets were mapped to gene symbols using the annotation provided with the dataset; when multiple probe sets mapped to the same gene, values were collapsed by the median. Differential expression between inner hair cells (IHCs) and outer hair cells (OHCs) was assessed on log2 expression values. For module scoring, gene expression was standardized within the microarray dataset (z-scored across the seven samples).

For the NIHL RNA-seq study (Warnecke et al., 2021), we re-analysed the TMM-normalized gene count matrix provided as a GEO (Supplementary File) (GSE187029_msc_experiment_tmm_raw_counts.txt.gz), comprising whole-cochlea RNA from noise-exposed mice with or without MSC therapy (n = 3 per group). Genes with low expression were filtered (CPM > 1 in at least two samples). Differential expression (MSC vs. untreated NIHL) was re-computed using a negative binomial generalized linear model (log link) with a group indicator and a library-size offset (log total counts per sample). A common dispersion parameter (α) was estimated across expressed genes using a method-of-moments approach (median α = 0.067). Two-sided Wald p-values were adjusted using Benjamini–Hochberg FDR. These gene-level statistics were used for downstream module summaries and visualization (Figure 2).

Four coherent transcriptional modules were defined: (i) excitability/synaptic transmission, (ii) plasticity/neurite growth, (iii) trophic–metabolic support and (iv) injury/inflammation. Seed genes were nominated from cochlear and SGN literature and then expanded/verified using functional enrichment results (Section 2.6). To quantify module activity within each dataset, we used a simple per-sample gene-set scoring approach conceptually related to GSVA and GSEA (Hänzelmann et al., 2013; Subramanian et al., 2005): gene expression was normalized and log2-transformed as appropriate for each platform, standardized per gene within the dataset (z-score across samples), and module scores were computed as the mean z-score across genes in the module. Because datasets were generated on heterogeneous platforms and in different contexts, we do not merge expression matrices across studies; cross-study comparisons are therefore presented as within-dataset standardized module scores and standardized effect sizes (e.g., Cohen’s d) rather than absolute quantitative equivalence. For external validation plots in purified SGNs (GSE132925), we additionally report raw module expression as the median log2(CPM + 1) across module genes; median- versus mean-based aggregation yielded highly correlated CNFC trajectories (Supplementary Table S1). The composite CNFC score was defined as Excitability + Trophic/Metabolic − Injury/Inflammation; plasticity was tracked as an independent axis because it can reflect both developmental consolidation and injury-driven regrowth programmes (see Tables 1, 2).

Gene Ontology (GO) and KEGG pathway enrichment analyses were performed using g:Profiler with g:SCS multiple-testing correction (Raudvere et al., 2019). Enrichment queries used the Mus musculus gene universe and were applied to differentially expressed genes in each dataset to provide an independent functional check on the curated module definitions and to transparently document pathway support for module membership. GO and KEGG annotations were referenced from the Gene Ontology resource and KEGG pathway databases, respectively (The Gene Ontology Consortium, 2021; Kanehisa and Goto, 2000).

To estimate the changing cellular composition of the EGFP− fraction across development, we computed three cell-type signature scores (neuronal, supporting-cell and immune) as the mean z-scored expression of marker genes curated from published cochlear atlases. Signature scores were compared across E16, P0, P4 and P7 and related to the excitability module to test whether excitability trends could be explained by changing neuronal enrichment.

To facilitate reuse across platforms and studies, we defined a minimal 24-gene panel by selecting highly referenced, functionally representative genes with consistent detection across datasets (6 excitability, 5 plasticity, 7 trophic/metabolic and 6 injury/inflammation genes; Table 3). Minimal-panel module scores and CNFC were computed using the same procedures as the full modules. Agreement between full and minimal CNFC scores was assessed by correlation across all samples (Figure 3B).

To benchmark CNFC against an unbiased transcriptional summary, we performed principal-component analysis (PCA) on log2(CPM + 1) expression for purified SGNs (GSE132925) after filtering low-variance genes. We tested whether CNFC and individual module scores aligned with the dominant transcriptomic maturity axis (PC1) by correlation analysis (Figure 4). To address concerns about equal module weighting, we also estimated data-driven weights by multivariate regression predicting PC1 from standardized module scores and compared weighted versus unweighted CNFC (Supplementary Table S1).

In the EGFP− fraction of the developing cochlea (GSE60019), excitability and synaptic transmission genes increased from embryonic to early postnatal stages, consistent with maturation of auditory neuronal properties. Representative examples included ion channels (e.g., Kcnc1, Scn8a), glutamate receptor components (e.g., Gria1) and synaptic machinery (e.g., Snap25; Figure 5). Importantly, all developmental comparisons were performed on normalized expression (CPM, log2-transformed) rather than raw counts to account for sequencing depth.

To test whether these excitability trends could reflect changing cell composition within the EGFP− fraction, we scored neuronal, supporting-cell and immune marker signatures. Neuronal signature scores increased in parallel with the excitability module (Figures 6A,B), suggesting that both transcriptional maturation and increasing neuronal enrichment contribute to the observed excitability trajectory.

In the EGFP+ hair-cell fraction, genes associated with actin remodelling and structural dynamics (e.g., Cfl1/Cfl2) declined across early postnatal development (Figure 7). While Cfl1/Cfl2 are representative actin regulators, we interpret this pattern conservatively as consolidation of developmental growth/remodelling programmes rather than a direct proxy for regenerative potential, which can involve distinct injury-induced gene programmes.

In adult hair-cell microarrays (GSE56866), inner and outer hair cells displayed distinct expression of trophic ligands, extracellular matrix regulators and metabolic mediators (Figure 8), supporting the idea that the sensory epithelium provides an active microenvironment shaping neural competence.

We re-analysed the NIHL + MSC cohort (GSE187029) from the TMM-normalized count matrix using a negative binomial GLM (Methods 2.4). Relative to untreated noise-exposed cochleae, MSC-treated cochleae showed strong down-regulation of the excitability module (mean log2FC = −2.59 across 12 mapped genes; directional Stouffer gene-set p = 6.8 × 10–128; Supplementary Table S1). This included large shifts in receptor and synaptic genes such as Gabra6 (log2FC = −7.07, FDR = 1.7 × 10–39), Grin2a (log2FC = −5.13, FDR = 1.8 × 10–19), Grin2c (log2FC = −3.76, FDR = 2.4 × 10–5), Gria1 (log2FC = −1.75, FDR = 6.3 × 10–20), Slc17a7 (log2FC = −2.08, FDR = 2.6 × 10–3), and synaptic release machinery (e.g., Syt1 log2FC = −3.38, FDR = 8.4 × 10–37; Snap25 log2FC = −2.69, FDR = 1.2 × 10–17) (Figure 2). Plasticity and trophic/metabolic gene sets also shifted downward with smaller magnitude (plasticity mean log2FC = −0.83; p = 8.1 × 10–9; trophic/metabolic mean log2FC = −0.58; p = 1.4 × 10–3; Supplementary Table S1). Because this dataset is whole-cochlea RNA, these signatures could reflect either within-neuron transcriptional regulation or cell-composition shifts; moreover, transcriptional suppression of excitability could be adaptive (dampening excitotoxic drive) or detrimental (reduced functional capacity). We therefore interpret directionality cautiously and frame these results as evidence that repair-oriented interventions can reshape programme-level signatures beyond simple survival metrics.

CNFC modules were constructed from literature-supported seed genes and then cross-checked/expanded using GO and KEGG enrichment of dataset-specific differentially expressed genes (Supplementary Table S1). This two-step process prioritizes interpretability while retaining an explicit link to unbiased functional annotation, and is intended to provide a hypothesis-generating scaffold rather than an exhaustive discovery of novel pathways.

Across the developmental FACS and adult microarray datasets, the minimal 24-gene panel closely tracked full-module CNFC scores (Pearson r = 0.83 within FACS; r = 0.86 within microarrays; overall r = 0.80; Figure 3B), supporting cross-platform portability. To assess generalisability beyond the primary datasets, we applied the minimal panel to an independent bulk RNA-seq dataset of purified mouse SGNs profiling five developmental ages (E15.5, P1, P8, P14 and P30) with glial (P8) and hair-cell (P12) reference samples (GSE132925) (Li et al., 2020). In purified SGNs, excitability increased with age, trophic/metabolic support remained high and injury/inflammation scores remained low, yielding an increasing CNFC trajectory from embryonic to mature stages (Figure 9). Glial and hair-cell reference samples showed substantially lower excitability and lower CNFC values than age-matched SGNs, supporting neuronal specificity of the competence trajectory.

Principal-component analysis (PCA) of purified SGNs (GSE132925) on transcriptome-wide log2(CPM + 1) expression identified a dominant maturity axis (PC1 explained 45.3% variance). CNFC correlated strongly with PC1 across SGN samples (Pearson r = 0.94; Figure 4B), supporting that the curated CNFC modules capture a major unbiased transcriptional trajectory. To address potential non-equivalence of module weighting, we estimated data-driven coefficients by regressing PC1 on standardized module scores. The resulting weights emphasised excitability and trophic/metabolic support and down-weighted plasticity (Figure 4C). A weighted CNFC remained highly correlated with the unweighted composite (Pearson r = 0.97; Supplementary Table S1), supporting use of the simple interpretable formulation.

To test whether CNFC behaves meaningfully in independent injury/degeneration contexts (beyond the NIHL/MSC cohort and beyond developmental maturation), we analysed two additional public datasets not used for module construction: GeoMx spatial transcriptomics of spiral ganglion regions after noise exposure (GSE312646) and rat spiral ganglion RNA-seq after kanamycin-induced deafening (GSE194063) (Methods 2.4.1–2.4.2; Figure 10; Supplementary Table S1). In GeoMx neuronal ROIs, noise exposure significantly reduced excitability scores (Δ = −0.55 z units; p = 7.8 × 10–4) and lowered CNFC (Δ = −1.36; p = 2.8 × 10–3), with a trend toward higher injury/inflammation scores (Δ = +0.43; p = 0.051) despite limited gene-panel overlap. Plasticity scores also declined modestly (Δ = −0.32; p = 0.019). In the rat spiral ganglion dataset, hearing status (vs deafened) was associated with higher excitability (β = +0.79; p = 0.005) and trophic/metabolic support (β = +0.62; p = 6.2 × 10–4), lower injury/inflammation (β = −1.09; p = 0.0015), and a robust increase in CNFC (β = +2.51; p = 1.8 × 10–5) after controlling for age. Gene-level inspection revealed a heterogeneous plasticity response: growth-associated regulators (Hmga2, Gap43) were induced in deafened spiral ganglia, whereas several neurite/cytoskeletal genes (Dcx, Stmn2, Tubb3) were reduced (Figure 10C), supporting partial and non-uniform re-engagement of plasticity programmes after injury.

A long-standing clinical intuition is that cochlear implant performance depends on how many auditory neurons survive after deafness. However, the relationship between spiral ganglion survival and speech recognition is variable and often modest (Cheng and Svirsky, 2021), and prediction of speech outcomes remains difficult even when multiple clinical factors are considered (Bartholomew et al., 2024). Our integrative analysis supports the concept that a functional competence state—encompassing excitability, synaptic signalling and plasticity programmes—may be an important latent variable. Under this view, neurons (and the surrounding cochlear microenvironment) can remain structurally present while functionally compromised in their ability to encode temporally precise information and to adapt to chronic electrical stimulation.

Auditory perception requires rapid, phase-locked firing and high-rate spiking. Developmental acquisition of these properties is known to involve fast potassium channels, voltage-gated sodium channels, and coordinated synaptic maturation (Boulet et al., 2016; Wang et al., 1998; Adamson et al., 2002; Hossain et al., 2005; Browne et al., 2017). Within the EGFP- fraction, we observed developmental increases in glutamatergic receptor genes such as Gria1 and in a subset of fast-spiking channel genes such as Kcnc1 (Figure 7). Although the EGFP- fraction is heterogeneous and cannot be equated with purified SGNs, these signatures are consistent with maturation of a neural transcriptome component within the developing cochlea. In the NIHL dataset, MSC therapy suppressed multiple excitability-related genes (including Grin2a/Grin2c, Kcnc1 and Scn8a) (Figure 2), suggesting that post-injury interventions may act, at least in part, by rebalancing excitability and excitotoxicity pathways implicated in synaptopathy (Kujawa and Liberman, 2009; Lin et al., 2011; Liberman and Kujawa, 2017).

A major barrier to improving CI outcomes is the physical distance between electrode contacts and target neural processes (Finley et al., 2008; Holden et al., 2016). In principle, neurite regeneration toward the electrode array could tighten electrode–neuron coupling and reduce stimulation thresholds. Our developmental analysis highlights a robust decline of actin-remodelling and growth-associated programmes (e.g., Cfl1/Cfl2 and Hmga2), consistent with a postnatal consolidation of structural plasticity. Importantly, developmental down-regulation does not imply that all plasticity-linked genes are uniformly silent after injury: in the rat kanamycin-deafening dataset (GSE194063) we observed selective induction of growth-associated regulators (Hmga2, Gap43) alongside suppression of several neurite/cytoskeletal genes (Dcx, Stmn2, Tubb3) (Figure 10C; Supplementary Table S1). This mixed pattern is consistent with an attempted but constrained regenerative response, and reinforces the rationale for interventions that both re-activate growth programmes and preserve excitability/synaptic competence—such as neurotrophin delivery or modulation of growth-cone signalling (Ramekers et al., 2012; Wan et al., 2014; Suzuki et al., 2016; Wang and Green, 2011).

Neurons do not exist in isolation: the sensory epithelium and adjacent supporting cells provide neurotrophic ligands, extracellular matrix scaffolds, and metabolic buffering that shape neuronal resilience (Ramekers et al., 2012; Nayagam et al., 2011; Reijntjes and Pyott, 2016). The adult IHC/OHC microarray analysis emphasised IHC-biased expression of Ntf3 and Gdnf (Figure 8; Liu et al., 2014), aligning with the critical role of hair-cell-derived neurotrophin signalling in maintaining type I SGN innervation and in promoting synaptic repair after injury (Ramekers et al., 2012; Wan et al., 2014; Suzuki et al., 2016; Wang and Green, 2011). These observations motivate a microenvironment-centric view of cochlear neural health, where preserving (or replacing) sensory-epithelial trophic support may protect functional competence even when hair-cell transduction is lost.

This study has several important limitations. First, the developmental dataset is based on FACS-separated hair cells and a heterogeneous EGFP− fraction rather than purified SGNs; consequently, EGFP− trajectories likely reflect both transcriptional maturation and changing cellular composition. We partially addressed this with cell-type signature scoring (Figure 6), but single-cell or nucleus-resolved approaches will be needed for a more direct SGN subtype-specific framework.

Second, CNFC modules are intentionally curated and mechanistically interpretable rather than derived solely from unsupervised clustering. We adopted a curated approach because our primary goal is a portable scoring vocabulary that can be applied across heterogeneous cochlear studies and platforms; in small or mixed-cell datasets, unsupervised modules (e.g., WGCNA) can be sensitive to sample size, cell-type composition and technical covariates and may not map one-to-one across studies or species. Accordingly, CNFC should be viewed as a hypothesis-generating scaffold. To reduce subjectivity, we provide transparent gene lists (Supplementary Table S1), enrichment support, an unbiased transcriptome-wide benchmark (PCA in purified SGNs; Figure 4), and sensitivity analyses for weighting and scoring.

Third, because datasets were generated on different platforms and in different contexts, we avoid direct pooling or aggressive batch correction and rely on within-dataset standardized scoring; absolute CNFC values should therefore not be compared across studies without harmonised processing. Finally, while we demonstrate CNFC behaviour across multiple injury/degeneration contexts (NIHL ± MSC, GeoMx NIHL, and kanamycin deafening), the present work remains based on animal transcriptomes and does not include patient-level CI outcomes. Future studies should integrate larger SGN-resolved injury/regeneration datasets and test whether CNFC-like scores derived from clinically accessible biospecimens (e.g., perilymph) prospectively associate with CI performance.

Our analysis is intended to generate testable, mechanistically anchored hypotheses rather than to over-interpret public datasets. Within that spirit, the CNFC framework suggests several concrete predictions that can be evaluated experimentally or in clinical cohorts:

Recent perilymph biomarker studies and systematic reviews support the feasibility of sampling cochlear fluids and linking molecular states to CI outcome variability (Wichova et al., 2021; Wanniarachchi et al., 2025). By providing an explicit module-based vocabulary and an initial minimal gene panel, our framework aims to accelerate the design of such validation studies and to connect cellular neuroscience mechanisms to clinically actionable neuroprosthetic hearing phenotypes.