Single-cell RNA sequencing data processing and quality control

Single-cell RNA sequencing (scRNA-seq) data were obtained from the Gene Expression Omnibus (GEO) database (accession number GSE153760) [17], comprising 8 atopic dermatitis (AD) donors and 7 healthy control donors (15 donors total) and processed in R (v4.4.1) using the Seurat package (v5.2.1) [18]. All samples were sequenced using the 10x Genomics Chromium platform on an Illumina HiSeq 3000 instrument. Prior to quality control, the dataset comprised approximately 15,000 cells in total. For each cell, quality-control metrics were computed, including the number of detected genes, the proportion of mitochondrial transcripts (genes matching the pattern prefix “MT-”), and the aggregate expression of hemoglobin genes (HBA1, HBA2, HBB, HBD, HBE1, HBG1, HBG2, HBM, HBQ1, HBZ) to flag potential erythrocyte contamination or cellular stress. Cells were retained if they contained 300–6,000 detected genes and fewer than 20% mitochondrial reads, resulting in 47,147 cells (AD: 28,768 cells; HC: 18,379 cells) retained for downstream analysis. Following quality-control filtering, expression matrices were normalized using Seurat’s LogNormalize method with a scale factor of 5,000 to account for differences in sequencing depth across cells. To mitigate potential batch effects across samples, we applied Harmony integration. Post-integration UMAP embeddings confirmed well-mixed sample distributions across major cell-type compartments, supporting the robustness of the data integration. Processed expression matrices, metadata, and dimensionality reduction embeddings are publicly available at GitHub: https://github.com/Vesemir-CN/AD-and-HC-single-cell-machine-learning-and-MD-dataset.

Dimensionality reduction, integration, and clustering

Highly variable genes were identified using the variance-stabilizing transformation (VST) method, retaining the top 5,000 variable features for downstream analyses. Principal component analysis (PCA) was performed on scaled data. To mitigate batch effects across samples, Harmony integration was applied using orig.ident as the batch key. Clustering was performed on the Harmony-corrected space using FindNeighbors (dims = 1:30) and FindClusters (resolution = 0.5). Uniform Manifold Approximation and Projection (UMAP) was used for two-dimensional visualization based on the same 30 Harmony-corrected principal components.

Cell type annotation and subset analysis

Cell clusters were annotated based on canonical marker genes and literature-curated cell-type signatures. For keratinocyte-focused analyses, keratinocyte cells were isolated and reanalyzed using the same preprocessing and clustering workflow, yielding 14 keratinocyte subclusters (Kcs1–Kcs14).

Differential expression and pathway analysis

Differential expression analysis was conducted using the Wilcoxon rank-sum test implemented in the presto package. Genes meeting both log2FC > 0.5 and adjusted P < 0.05 were considered statistically significantly differentially expressed. For pathway-level interpretation, gene set enrichment analysis (GSEA) was performed using clusterProfiler with the MSigDB Hallmark (H) gene set collection. Pathways with FDR < 0.05 were considered significant, and enrichment magnitude were summarized using normalized enrichment scores (NES).

Cell type prioritization using Augur

To prioritize cell types most responsive to CA2-associated stratification, Augur was applied as a cell-type prioritization framework [19]. Cells were stratified into CA2-high and CA2-low groups based on a median split of normalized CA2 expression. Within each annotated cell type, Augur trained random forest classifiers to quantify the separability between CA2-high and CA2-low cells, reporting cross-validated area under the ROC curve (AUC). Cell types were ranked by AUC, with higher values indicating greater transcriptional separability under CA2 stratification.

Machine learning-based consensus gene selection

To ensure biological independence and avoid pseudoreplication, single-cell expression profiles were aggregated into donor-level metacells using the hdWGCNA package (k and target metacell numbers auto-calculated based on group sizes), such that each metacell represents a transcriptionally coherent profile from a single donor. All machine learning analyses were performed on these metacells. We implemented a nested cross-validation framework via the easyScml pipeline (https://github.com/Vesemir-CN/easyscml). In the outer loop, donor-stratified 10-fold cross-validation (10 repeats) ensured that all metacells from the same donor were assigned exclusively to either the training or test set within each fold, preventing donor-level data leakage. The inner loop was reserved for hyperparameter optimization. Twenty machine learning algorithms were trained, and all 190 pairwise combinations (excluding self-pairing) were evaluated [20,21]; the combination heatmap diagonal displays single-algorithm performance for reference. Combinations achieving mean cross-validated AUC ≥ 0.8 were designated elite, and genes receiving votes from ≥ 50% of elite combinations were retained as consensus biomarkers. SHAP analysis was performed using an XGBoost classifier trained on the full metacell dataset to quantify directional feature contributions: positive SHAP values indicate contribution toward AD prediction; negative values toward HC. Point color in the beeswarm plot reflects normalized expression (blue: low; red: high). Visualizations were generated with ggplot2 (v3.4.0).

Statistical analysis and visualization

All statistical analyses were performed in R, with multiple-testing correction applied using the Benjamini–Hochberg false discovery rate procedure where applicable. Visualization was performed using ggplot2, pheatmap, and custom plotting functions. UMAP projections were generated using Seurat, and pathway results were visualized using clusterProfiler functions.

Cell state distance analysis using scDist

To quantify transcriptomic distances between CA2-stratified states while accounting for sample-level variation, scDist was applied to the keratinocyte subset [22]. CA2-high versus CA2-low status (newgroup) was modeled as a fixed effect, and sample identity (orig.ident) was included as a random effect. Keratinocyte subclusters (Kcs1–Kcs14) were used as the cell-state labels. Distances were computed in a reduced-dimensional space (d = 20 principal components), yielding a quantitative estimate of CA2-associated perturbation magnitude across keratinocyte subpopulations.

scHSC single-cell fine clustering

Single-cell Hierarchical Spectral Clustering (scHSC) was used as an additional unsupervised fine-clustering strategy [23]. Briefly, scHSC constructs a k-nearest neighbor (k-NN) graph from normalized expression profiles, performs spectral embedding, and applies hierarchical clustering in the embedded space. The optimal number of clusters was determined by the silhouette coefficient. Analyses were performed in R (v4.4.2) using the scHSC package with default parameters unless otherwise specified.

Molecular docking analysis

Molecular docking was conducted to evaluate the binding feasibility and predicted interaction mode of baicalin with carbonic anhydrase II (CA2) [24]. The CA2 crystal structure was retrieved from the RCSB Protein Data Bank (PDB ID: 1LUG) and prepared by removing crystallographic water molecules and non-essential heteroatoms. The baicalin structure was obtained from PubChem and converted to a 3D conformer with energy minimization in Chem3D (MM2 force field), followed by format conversion using AutoDockTools (v1.5.7). Docking was performed with AutoDock Vina using a grid box (48.8 × 46.7 × 56.9 Å) centered on the CA2 active site. Key parameters were set as follows: exhaustiveness = 10, energy_range = 4 kcal/mol, and n_modes = 10. The top-ranked binding pose was selected for subsequent simulation, and protein–ligand interactions were inspected in MOE.

Molecular dynamics simulations

Molecular dynamics (MD) simulations were performed using GROMACS (v2025.3) to assess the conformational stability of the baicalin–CA2 complex in explicit solvent. The AMBER14SB force field was applied to the protein, and the system was solvated with TIP3P water molecules in a periodic boundary condition simulation box. Ligand topology and force field parameters were generated using Sobtop (v1.0-dev5). The system was charge-neutralized and supplemented with 150 mM NaCl to mimic physiological ionic conditions. Following energy minimization, the system was equilibrated sequentially under NVT and NPT ensembles, followed by a 100 ns production run with a 2 fs integration time step. Long-range electrostatics interactions were treated using the particle mesh Ewald (PME) method, and bonds involving hydrogen atoms were constrained using the LINCS algorithm. Production trajectories were analyzed to compute root mean square deviation (RMSD), radius of gyration (Rg), root mean square fluctuation (RMSF), solvent-accessible surface area (SASA), and the number of protein–ligand hydrogen bonds.

Binding free energy estimation (MM-PBSA)

Binding free energy of the baicalin–CA2 complex was estimated using the molecular mechanics Poisson–Boltzmann surface area (MM-PBSA) method applied to snapshots extracted from MD trajectories [25]. The total binding free energy was decomposed into molecular mechanics terms (van der Waals and electrostatic interactions) and solvation free energy terms (polar and non-polar components). Per-residue energy decomposition was additionally performed to identify residues contributing most strongly to the estimated binding free energy.

Free energy landscape (FEL) analysis

To characterize the dominant conformational states sampled during MD simulation, free energy landscapes (FELs) were constructed using backbone RMSD and radius of gyration (Rg) as reaction coordinates [26]. Free energy surfaces were derived from the Boltzmann-weighted probability distributions of these coordinates to visualize major low-energy basins and conformational preferences of the baicalin–CA2 complex.

A quality-controlled, integrated single-cell atlas reveals separable cell communities in AD and control skin

To provide a robust single-cell reference for downstream target prioritization and cell-population localization (Fig 1), we constructed an integrated AD–HC skin atlas through harmonized preprocessing, integration, and unsupervised clustering. The resulting atlas, visualized as a circular embedding in which each arc segment represents a major cell-type compartment, displayed well-separated cellular communities with distinct boundaries across nine annotated lineages, supporting adequate global resolvability of major skin cell populations (Fig 2A). With this atlas structure in place, we next evaluated cross-group comparability within the same reference space. AD and HC cells occupied the same major cellular compartments on the shared embedding, indicating that subsequent AD–HC contrasts can be performed under a consistent cellular framework rather than being confounded by missing lineages in one group (Fig 2B). At the level of tissue cellular ecology, cell-type composition differed markedly between AD and HC: keratinocytes accounted for a substantially larger proportion in AD relative to HC, whereas immune cell populations—including T cells and dendritic cells—showed comparatively higher representation in HC, consistent with the epidermal hyperplasia and immune infiltration patterns characteristic of AD-associated tissue remodeling (Fig 2C). Annotation reliability was then assessed using canonical marker expression patterns. Representative lineage markers exhibited cell-type–specific enrichment across clusters—exemplified by epidermal keratinocyte markers (e.g., KRT15, DMKN), immune markers for T cells (e.g., CD3D, TRBC1), and myeloid/mast-cell–associated markers (e.g., CD1C; TPSAB1/CPA3)—supporting reliable downstream cell-type labeling based on canonical marker concordance (Fig 2D). Finally, we examined whether group-associated transcriptional differences were cell-type interpretable, rather than reflecting diffuse noise. Comparing AD to HC within each annotated cluster, distinct cell-type–specific differential expression patterns emerged consistent with expected biological roles: keratinocyte clusters exhibited upregulation of structural epidermal genes (e.g., KRT15, DMKN) alongside downregulation of differentiation-associated transcripts, while immune compartments—including T cells, dendritic cells, and mast cells—displayed upregulation of markers associated with activation and inflammatory state (e.g., CD3G, TRBC1, TPSAB1). These cell-type–interpretable transcriptional differences, rather than diffuse noise, provide a structured computational framework for subsequent CA2-associated feature prioritization and stratification analyses (Fig 2E).

Machine learning‑based consensus screening and SHAP interpretation prioritize a panel of robust AD‑associated genes

To prioritize candidate biomarkers with high cross-validated discriminative performance, we benchmarked 20 diverse machine learning algorithms on donor-level metacells under a nested, donor-stratified cross-validation framework. Single-algorithm evaluation revealed substantial performance heterogeneity across methods, with a subset of algorithms exceeding the predefined elite threshold of AUC = 0.8 (Fig 3A). To capture complementary algorithmic strengths, we systematically constructed all 190 pairwise algorithm combinations and assessed their combinatorial performance. Scatter plot visualization of pairwise AUCs demonstrated that high-performing combinations clustered distinctly from lower-performing pairs (Fig 3B), and density distribution analysis confirmed that the majority of elite combinations were concentrated above the AUC = 0.8 threshold, with the median combination AUC exceeding that of most individual algorithms (Fig 3C). Dumbbell plot comparison further illustrated that participation in elite combinations consistently elevated mean AUC relative to single-algorithm baselines for most methods (Fig 3D), while algorithm participation frequency analysis identified a core set of algorithms that disproportionately contributed to elite combinations, suggesting their superior generalizability in this dataset (Fig 3E). The full combination performance landscape was summarized in a pairwise AUC heatmap, which revealed structured blocks of high-performing algorithm pairs and confirmed the robustness of the combinatorial evaluation strategy (Fig 3F). Gene-level voting across all elite combinations then identified six consensus biomarkers surpassing the 50% vote threshold, with CA2 receiving the highest vote frequency among all candidates (Fig 3G). To interpret the predictive contribution of each consensus gene, SHAP analysis was performed using an XGBoost classifier; global feature importance ranking confirmed CA2 as the dominant predictive feature by mean absolute SHAP value (Fig 3H), and the beeswarm summary plot demonstrated that elevated CA2 expression was consistently associated with positive SHAP values, indicating a robust directional contribution toward AD prediction across individual metacells (Fig 3I). Finally, expression distribution analysis across AD and HC groups validated that all six consensus genes exhibited statistically significant differential expression, with CA2 showing the most pronounced and consistent upregulation in AD (Fig 3J; Wilcoxon rank-sum test, p < 0.001).

CA2 stratification prioritizes keratinocytes as the most transcriptionally separable cell type and associates with broad pathway enrichment signals

Building on the nomination of CA2 and its disease-associated expression support (Fig 3), we next asked which major cell types exhibit the most substantial transcriptomic separability under CA2 stratification (CA2-high vs CA2-low). To quantify this in a cell-type–aware manner, we applied Augur, which measures within–cell type separability using an AUC-based prioritization score.

Projecting Augur AUC onto the integrated UMAP revealed marked spatial heterogeneity in separability across the atlas, with the strongest signal mapping to the keratinocyte-dominated region (Fig 4A). Consistently, cell type–level summaries showed that keratinocytes ranked highest in AUC, suggesting that keratinocytes exhibited the strongest CA2-stratified transcriptional separability among all major cell types, as ranked by Augur AUC (Fig 4B).

To contextualize this prioritization, we examined the distribution of CA2 expression across major cell types. A dot-plot overview showed that keratinocytes display the highest CA2 signal in terms of both mean expression and the fraction of expressing cells, consistent with their top-ranked separability under CA2-high vs CA2-low stratification (Fig 4C).

We then asked what pathway-level enrichment patterns accompany CA2 stratification at the pathway level. Gene set enrichment analyses identified coordinated shifts across multiple functional modules rather than an isolated pathway change. KEGG GSEA highlighted significant enrichment involving metabolic reprogramming, epithelial junction and barrier remodeling, and proliferative quiescence (Fig 4D) (S1 Table). In parallel, HALLMARK GSEA implicated broad programs spanning energy metabolism and signaling/inflammatory axes (Fig 4E) (S1 Table). Together, these results position keratinocytes as the most CA2-sensitive major cell type in terms of transcriptomic separability and suggest that CA2 stratification is associated with multi-program reconfiguration across curated pathway resources.

Subpopulation-resolved prioritization identifies Kcs9 as the most CA2-associated perturbed keratinocyte state

To localize CA2-associated transcriptional differences within the keratinocyte compartment, we constructed a keratinocyte sub-atlas comprising 14 subclusters (Kcs1–Kcs14). Visualization of the keratinocyte embedding split by group (AD vs HC) showed that major subclusters were represented across conditions, supporting adequate cross-group coverage for downstream CA2-based stratification analyses (Fig 5A). Marker-gene dot plots further summarized the transcriptional features distinguishing Kcs states, providing a reference map for interpreting where CA2-linked signals concentrate within the keratinocyte landscape (Fig 5B).

We first quantified CA2-stratified separability within each keratinocyte state using Augur. Projecting Augur AUC onto the keratinocyte UMAP revealed substantial heterogeneity across the sub-atlas, suggesting that separability between CA2-high vs CA2-low is subcluster-dependent rather than uniformly distributed across keratinocytes (Fig 5C). Consistently, subcluster-level AUC summaries ranked keratinocyte states by CA2-associated separability, highlighting a subset of high-sensitivity states and nominating Kcs9 among the top prioritized populations (Fig 5D).

To quantify perturbation magnitude in multivariate expression space, we next applied scDist. Mapping scDist distance onto the keratinocyte UMAP showed spatially concentrated regions of elevated distance, consistent with discrete subpopulation-associated shifts rather than diffuse cell-to-cell noise (Fig 5E). Faceted UMAPs further demonstrated that these distance “hotspots” localize to specific keratinocyte subclusters (Fig 5F). When summarized across Kcs1–Kcs14, scDist distances robustly ranked keratinocyte states by CA2-associated transcriptomic distance, with Kcs9 exhibiting the largest distance estimate, thereby nominating it as the most CA2-associated perturbed keratinocyte state (Fig 5G).

Finally, to interpret the biological programs accompanying CA2 stratification within the prioritized state, we performed pathway enrichment within Kcs9. HALLMARK GSEA showed enrichment signals across metabolic and signaling/stress-response programs, including modules related to oxidative phosphorylation and pathway-level signaling signatures (Fig 5H). In parallel, KEGG GSEA highlighted enrichment patterns involving cellular adhesion/junctional organization and multiple metabolism- and immune-related pathways, consistent with the notion that CA2 stratification in Kcs9 is associated with enrichment patterns spanning multiple gene-set modules across curated resources (Fig 5I) (S1 Table).

Structure-based docking and molecular dynamics provide computational support for binding feasibility of baicalin to CA2

To provide orthogonal, structure-based support for CA2 as a plausible baicalin-interacting candidate, we performed molecular docking followed by 100-ns molecular dynamics (MD) simulations and post hoc binding-energy analyses (Fig 6A–K). Docking yielded a coherent binding pose in which baicalin occupied a defined pocket on CA2 (Fig 6A). Local interaction inspection suggested a hydrogen-bond–supported polar contact network involving His3, His4, Gly63, and His64, with interaction distances within a typical hydrogen-bonding range (approximately 1.9–3.4 Å), consistent with a geometrically plausible docked pose (Fig 6B). We then assessed complex stability during MD simulation. RMSD trajectories indicated an initial equilibration phase followed by a plateau-like regime; the protein RMSD remained stable throughout the trajectory, fluctuating around ~0.17–0.18 nm, while the complex RMSD fluctuated around ~0.18–0.20 nm without a pronounced drift, whereas the ligand RMSD exhibited larger fluctuations throughout the trajectory, consistent with greater conformational flexibility of the free ligand relative to the bound protein (Fig 6C). In parallel, the radius of gyration (Rg) remained largely stable at approximately 1.75–1.80 nm throughout the 100-ns trajectory, with axis-resolved components (Rg/sX ~ 1.44 nm, Rg/sY ~ 1.49 nm, Rg/sZ ~ 1.35 nm) similarly showing minimal drift, collectively suggesting maintenance of global compactness of CA2 during the simulation trajectory (Fig 6D). Solvent accessible surface area (SASA) similarly showed no marked shift across the trajectory, fluctuating around a mean of approximately ~127 nm² within a range of ~121–133 nm², suggesting that no large-scale changes in protein surface exposure were observed during the simulation (Fig 6E). Hydrogen-bond analysis further showed persistent but dynamic protein–ligand hydrogen bonding over time (typically ~2–3 hydrogen bonds, ranging from 0 to 7), consistent with persistent yet dynamic hydrogen-bonding patterns during simulation (Fig 6F). RMSF remained low for most residues (predominantly <0.20 nm), with the exception of N-terminal residues (~residues 1–5) exhibiting the highest fluctuations (~0.54 nm) and C-terminal residues (~residues 220–265) showing moderately elevated fluctuations (peaking at ~0.30 nm), suggesting low overall residue-level fluctuation with localized flexibility at both terminal regions (Fig 6G). To quantify binding energetics, MM-PBSA yielded an estimated binding free energy for the baicalin–CA2 complex (ΔG_bind ≈ −22.082 kcal/mol), suggesting a potentially favorable interaction within the limitations of this approximate method, with van der Waals interactions (VDW ≈ −46 kcal/mol) representing the dominant favorable contribution, supplemented by electrostatic interactions (COU ≈ −15 kcal/mol), collectively offset by unfavorable polar solvation (PB ≈ +40 kcal/mol), while nonpolar solvation (SA ≈ −3 kcal/mol) contributed modestly favorably (Fig 6H). Per-residue free-energy decomposition further highlighted a subset of pocket residues with stronger favorable contributions, with Phe230 and Asn231 exhibiting the largest individual contributions, nominating these residues as computationally predicted interaction hotspots that may contribute disproportionately to the estimated binding free energy (Fig 6I). FEL analysis revealed a single dominant low-energy basin with a well-defined funnel-shaped profile (centered at Rg ~ 1.750 nm and RMSD ~0.05–0.10 nm, free energy ~0 kJ/mol), surrounded by high-energy regions reaching ~13–14 kJ/mol, suggesting the presence of a dominant low-energy conformational basin sampled during the trajectory (Fig 6J). Two-dimensional Gibbs free energy landscape projected onto PC1 and PC2 from principal component analysis (PCA) further confirmed a single predominant low-energy conformational state sampled during simulation, consistent with a stable and well-defined bound-state ensemble (Fig 6K). Collectively, these docking and MD results provide consistent in silico observations supporting the computational plausibility of baicalin binding to CA2 and motivate downstream experimental investigation of this predicted interaction.

Mechanistic interpretation: A coupled “state-network” response in keratinocytes

A useful way to interpret our findings is through a state-network hypothesis: in the inflammatory microenvironment of AD, keratinocytes may transition into adaptive states characterized by concurrent structural remodeling and metabolic/proteostatic stress. In this framework, enrichment of junctional/cytoskeletal and adhesion-related programs can be viewed as a response to barrier disruption and repeated injury–repair cycles, while concurrent shifts in oxidative phosphorylation, glycolysis, and protein-processing pathways are consistent with increased energetic demand and cellular stress management required to sustain these remodeling programs.

Within such a coupled network, CA2—an enzyme central to pH and bicarbonate/ion homeostasis—may either functionally contribute to, or serve as a transcriptional marker of, keratinocyte adaptation to inflammatory stress. Critically, independent experimental evidence supports the disease relevance of CA2 in AD epidermis: Kamsteeg et al. [27] demonstrated that CA2 mRNA and protein are significantly upregulated in lesional AD skin compared with both psoriatic lesions and healthy controls, with expression localized predominantly to suprabasal, differentiated keratinocytes. Importantly, this upregulation was shown to be driven by Th2 cytokines (IL-4 and IL-13) rather than Th1 stimulation, directly linking CA2 induction to the canonical AD inflammatory microenvironment rather than reflecting a cell-autonomous, genetically programmed difference. This experimentally established expression pattern is consistent with the CA2-associated transcriptomic signal identified in our single-cell analyses, and provides external biological plausibility for the computational nominations reported here. pH and ion handling are tightly linked to epidermal differentiation, stratum corneum enzymatic activity, and barrier function [28,29]; elevated skin surface pH has been reported in AD patients, and alkalinization of the stratum corneum has been shown to impair barrier homeostasis [27], further contextualizing CA2 dysregulation within the barrier–inflammation axis characteristic of AD. Thus, the co-enrichment of modules of metabolic reprogramming, epithelial junction and barrier remodeling, and proliferative quiescence observed in CA2-stratified keratinocytes is not simply a list of unrelated pathways, but a coherent signature of keratinocytes operating under combined barrier-repair demand and homeostatic pressure.

Notably, the prioritization of Kcs9 by both separability (Augur) and transcriptomic distance (scDist) suggests that this subpopulation represents a high-sensitivity keratinocyte state in which CA2 stratification aligns with coordinated remodeling and stress-adaptation programs. Conceptually, Kcs9-like cells may therefore constitute a ‘response-prone’ epidermal node where barrier-associated structural programs and homeostatic stress programs show concurrent enrichment signals, making them informative for disease stratification and mechanistic testing. The biological identity of Kcs9 is further informed by the convergence of its five cluster-defining marker genes—FOXC1, KRT23, FST, CDA, and FGF7—with its GSEA enrichment architecture. FST, a secreted BMP/Activin antagonist documented in suprabasal keratinocyte compartments of inflamed skin [30], aligns with positive enrichment of TIGHT_JUNCTION, ADHERENS_JUNCTION, NOTCH, and WNT_BETA_CATENIN signals, positioning Kcs9 within a barrier-competent suprabasal differentiation context; notably, NOTCH/WNT attenuation drives keratinocyte fate divergence toward inflammatory states, suggesting that their sustained enrichment in Kcs9 reflects active barrier maintenance rather than default differentiation. KRT23 co-expression further supports a spinous-layer identity consistent with KRT1/KRT10-anchored coordinates in single-cell atlases. FGF7-associated enrichment of PROTEIN_SECRETION, MTORC1_SIGNALING, and MAPK—alongside negative enrichment of CYTOKINE_CYTOKINE_RECEPTOR_INTERACTION—indicates growth factor–biased rather than cytokine-amplifying secretory output, mechanistically distinct from the LAMB3–CD44 and integrin-dominated pro-inflammatory KC communication programs in AD lesional skin. CDA-linked dual enrichment of OXIDATIVE_PHOSPHORYLATION across HALLMARK and KEGG collections, together with GLYCOLYSIS and negative PURINE_METABOLISM enrichment, parallels the multi-pathway metabolic reprogramming causally attributed to AD keratinocytes by Mendelian randomization, suggesting cell-autonomous metabolic adaptation rather than bystander dysregulation. FOXC1 co-occurrence with negative E2F_TARGETS, MYC_TARGETS, and DNA_REPLICATION enrichment indicates proliferative quiescence, distinguishing Kcs9 from PCLAF-high proliferating KC subpopulations in AD and MKI67 + /TOP2A+ cycling states in psoriasiform dermatitis [30]. This suprabasal identity is independently anchored by CA2 localization to suprabasal differentiated keratinocytes in AD epidermis [27], with carbonic anhydrase family causal relevance reinforced by Mendelian randomization evidence identifying CA4 as an independent AD risk gene. Jointly, these signals support a unified model in which Kcs9 represents a suprabasal, barrier-repair–oriented keratinocyte state deploying paracrine growth factor signaling and metabolic reprogramming as primary adaptive responses to AD inflammatory stress—rather than hyperproliferation or direct cytokine amplification—thereby generating specific, experimentally testable predictions regarding its functional role.

In this context, elevated CA2 could represent (i) a potential functional contributor to maintaining intracellular/extracellular acid–base balance during stress, (ii) a proxy for a broader keratinocyte stress–remodeling state, or (iii) both, depending on the specific biology and lineage position of the Kcs9-like population. Our docking/MD analyses are consistent with CA2 being not only a computationally inferred state-associated gene but also a potentially tractable in silico target for baicalin; however, the direction and magnitude of functional modulation, and whether modulation would affect barrier-repair programs or influence inflammatory signaling contexts, require targeted experimental validation.

Translational implications: A path from epidermal state to testable intervention

Our results suggest a translational path centered on epidermal accessibility, state-based stratification, and mechanistic complementarity to immune-targeted therapies—without implying immediate clinical efficacy. First, because the epidermis is directly accessible, an epidermal target such as CA2 and a keratinocyte state such as Kcs9 conceptually support exploration of topical or localized delivery strategies, which may reduce systemic exposure and enable direct modulation of barrier-adjacent programs. Second, CA2-high status or a Kcs9-like transcriptional signature could potentially serve as computationally derived hypothesis for a biomarker framework for patient stratification or pharmacodynamic readouts (e.g., identifying individuals whose disease is dominated by a keratinocyte stress–remodeling state), though this requires validation in independent cohorts and longitudinal sampling. Third, the keratinocyte-centered mechanism suggested by our computational findings may complement established immune-directed approaches by targeting the barrier–inflammation feedback loop from the epidermal side, which may be relevant to residual disease activity that can persist despite cytokine blockade in some patients.

The computational nominations presented here define a prioritized but unvalidated hypothesis requiring stepwise experimental corroboration before any mechanistic or therapeutic conclusions can be drawn. We propose a staged validation framework: first, CA2 expression and Kcs9-like signatures should be confirmed in independent AD cohorts and spatially localized via spatial transcriptomics or multiplex imaging. Second, direct baicalin–CA2 binding should be evaluated using orthogonal biophysical assays (SPR/MST/ITC) and cellular target engagement approaches (CETSA/DARTS), with CA2 enzymatic activity measured to establish mechanistic directionality. Third, barrier-relevant functional readouts should be assessed in Kcs9-enriched keratinocyte model systems. Finally, if in vitro engagement is confirmed, epidermal-compatible baicalin formulations should be evaluated in established AD animal models.

Generalizability of the framework and relevance of the baicalin–CA2–Kcs9 axis beyond AD

The computational framework presented here—integrating pharmacogenomic target mining, dual machine-learning feature selection, single-cell state–resolved perturbation prioritization, and structure-based feasibility assessment—is not inherently disease- or drug-specific, and is in principle applicable to the systematic identification of other molecule–target–cell state axes in AD. The same workflow could in principle be applied to other bioactive compounds reported in AD (e.g., quercetin, berberine) against the same scRNA-seq reference atlas to nominate additional molecule–target–cell state axes, or extended to non-keratinocyte compartments such as fibroblasts or mast cells, with relatively modest methodological modification.

Regarding cross-disease relevance, the baicalin–CA2–Kcs9 axis may warrant investigation in other skin diseases characterized by epidermal barrier disruption and keratinocyte dysregulation. Psoriasis is a particularly relevant context: CA2 expression has been reported in psoriatic lesional skin, albeit at lower levels than in AD [27], suggesting that CA2-associated keratinocyte states may reflect a shared epidermal stress–remodeling response modulated by inflammatory context. Other conditions involving barrier dysfunction—including ichthyoses and allergic contact dermatitis—represent additional candidates addressable using publicly available scRNA-seq datasets. We acknowledge that cross-disease validation remains an important direction for future work.