Cell‐free transcriptomic profiles and mechanism insights in female androgenetic alopecia

Dear Editor, Our study presents a novel predictive machine learning model that demonstrates the potential of plasma cell-free RNA (cfRNA) for diagnosing and prognosing female androgenetic alopecia (FAGA). We identified cell-free DNAJB9 as significantly associated with FAGA through bioinformatic analysis and machine learning followed by RT-qPCR validation (Figure 1A). FAGA manifests heterogeneously,1 often as diffuse thinning of the crown and frontal scalp.2 It's pathogenesis critically involves androgen-hair follicle interactions and WNT and JAK-STAT signalling.3 The cfRNA in bodily fluids have shown diagnostic/prognostic potential for various diseases.4 Machine learning is increasingly used to analyse complex cfRNA data.5 However, the potential association between cfRNA and FAGA remains unclear. For subsequent analyses comparing disease severity, we focused on the ‘upper’ group as patients in the top 25% of the FAGA-Index (scores > 5.53) and the ‘lower’ group as those in the bottom 25% (scores < 1.92). Blood test results (Figure S1 and Table S2) showed no significant differences in various haematological and biochemical indicators between the FAGA and control groups. However, testosterone exhibited a significantly lower level in the ‘upper’ group (Figure S5), supporting that the FAGA-Index effectively enhances the stratification of patients by severity and may facilitate identification of other potential biomarkers in FAGA progression. Greater variation in principal component analysis (PCA) of cfRNA expression profiles between upper and lower FAGA subgroups, compared to that between FAGA and control groups, also suggested increased heterogeneity or molecular diversity within FAGA subtypes (Figure 1C and D). The RNA biotypes were categorised based on Ensembl classifications with minor adjustments (Figure 1E). Analysis of differentially expressed genes (DEGs) showed that CYTB, RNY1, and TMSB4X were notably upregulated in FAGA patients, whilst EEF1A1 was significantly downregulated (Figure 2A; Table S5). Furthermore, genes including ND2, ATP6, ND6, and PARLP1 exhibited significant expression changes across varying disease severities (Table S7), suggesting their potential association with FAGA progression (Figure 2B). Functional enrichment analysis of these DEGs implicated pathways related to sensory perception, nuclear division, chromosome segregation, and mitosis in FAGA (Figure 2C and D; Tables S6 and S8). Pathway activity analysis reinforced the potential importance of JAK-STAT and WNTs pathways in FAGA (Figure 3A–C; Tables S17–S19). A comparison of transcription factor (TF) activities between FAGA patients and controls (Figure 3D; Table S9) revealed significantly increased activity of NCOA3 and MAX. However, no significant differences in TF activities were observed between upper- and lower-FAGA subgroups (Table S10). Crucially, we observed a significant negative correlation between cell-free DNAJB9 expression and the FAGA-Index, suggesting that low DNAJB9 expression may be associated with increased FAGA severity (Table S11). Protein–protein interaction network (PPI) analysis further revealed that in FAGA versus controls, upregulated genes drive endocrine compensation and mitochondrial stress responses, while downregulated genes suppress growth signalling (MET/mTORC1) and RNA metabolism (Figure S2). Comparing upper- versus lower-FAGA, there is increased mitochondrial/endocrine activity alongside impaired translation and calcium homeostasis (Figure S3). These results indicate worsening pathway dysregulation with FAGA progression (Tables S12–S15). Subsequently, we developed a predictive model for FAGA using machine learning, employing the GeneLLM downstream classification framework,6 a state-of-the-art method for cfRNA classification. After initial feature extraction, deep feature mining was conducted to uncover latent patterns within gene expression profiles indicative of FAGA. The cfRNA RPKM matrix was partitioned into training (40%), validation (40%), and testing (40%) sets (Figure 4A). The large, completely held-out test set offers a rigorous internal validation of the model's performance on unseen data from the same population. To ensure robustness, hyperparameter optimisation was conducted using 10-fold cross-validation. The model achieved an Area Under the Curve (AUC) of .707 for distinguishing FAGA patients from controls (Figure 4B) and .714 for separating high versus low FAGA-Index scores (Figure 4C). It identified several genes associated with FAGA, including VGLL3, CYP1A1, antisense to PDE7B, and notably DNAJB9 (Figure 4D). Features correlated with FAGA severity, such as the long non-coding RNA ARL14EP-DT and pseudogene TJAP1P1, were also highlighted (Figure 4E). We performed correlation analysis between the expression levels of the six candidate biomarkers and various blood parameters (Figure S6; Table S20). Nevertheless, the correlation coefficient indicated only weak to moderate associations, implying that more validation is needed between biochemical and molecular diagnosis. Further RT-qPCR validation in both internal and external cohorts confirmed that cell-free VGLL3, antisense to PDE7B, and DNAJB9 were significantly downregulated in the FAGA patients, while lncRNA ARL14EP-DT and pseudogene TJAP1P1 were significantly downregulated in the upper FAGA subgroup (Figure S4). Notably, DNAJB9 is a DNAJ/HSP40 heat shock protein essential for cellular stress responses.7 HSP40 family proteins are implicated in regulating androgen receptor (AR) activity, often maintaining AR in an inactive state.8, 9 Reduced DNAJB9 expression may potentially disrupt AR signalling in hair follicles, especially under stress. In summary, as the first investigation integrating cfRNA bioinformatic analysis with machine learning, this study establishes a crucial proof-of-concept for the utility of cfRNA in FAGA diagnosis and prognosis, addressing a critical gap in the field and providing a solid foundation for future work. Our exploratory model, whilst moderate in its predictive power, proved effective in highlighting cell-free DNAJB9 as a FAGA biomarker and a candidate for therapeutic intervention, meriting further investigation. Project supervision was overseen by H.J. Y.L. S.D., Y.J., H.J., and Y.L. were responsible for conceptualisation. L.J., M.L., and Y.L. managed clinical recruitment design. Y.J. designed and conducted the wet-lab experiments. Data analysis and interpretation were performed by S.D. and Z.A. The primary manuscript writing and revisions were undertaken by L.J., M.L., S.D., and Y.J. Coordination of human samples and data collection was managed by Y.Z., L.J., H.J., M.L., Y.L., Z.L., C.Z., and R.L. All authors contributed to discussions on the results and provided feedback on the manuscript. This research was funded by Ministry of Education of People's Republic of China (Grant 231103242232720 to Yufei Li), the East Hospital Affiliated to Tongji University Introduced Talent Research Startup Fund (Grant DFRC2019008 to Hua Jiang), the Clinical Research Plan (Grant numbers: SEHHH-2021(KJ)-0669-KJB-485 and 2023(KJ)-0143-KYB-111 to Yufei Li), the Shenzhen Science and Technology Program (Grant 20240724152335001 to Yongcheng Jin), and the Featured Clinical Discipline Project of Shanghai Pudong Fund (Grant PWYts2021-07 to Hua Jiang). No competing interests to disclose. All data preprocessing and downstream analyses were conducted using standard bioinformatics tools on a Linux CentOS 8-based High-Performance Computing (HPC) system and R version 4.1.3, as provided by OxTium Technology Co. Ltd. Details of the tools, including their names, versions, and specific usage, are outlined in the Methods section. Unless otherwise specified, default parameters were employed for all tools. All individuals in this study were thoroughly informed about the objectives, procedure, and possible risks, and written informed consent was obtained before participation. Furthermore, the study protocol received ethical review and approval from the Ethics Review Committee of Shanghai East Hospital (EC.D(BG).016.02.1). All the cfRNA sequencing data have been submitted to the NCBI SRA database, accessed under the BioProject accession PRJNA1146172. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.