5-Hydroxymethylcytosine profiles in circulating cell-free DNA serve as potential biomarkers for diagnosis and classification of adenomyosis

The datasets supporting the conclusions of this article are included within the article and its additional files. All other datasets used and analysed during the current study are available from the corresponding author upon reasonable request. The raw sequence data reported in this paper have been deposited in the Genome Sequence Archive (GSA) of the BIG Data Center, Beijing Institute of Genomics, Chinese Academy of Sciences under accession number HRA013666 ( https://ngdc.cncb.ac.cn/gsa-human ). Data are available from the date of publication without restriction. Source code is available at GitHub ( https://github.com/lzhang328/adenomyosis-5hmC-analysis ) and will be archived via Zenodo.

All participants provided written informed consent, and the study protocol was approved by the Peking University Third Hospital Medical Science Research Ethics Committee (approval no. M2021682). The study was conducted according to the guidelines of the Helsinki Declaration. In vitro validation employed the immortalised human endometrial epithelial cell line CTCC-008-0014 (Meisen, China). Mouse work was approved by the Peking University Animal Ethics Committee (BCJI0251) and complied with relevant ethical guidelines. The work was conducted according to ARRIVE guidelines. This study employed three independent, non-overlapping cohorts for distinct analyses ( Supplementary Table S1 ). For the 5hmC sequencing study, we prospectively collected plasma samples from 51 patients with histologically confirmed adenomyosis who underwent hysterectomy or lesion resection at Peking University Third Hospital between January 2020 and December 2021; 26 patients had matched adenomyotic tissue available. Plasma controls comprised archived samples from 46 healthy women collected at Peking University between 2016 and 2019, matched for age (±3 years) and menstrual phase. This temporal discrepancy was necessitated by the absence of healthy tissue biopsies, as healthy women do not undergo hysterectomy. Tissue controls consisted of normal myometrial specimens from 21 patients who underwent hysterectomy for benign conditions other than adenomyosis (leiomyoma, uterine prolapse) during 2020–2021. No participant contributed samples to multiple groups within this cohort. For the APP immunohistochemistry study, we selected archival formalin-fixed paraffin-embedded sections from an independent cohort of 38 adenomyosis patients (21 extrinsic, 17 intrinsic) who underwent surgery between 2018 and 2021, alongside control sections of normal myometrium with eutopic endometrium from 21 contemporaneous non-adenomyotic hysterectomy specimens; an additional 20 adenomyotic samples were used to evaluate migration-associated APP phenotypes. All sections were reviewed independently by two experienced gynaecological pathologists. For HOXD9 and 5hmC tissue profiling, adenomyotic lesions and adjacent normal myometrium containing ectopic endometrial glands were selected from 21 intrinsic-type, 6 extrinsic-type adenomyosis patients, and 100 non-adenomyotic benign controls, distinct from those employed in the aforementioned cohorts. Postmenopausal women and those with gynaecological malignancies were excluded from all cohorts. A hybrid retrospective-prospective hybrid follow-up was conducted between January 2021 and March 2021. Medical records were systematically reviewed for visit history, laboratory and imaging results, drug type, dose and duration, surgical modality, adjuvant therapy and post-operative course. Subsequently, all patients were contacted by telephone using a standardised questionnaire. Data captured included general health, menstrual pattern (dysmenorrhoea VAS 0–10, menstrual flow change), evolution of symptoms from menarche to diagnosis and after intervention, marital status, infertility history, assisted-reproductive-technology use and self-reported treatment satisfaction. The primary endpoint was disease recurrence (dysmenorrhoea VAS ≥4 after prior improvement ≥3 months or de novo ectopic lesions on imaging). Secondary endpoints were time-to-pregnancy and need for repeat surgery. Loss to follow-up was defined as failure to contact after three telephone attempts. The study was approved by the Peking University Third Hospital Ethics Committee (No. 2021-189-02). The aim of this study was to evaluate 5hmC as a diagnostic marker for adenomyosis and to distinguish molecular subtypes among patients. In total, 144 samples were included in the cohort analysis: 51 from patients with adenomyosis, 46 from healthy individuals, 26 adenomyotic tissue samples, and 21 normal myometrial tissues from patients who underwent hysterectomy for benign conditions other than adenomyosis. All 5hmC libraries were sequenced on the Illumina NextSeq 500 platform. To ensure that candidate plasma biomarkers captured the epigenetic signature of the adenomyotic lesion itself, we adopted a paired-tissue-driven—rather than random—splitting strategy: the training set consisted of 26 adenomyosis plasmas with matched adenomyotic tissues plus 23 healthy-control plasmas, whereas the validation set comprised all remaining plasmas. Firstly, we identified candidate differential 5hmC markers between patients with adenomyosis and healthy individuals in the training cohort. Secondly, using 26 matched tissue samples, we defined tissue-specific 5hmC signatures. These markers were then integrated to construct a diagnostic model, which was subsequently validated in the independent validation cohort. Finally, we performed molecular subtyping of patients with adenomyosis on the basis of 5hmC profiles derived from lesion tissue and identified the APP as a key regulatory gene. We established an APP-overexpressing cell model and carried out RNA sequencing, which revealed significant up-regulation of the transcription factor HOXD9 . Functional assays indicated that APP overexpression is associated with enhanced cell migration via increased HOXD9 expression, thus revealing a potential 5hmC–APP– HOXD9 regulatory axis that may be implicated in adenomyosis progression ( Fig. 1 a). Fig. 1 Research overview. (a).Whole-genome 5hmC sequencing was performed on 144 clinical samples, including plasma and tissue from patients with adenomyosis and matched controls. Differentially hydroxymethylated regions were used to establish a diagnostic classifier and to identify APP as a disease-specific marker. Functional follow-up in an APP-over-expressing cellular model showed up-regulation of the transcription factor HOXD9 and enhanced cell migration, suggesting that the 5hmC–APP–HOXD9 axis may drive adenomyosis progression. Research overview. (a).Whole-genome 5hmC sequencing was performed on 144 clinical samples, including plasma and tissue from patients with adenomyosis and matched controls. Differentially hydroxymethylated regions were used to establish a diagnostic classifier and to identify APP as a disease-specific marker. Functional follow-up in an APP-over-expressing cellular model showed up-regulation of the transcription factor HOXD9 and enhanced cell migration, suggesting that the 5hmC–APP–HOXD9 axis may drive adenomyosis progression. Five millilitres of peripheral blood was collected from each participant in a BD Vacutainer® EDTA tube for cfDNA extraction (Becton, Dickinson and Company, Cat# 367525). Blood samples were transported to the laboratory within 24 h and centrifuged at 1350 g for 12 min to isolate plasma. The upper plasma layer was transferred to 2 mL centrifuge tubes (AXYGEN, McT-200-C) and centrifuged at 1350 g for 12 min to remove residual leukocytes. The upper plasma layer was transferred to a new 2 mL centrifuge tube (AXYGEN, McT-200-C) and centrifuged again at 13500 g for 5 min to remove residual red blood cell fragments. The clarified plasma was immediately stored at −80 °C.Plasma cfDNA was extracted using the Quick-cfDNA Serum & Plasma Kit (ZYMO, Cat# D4076) and quantified with a Qubit 3.0 fluorometer (Thermo, Cat# Q33216 ). Extracted cfDNA was stored at −80 °C prior to library preparation. Agarose gel electrophoresis was used to verify the fragment size (∼180 bp) before construction of the 5hmC library. Tissue samples from patients, including adenomyosis lesions and normal myometrium tissue samples, were stored at −80 °C after surgical resection. After thawing, 25 mg of tissue was diced and collected. Genomic DNA was isolated from tissues using the quick-DNATM Miniprep Plus Kit (ZYMO, Cat# D4069) and quantified by Qubit 3.0 (Thermo, Cat# Q33216 ). Samples were then stored at −80 °C. Prior to library preparation, agarose gel electrophoresis was used to confirm the fragment size (∼1000 bp), after which the 5hmC library was prepared. Library construction used a chemically selective labelling method, 41 in which bacteriophage T4 β-glucosyltransferase was used to transfer engineered glucose fragments containing azide groups to hydroxyl groups of 5hmC in the human genome. Biotin was then used to chemically modify the azide group, enriching 5hmC-containing DNA fragments and enabling efficient capture of hydroxymethylated sites. First, according to the requirements of second-generation sequencing, Qubit 3.0 (Thermo, Cat# Q33216 ) was used for accurate quantification: 1–10 ng of cfDNA and 1–50 ng of genomic DNA (gDNA) were used, with gDNA fragmented by enzymatic reaction. The KAPA Hyper Prep Kit (KAPA, Cat# KK8514) was then used for end repair. The DNA was subsequently ligated with adaptors compatible with Illumina (Purkary, Cat# PKR2015, PKR2016, PKR2017, and PKR2018). The ligated DNA was reacted in a 25 μL solution containing 50 mM HEPES buffer (pH 8.0), 25 mM MgCl 2 , 100 μM UDP6-N3-GLC and 1 μM T4 β-GT (NEB, Cat# M0357L) at 37 °C for 2 h. The DNA was then purified using the DNA Clean & Concentrator™-5 (ZYMO, Cat# D4014). The purified DNA was fully mixed with 1 μL DBCO-PEG4-biotin (Click Chemistry Tools, 4.5 mM stock in DMSO) and reacted at 37 °C for 2 h. Similarly, the DNA was purified using the DNA Clean & Concentrator™-5 (ZYMO, Cat# D4014). Meanwhile, 2.5 μL of Thermo Life Technologies streptavidin beads (Thermo, Cat# 65305) was added directly to the mixture in 1 × binding buffer (5 mM Tris pH 7.5, 0.5 mM EDTA, 1 M NaCl, 0.2% Tween 20), followed by incubation at room temperature for 30 min with gentle rotation. Finally, the beads were rinsed 8 times for 5 min each with Buffers 1–4. The beads were then resuspended in RNase-free water (Tarara, Cat# 9012) for 14–16 PCR amplification cycles. The amplified products were purified using Pure Beads (KAPA, Cat# KK8001). The library concentration was measured with Qubit 3.0 (Thermo, Cat# Q33216 ). Paired-end 38 bp high-throughput sequencing was performed on the NextSeq 500 platform. Paired-end 38 bp high-throughput sequencing was performed on the NextSeq 500 platform. FASTQC (version 0.11.5) was used to assess the sequence quality. Raw reads were aligned to the human genome (version hg19) with bowtie2 (version 2.2.9) 42 and further filtered with Samtools (version 1.3.1) 43 (parameters used: Samtools view -f 2 –F 1548 -q 30 and Samtools rmdup) to retain unique non-duplicate matches to the genome. Pair-end reads were extended and converted into bedgraph format normalised to the total number of aligned reads using bedtools (version 2.19.1), 44 and then converted to bigwig format, using bedGraphToBigWig from the UCSC Genome Browser for visualisation in the Integrated Genomics Viewer. Potential 5hmC-enriched regions (hMRs) were identified using MACS (version 1.4.2) and the parameters used were macs14 -p 1e-3 -f BAM -g hs. 43 Peak calls were merged using bedtools merge and only those peak regions that appeared in more than 10 samples and that were less than 1000 bp were retained. Blacklisted genomic regions that tend to show artefactual signals, according to ENCODE, were also filtered. The hMRs for each patient were generated by intersecting the individual peak call file with the merged peak file. The hMRs within chromosome X and chromosome Y were excluded and used as input for the downstream analyses. Dimensionality reduction was performed using the non-parametric t-SNE algorithm implemented in the RtSNE package (v1.3.1). For heatmap visualisation, unsupervised hierarchical clustering was carried out with the pheatmap package (v1.0.12) using Euclidean distance and complete-linkage criteria; neither procedure assumes a normal distribution. For AM patients with matched tissue, these samples were assigned to the training set and all remaining samples to the validation set using train_test_split from the Scikit-Learn package (version 0.22.1) in Python (version 3.6.10); a logistic-regression CV (LR) model was then trained to establish diagnosis models. In the training cohort, we identified differentially 5hmC-enriched Regions (DhMRs) using DESeq2 package (version 1.30.0) in R (version 3.5.0), with filtering threshold (p < 0.05 and |log 2 FoldChange| ≥ 0.5). To avoid overfitting, 5 rounds of 10-fold cross-validation was performed. The training set was randomly split into 10 folds; 9 folds formed the training subset and the remaining fold served as the test subset. In each training subset, the DhMRs between adenomyosis samples and healthy controls with false discovery rate p < 0.01 and |log 2 FoldChange| ≥ 0.5 (t-test) remained as candidates. Then, we performed 100 repeats to further select markers using the logistic regression CV (LR) model, and a panel of DhMRs in each training subset that appeared in at least 95% of the iterations was retained. Thus, 10-fold cross-validation was repeated 100 times each round. 42 Finally, markers retained in ≥3 cross-validation rounds were used to build the final diagnostic model in the training set, which was then applied to the validation set. The regularisation α was chosen to maximise accuracy on a grid from 0.1 to 0.9. Receiver operating characteristic (ROC) analysis was used to evaluate model performance. We used the ChIPseeker R package (version 1.20.0) 45 to annotate the DhMRs, and genes closest to the marker regions were used for the following functional analyses. Gene ontology (GO) enrichment analysis (Biological Process) was performed using the ClueGO (version 2.5.5) and CluePedia (version 1.5.5) plug-ins in Cytoscape (version 3.7.2). We used the following parameters: Medium Network Specificity, Bonferroni step down p value correction, and two-sided hypergeometric test. The endometrial normal cell line CTCC-008-0014 was purchased from Meisen Cell (Zhejiang, PR China, Catalogue number:KC230110014). CTCC-008-0014 cells were cultured in DMEM/F12 (Meisen, Zhejiang, PR China) supplemented with 10% FBS, 10 μg/mL insulin, 1% penicillin-streptomycin, and 5 ng/mL EGF. RNAi knockdown was performed using Lipofectamine 3000 (Thermo Scientific, Cat# L3000-015) in endometrial normal cells with 25 nM siRNA according to the manufacturer's instructions. For HOXD9 knockdown, the following siRNAs were used: siCtrl: 5′-AACGTACGCGGAATACTTCGA-3′ and si-HOXD9: 5′-GCUGUUCGCUGAAGGAGGATT-3′. Fresh tissue specimens were homogenised in TRIzol (Invitrogen, Cat# 99939401). Chloroform (0.2 mL per 1 mL TRIzol) was added, the mixture vortexed for 15 s, incubated at room temperature for 3 min and centrifuged (12,000× g , 4 °C, 15 min). The aqueous phase was transferred to a new tube, mixed with an equal volume of isopropanol, held at −20 °C for 30 min and centrifuged (12,000× g , 4 °C, 10 min). The pellet was washed with 75% ethanol (in DEPC-treated water) and re-centrifuged (7500× g , 4 °C, 5 min). After air-drying for 5 min, the pellet was dissolved in 30 μL DEPC-treated water. RNA concentration and A260/A280 ratio were determined using a NanoDrop One spectrophotometer (Thermo, Cat# Q33216 ). For reverse transcription, 1 μg total RNA was combined with 4 μL 5× PrimeScript RT Master Mix (Accurate Biology, Cat# AG11706) in a final volume of 20 μL, briefly centrifuged, incubated at 37 °C for 15 min and inactivated at 85 °C for 5 s. Primers were designed with Primer Premier 5.0 and synthesised by Qil Biotech (Beijing, China); sequences are listed in Table 1 . Reactions (20 μL) were assembled as detailed in Supplementary Materials Table S2 and contained 10 μL TB Green Premix Ex Taq II (2×, Accurate Biology, Cat# AG11702), 0.4 μM forward and reverse primers, and 2 μL template cDNA (1:10 dilution). Amplification was performed on a CFX96 real-time PCR system (Bio-Rad) under the following conditions: initial denaturation at 94 °C for 6 min; 40 cycles of 94 °C for 1 min, 55 °C for 1 min, and 72 °C for 1 min; final extension at 72 °C for 1 min; ending at 4 °C. Table 1 Primers used in qPCR. Gene Forward primer (5' → 3′) Reverse primer (5' → 3′) GAPDH ATTCCATGGCACCGTCAAGG TCGCCCCACTTGATTTTGGA APP TCTCGTTCCTGACAAGTGCAA GCAAGTTGGTACTCTTCTCACTG HOXD9 GGCGCCACTACGGGATTAAG CAGGAACGAGTTGCACGAGA PCDH9 ACTATGGAGACCCCGTTGGA CACCTTGCTTGCTCCTCTGA Primers used in qPCR. Confluent monolayers were detached with 0.25% (w/v) trypsin–EDTA, collected by centrifugation (300× g , 3 min), and resuspended in complete growth medium. Cell density was quantified with a haemocytometer and adjusted to 5 × 10 6 cells mL −1 ; 2 mL aliquots were seeded into each well of a 6-well plate and incubated at 37 °C, 5% CO 2 until 100% confluence was re-established. A sterile 200 μL pipette tip was drawn across the monolayer against a ruler to create a uniform, straight wound. Detached cells were removed by three gentle PBS washes, and 2 mL serum-free medium was added to each well. The denuded area was immediately photographed (0 h) under an inverted light microscope and a fluorescence microscope equipped with a digital camera. Identical fields were re-imaged every 12 h until complete closure. Wound width was measured with ImageJ and expressed as mean ± SD from three independent experiments, each performed in duplicate. Cell migration was quantified in 24-well Transwell inserts (8 μm pore, Corning 3422). Log-phase cells were detached with 0.25% trypsin–EDTA, washed, and resuspended in serum-free medium. Viability (>95%) was verified by trypan-blue exclusion and the suspension adjusted to 5–6 × 10 5 cells ml −1 . A 100 μL aliquot (5–6 × 10 4 cells) was plated in the upper chamber; the lower well received 750 μL complete medium containing 20% FBS as chemoattractant. After 16 h at 37 °C/5% CO 2 , inserts were removed, medium aspirated, and non-migrated cells wiped from the upper membrane with a cotton swab. Membranes were rinsed three times with PBS, fixed and stained for 1 h in 0.1% crystal violet dissolved in 4% paraformaldehyde, and washed again with PBS. Following air-drying, three random 200 × fields per membrane were imaged on an inverted microscope and migrated cells counted with ImageJ (v1.53, NIH). Data are expressed as mean migrated cells per field. Total proteins were extracted from cells using RIPA buffer (Solaibao, Cat# R0010). The protein concentration of the whole extracts was determined using a BCA protein detection kit (Invitrogen, Cat# 23227). Approximately 50 μg of protein extract per sample was separated by SDS-polyacrylamide gel electrophoresis and subsequently transferred onto a PVDF membrane (Merck Millipore, Cat# ISEQ00010). The membrane was blocked with 5% skim milk and incubated overnight at 4 °C with anti-APP (1:1000, Abcam, Cat# ab32136,Purification technique: Affinity purification Protein A. The antibody detects the C-terminal fragment of APP, validated by knockdown experiment) anti-HOXD9 (1:2000, OriGene Technologies, Cat# TA504872, Purification:affinity purified by Protein A, validated by knockdown experiment), and anti-GAPDH (1:3000, Solebao, Cat# SE247). Following this incubation, the membranes were treated with a horseradish peroxidase (HRP)-conjugated secondary antibody (1:5000, Bio-Rad, Cat# 1706515) for 90 min. After extensive washing in PBST, protein expression levels were quantified using Quantit-ONE software (Bio-Rad Laboratories, USA) along with an ECL chemiluminescence kit (Bio-Rad, Cat# 1705061). All procedures were approved by the Peking University Animal Ethics Committee (BCJI0251), and all animal use complied with relevant ethical guidelines. Thirty-four virgin female ICR mice received oral tamoxifen (1 mg kg −1 day −1 in peanut oil) from postnatal days 2–5 to induce adenomyosis. At 11 weeks, mice underwent bilateral ovariectomy under isoflurane anaesthesia; after 7 days' recovery, three artificial menstrual cycles were administered (oestrogen: 100 ng β-oestradiol, 3 days; combined: 10 ng β-oestradiol + 1 mg progesterone, 3 days; progestogen: 1 mg progesterone, 3 days plus 24 h withdrawal). During the final progestogen phase, mice were randomised to daily intraperitoneal PBS (0.1 mL, n = 11–12) or deferoxamine mesylate (100 mg kg −1 , n = 11–12) for 3 days. Uteri were collected 48 h post-injection. Successful induction was confirmed by uterine enlargement with endometrial invasion ≥1 mm into the myometrium; these changes were attenuated by deferoxamine (DFO) ( Supplementary Figures S1 and S2 ). The female mice were provided by the Laboratory Animal Center of Peking University and raised in the barrier environment of Peking University Health Science Center (temperature 22 ± 2 °C, humidity 50%–60%, 12-h light–dark cycle, free feeding and drinking). The female mice were in good health, with no reproductive tract malformations or infections. No animals were excluded. Uterine tissue was fixed in 4% (w/v) paraformaldehyde (Servicebio, Cat# G1101) for 24 h, dehydrated through graded ethanol solutions, cleared in xylene (Aladdin, Cat# X112054) and embedded in paraffin. Sections (3–8 μm) were cut, dewaxed in xylene and rehydrated through descending ethanol concentrations. Masson staining was performed according to the manufacturer's instructions (Solarbio, Cat# G1340). Briefly, nuclei were stained with haematoxylin (Gene Technology, Cat# GT100540 ) for 10 min, followed by Ponceau–acid fuchsin (Solarbio, Cat# G1342) for 5 min. Sections were then treated with phosphomolybdic–phosphotungstic acid solution (Solarbio, Cat# G1343) for 5 min. A weak-acid working solution (distilled water:weak-acid stock = 2:1) was used to rinse slides for 30 s between each step. After a final rinse, sections were incubated in aniline blue (Solarbio, Cat# G1344) for 2 min, briefly washed in 1% acetic acid (30 s), rapidly dehydrated in 95% ethanol (2–3 s) and twice in absolute ethanol (5–10 s each), cleared in xylene (2 × 1–2 min) and mounted with neutral balsam. Collagen (blue) was visualised under an Olympus BX53 light microscope and quantified with ImageJ v1.53 as the percentage of blue-stained area relative to the total tissue area. For immunohistochemical analysis, ten consecutive sections were cut from each paraffin block. Following deparaffinization, rehydration, and antigen retrieval, endogenous peroxidase activity was quenched using 3% hydrogen peroxide. Non-specific binding was blocked with bovine serum albumin (BSA, Servicebio, Cat#: G5001). Tissue sections were then incubated overnight at 4 °C with primary antibodies (detailed in Supplementary Table S3 ). Subsequently, sections were incubated with a secondary antibody using a two-step detection kit (Servicebio, Cat#: GB23303, diluted 1:200), followed by visualisation with 3,3′-diaminobenzidine (DAB, Servicebio, Cat#: G1211). Haematoxylin was used for counterstaining, and slides were dehydrated and mounted for microscopic examination. Six random high-power fields (×400 magnification) containing glandular epithelium were captured using a light microscope (Nikon, Cat#: E100). A panoramic digital scan was performed using a Nikon DS-U3 imaging system, and representative areas were selected by a pathologist from Peking University Third Hospital. All staining procedures were performed synchronously across groups to ensure consistency. Staining was evaluated using the histochemistry score (H-score = positive cell ratio score × staining intensity score). In each section, 5 to 10 microscopic high-power fields were randomly selected in areas with glandular cells, and the average score of each field was taken as the score of the section. Statistical analyses were conducted with GraphPad Prism 8.0 and R 3.6.3. Normality was evaluated by the Shapiro–Wilk test; normally distributed data were analysed with two-tailed t-tests (paired or unpaired), one-way ANOVA followed by Tukey's post-hoc test, or two-way repeated-measures ANOVA with Bonferroni correction, while non-normal data were analysed with the Kruskal–Wallis test and Dunn's correction. Multiple comparisons were controlled using the Benjamini–Hochberg false-discovery-rate procedure. Correlations were assessed with two-tailed Pearson's product–moment correlation. A p < 0.05 was considered statistically significant; exact p-values are reported in the figure legends. The sponsors had no role in study design, data collection, analysis or interpretation, manuscript writing, or the decision to submit for publication.

Table 2 shows the baseline characteristics of the patients. The adenomyosis and control groups comprised 51 and 46 participants, respectively, with mean ages of 45.29 and 48.30 years. All adenomyosis diagnoses were histopathologically confirmed. There were 27 (52.94%), 26 (50.98%), and 5 (9.80%) patients with adenomyosis complicated by severe dysmenorrhoea (VAS ≥7), menorrhagia, and infertility, respectively. Deep infiltrating endometriosis (DIE) was found in 12 patients (23.53%), chocolate ovarian cysts in 13 (25.49%), and uterine myomas in 31 (60.78%). Diagnostic accuracy was 78.38% for CA125 (29/37), 22.22% for CA199 (8/36), 80.39% for TVS (41/51), and 91.43% for MRI (32/35; 20 intrinsic, 15 extrinsic). Fifteen patients (29.41%) had received GnRHa pretreatment within 6 months before surgery. Table 2 Adenomyosis patient's characteristics. Adenomyosis (N = 51) Age (years) 45.29 ± 5.84 MRI classification  Type A (intrinsic) (%) 20/51 (39.22)  Type B (extrinsic) (%) 15/51 (29.41)  Type C (NA) 16/51 (31.37) Menstruation  Severe dysmenorrhoea (%) 27/51 (52.94)  Menorrhagia (%) 26/51 (50.98) History of gestation  Gravidity (%) 43/51 (84.31)  Vaginal delivery (%) 22/51 (43.14)  Cesarean section (%) 18/51 (35.29)  Dilatation and curettage (%) 29/51 (56.86)  Abortion or stillborn (%) 5/51 (9.80)  Infertility (%) 5/51 (9.80) Complication  DIE (%) 12/51 (23.53)  Ovarian chocolate cyst (%) 13/51 (25.49)  Uterine myoma (%) 31/51 (60.78) Examination  CA125 (accuracy) (%) 29/37 (78.38)  CA199 (accuracy) (%) 8/36 (22.22)  Haemoglobin (g/l) 112.51 ± 27.45  TVS (accuracy) (%) 41/51 (80.39)  MRI (accuracy) (%) 32/35 (91.43) Preoperative treatment  GnRHa (%) 15/51 (29.41)  None (%) 36/51 (70.59) Adenomyosis patient's characteristics. To explore the potential diagnostic value of plasma cfDNA 5hmC and to find more effective biomarkers, patients with adenomyosis and healthy subjects were divided into a training cohort (N = 49) and a validation cohort (N = 48). The patients with adenomyosis in the training cohort all had matched tissues. Meanwhile, we compared the two adenomyosis subtypes, enabling diagnosis and classification by liquid biopsy ( Fig. 2 a). Principal component analysis of plasma cfDNA 5hmC profiles completely separated patients from controls ( Fig. 2 b). Differential analysis (p < 0.05, |log 2 FoldChange| ≥ 0.5) identified 1500 DhMRs, with 966 hyper- and 534 hypo-hydroxymethylated loci in patients versus controls ( Supplementary Table S3 , Fig. 2 c). Genome-wide profiling revealed that differential 5hmC peaks are enriched within ±2 kb of the transcription start site (TSS) ( Fig. 2 d) and are significantly lower in patients with adenomyosis than in healthy controls. Most of these differential peaks are distributed within introns, intergenic regions, and promoters ( Fig. 2 e and f). Representative DhMRs—hyper-hydroxymethylated CCDC149 (p = 0.015) and hypo-hydroxymethylated PSMB1 (p = 1.0 × 10 −4.9 )—are shown in Fig. 2 g and h. Finally, the top 200 most-variable DhMRs distinguished patients from healthy individuals in the training cohort ( Fig. 2 i). Fig. 2 Distribution of 5hmC in plasma of adenomyosis and healthy subjects. (a). Schematic diagram of research design overview. (b). PCA differentiates healthy people from patients with adenomyosis (healthy in green, patients with adenomyosis in red). (c). Volcano map (healthy versus adenomyosis). HMR significantly changed (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are marked in red (up) or green (down). The black dots represent the HMR with no difference. (d). Meta-gene heat-maps of 5hmC signal centred ±2 kb around the transcription start site (TSS) in healthy (top) and adenomyosis (bottom) endometrium. Colour intensity represents normalised 5hmC read density. (e, f). The pie chart shows the percentage of 5hmC peaks for up-regulated and down-regulated genes in each category of genomic traits. The Promoter region is defined as 2 KB around TSS. (g, h). Box diagram of CCDC149 and PSMD1 in healthy persons and adenomyosis group. (i). Heatmap of 23 healthy and 26 patients with adenomyosis based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) (healthy in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. Distribution of 5hmC in plasma of adenomyosis and healthy subjects. (a). Schematic diagram of research design overview. (b). PCA differentiates healthy people from patients with adenomyosis (healthy in green, patients with adenomyosis in red). (c). Volcano map (healthy versus adenomyosis). HMR significantly changed (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are marked in red (up) or green (down). The black dots represent the HMR with no difference. (d). Meta-gene heat-maps of 5hmC signal centred ±2 kb around the transcription start site (TSS) in healthy (top) and adenomyosis (bottom) endometrium. Colour intensity represents normalised 5hmC read density. (e, f). The pie chart shows the percentage of 5hmC peaks for up-regulated and down-regulated genes in each category of genomic traits. The Promoter region is defined as 2 KB around TSS. (g, h). Box diagram of CCDC149 and PSMD1 in healthy persons and adenomyosis group. (i). Heatmap of 23 healthy and 26 patients with adenomyosis based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) (healthy in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. Pathway analysis of genes bearing differential 5hmC (DhMGs) in patients with adenomyosis ( Supplementary Table S3 , Table 2 ) revealed enrichment of several canonical pathways ( Fig. 3 a and b). The top downregulated genes were related to phosphatidylinositol 3-kinase activity ( Fig. 3 a), while upregulated genes were enriched in positive regulation of cell migration ( Fig. 3 b). The protein interaction network ( Fig. 3 c and d) showed these genes: JAK2, CDC42, MAPK1 , EGFR, HIF1A, and ITGB1. Fig. 3 GO enrichment analysis and function exploration of 5hmC markers using Metascape software (p < 0.05 & |log 2 FoldChange| ≥ 0.5). (a). Metascape bar graph for viewing down markers in GO enrichment clusters. (b). Metascape bar graph for viewing up markers in GO enrichment clusters. (c). PPI network analysis of PI3K-mTOR-AKT signalling pathway. (d). PPI network analysis of positive regulation of cell migration. GO enrichment analysis and function exploration of 5hmC markers using Metascape software (p < 0.05 & |log 2 FoldChange| ≥ 0.5). (a). Metascape bar graph for viewing down markers in GO enrichment clusters. (b). Metascape bar graph for viewing up markers in GO enrichment clusters. (c). PPI network analysis of PI3K-mTOR-AKT signalling pathway. (d). PPI network analysis of positive regulation of cell migration. To determine whether plasma cfDNA 5hmC profiles mirror intrauterine pathology, we applied 5hmC-Seal to adenomyotic lesions and normal myometrium from healthy controls. PCA based on tissue 5hmC levels clearly separated adenomyosis samples from controls ( Fig. 4 a). Differential analysis (p < 0.05, |log 2 FoldChange| ≥ 0.5) identified 2058 DhMRs, with 1055 hyper- and 1003 hypo-hydroxymethylated regions in patient tissues ( Supplementary Table S4 , Fig. 4 b). Genome-wide distribution showed that the majority of differential 5hmC peaks were concentrated within ±2 kb of the transcription start site (TSS) ( Fig. 4 c). This region exhibited a marked difference between the groups, with 5hmC levels being significantly lower in adenomyotic lesions than in non-adenomyosis myometrium, mirroring the pattern observed in plasma cfDNA ( Fig. 2 d). Unsupervised clustering of the top 200 most-variable DhMRs clearly segregated the two diagnostic groups ( Fig. 4 d). Next, to verify the correlation between tissue genomic DNA and plasma cfDNA 5hmC, we performed correlation analysis on the samples of the same tissue and plasma. We found that 5hmC markers in plasma cfDNA correlated with 5hmC markers in pathological tissue genomic DNA ( Fig. 4 e). For example, SIRT1 was significantly correlated (R = 0.551) ( Fig. 4 f, Supplementary Figure S3a ). In order to search for characteristic markers of 5hmC reflected in plasma from focal tissues, we screened 141 DhMGs co-expressed in tissues and plasma, of which 51 genes were down-regulated and 90 genes were up-regulated ( Fig. 4 g and h, Supplementary Table S5 ). We analysed pathway enrichment of the 141 DhMGs and found significant associations with vascular smooth-muscle contraction, neuronal differentiation, focal-adhesion assembly and cell-substrate junction organisation ( Fig. 4 i and j). Genes identified included AUTS2 (neurodevelopment-related), BCAS3 (oestrogen-related), and FOXO3 (immune-related) ( Fig. 4 j). Fig. 4 Correlation analysis of tissue and plasma characteristics . (a). PCA differentiation between controls and patients with adenomyosis (controls in green, patients with adenomyosis in red). (b). Volcano plot. Significantly altered genes (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are highlighted in red (up) or green (down) using the control group as the reference. Black dots represent the genes that are not differentially expressed. (c). Meta-gene profiles and heat-maps of 5hmC density centred ±2 kb around the transcription start site (TSS) in control (upper panel) and adenomyosis (lower panel) tissues. The average signal (left y-axis) and heat-map scale (right) illustrate the significant reduction of 5hmC at TSS regions in adenomyosis compared with controls. (d). Heatmap of 22 controls and 26 patients with adenomyosis based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) (controls in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. (e). Tissue was associated with plasma markers (the abscissa is tissues; the ordinate is plasma). (f). SIRT1 is correlated with tissue in plasma. (g, h). Biomarkers in plasma overlap the biomarkers in tissue (plasma in red, tissue in green). (i). 141 markers in GO enrichment bar plot (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (j). 141 markers in GO enrichment and Gene-Concept Network. The node size is proportional to the p-value calculated from the network. Correlation analysis of tissue and plasma characteristics . (a). PCA differentiation between controls and patients with adenomyosis (controls in green, patients with adenomyosis in red). (b). Volcano plot. Significantly altered genes (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are highlighted in red (up) or green (down) using the control group as the reference. Black dots represent the genes that are not differentially expressed. (c). Meta-gene profiles and heat-maps of 5hmC density centred ±2 kb around the transcription start site (TSS) in control (upper panel) and adenomyosis (lower panel) tissues. The average signal (left y-axis) and heat-map scale (right) illustrate the significant reduction of 5hmC at TSS regions in adenomyosis compared with controls. (d). Heatmap of 22 controls and 26 patients with adenomyosis based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) (controls in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. (e). Tissue was associated with plasma markers (the abscissa is tissues; the ordinate is plasma). (f). SIRT1 is correlated with tissue in plasma. (g, h). Biomarkers in plasma overlap the biomarkers in tissue (plasma in red, tissue in green). (i). 141 markers in GO enrichment bar plot (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (j). 141 markers in GO enrichment and Gene-Concept Network. The node size is proportional to the p-value calculated from the network. We generated genome-wide 5hmC profiles for patients in the validation set, including 23 healthy and 25 patients with adenomyosis. While markers such as SIRT1 demonstrated strong tissue–plasma correlation (r = 0.551, p < 0.01, Fig. 4 f), individual marker performance was limited (SIRT1 AUC = 0.60, Supplementary Figure S4 ), supporting a multi-marker approach. Using recursive feature elimination based on the logistic regression CV estimator, we reduced the number of 5hmC markers from 141 to 10, achieving the optimal cross-validation score ( Fig. 5 a). PCA and unsupervised clustering showed that the ten 5hmC markers ( Table 3 ) distinguished patients with adenomyosis from healthy subjects in both the training and validation cohorts ( Fig. 5 b–e).The diagnostic model built on these ten 5hmC markers performed perfectly in the training set (AUC = 1.00; sensitivity and specificity both 1.00; Fig. 5 f) and maintained high performance in the validation set (AUC = 0.91; sensitivity 0.88; specificity 0.87; Fig. 5 g). These 10 5hmC markers achieved discrimination in the training cohort (AUC = 1.00) and validation cohort (AUC = 0.91) ( Fig. 5 h). Full model parameters and accuracy metrics are given in Table 4 . We also calculated the individual AUC for each of the ten 5hmC markers in the training and validation cohorts ( Supplementary Figure S5a ). Fig. 5 5hmC markers' diagnosis adenomyosis in the training and validation cohort . (a). Workflow for building the diagnostic model. (b). PCA identification of healthy and patients with adenomyosis in the training cohort (healthy in green, patients with adenomyosis in red). (c), Heatmaps of ten 5hmC markers with adenomyosis in the training cohort (healthy in green, patients with adenomyosis in red) Unsupervised hierarchical clustering was performed across genes and samples. (d). PCA identification of healthy and patients with adenomyosis in the validation cohort (healthy in green, patients with adenomyosis in red). (e). Heatmaps of ten 5hmC markers with adenomyosis in the validation cohort (healthy in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. (f, g). Confusion matrices of the DhMR-based diagnostic model in the training (f) and validation (g) cohorts. (h). ROC curves for adenomyosis diagnosis: training cohort (AUC = 1.00), validation cohort (AUC = 0.91). Table 3 Characteristics and model coefficients of 10 5hmC markers. Markers coef std err z p > |z| 2.5% 97.5% const −1.7485 0.686 −2.547 0.011 −3.094 −0.403 JAK2 0.0739 0.027 2.717 0.007 0.021 0.127 ACOT12 0.1469 0.05 2.937 0.003 0.049 0.245 NAPB 0.1364 0.047 2.927 0.003 0.045 0.228 PAIP1 0.0316 0.021 1.487 0.137 −0.01 0.073 PBX1 0.0867 0.028 3.128 0.002 0.032 0.141 QKI 0.0692 0.024 2.854 0.004 0.022 0.117 SIRPA 0.0486 0.022 2.238 0.025 0.006 0.091 SPRY1 −0.0763 0.026 −2.902 0.004 −0.128 −0.025 ST3GAL4 0.061 0.023 2.632 0.008 0.016 0.106 STK17B 0.0722 0.025 2.845 0.004 0.022 0.122 Note: SE: standard error of coefficient; Z: The Z-statistic of Wald. Table 4 Diagnostic performance of the 10-5hmC-marker model in training and validation sets. Metric Percentage (95% CI) Training set value Validation set value Sensitivity 100 (100–100) 87 (69.5–100) Specificity 100 (100–100) 88 (76–100) AUC 100 (100–100) 91 (83–99) NPV 100 (98.5–100) 86.9 (84–89.8) PPV 100 (99–100) 88 (85–91) Accuracy 100 (99.3–100) 87.5 (85.1–90) 5hmC markers' diagnosis adenomyosis in the training and validation cohort . (a). Workflow for building the diagnostic model. (b). PCA identification of healthy and patients with adenomyosis in the training cohort (healthy in green, patients with adenomyosis in red). (c), Heatmaps of ten 5hmC markers with adenomyosis in the training cohort (healthy in green, patients with adenomyosis in red) Unsupervised hierarchical clustering was performed across genes and samples. (d). PCA identification of healthy and patients with adenomyosis in the validation cohort (healthy in green, patients with adenomyosis in red). (e). Heatmaps of ten 5hmC markers with adenomyosis in the validation cohort (healthy in green, patients with adenomyosis in red). Unsupervised hierarchical clustering was performed across genes and samples. (f, g). Confusion matrices of the DhMR-based diagnostic model in the training (f) and validation (g) cohorts. (h). ROC curves for adenomyosis diagnosis: training cohort (AUC = 1.00), validation cohort (AUC = 0.91). Characteristics and model coefficients of 10 5hmC markers. Note: SE: standard error of coefficient; Z: The Z-statistic of Wald. Diagnostic performance of the 10-5hmC-marker model in training and validation sets. Patients with adenomyosis were evaluated by MRI imaging. Intrinsic, extrinsic, and control tissue samples were analysed using 5hmC markers. PCA revealed separation between each adenomyosis subtype and controls ( Fig. 6 a). DhMRs were most frequent in introns, intergenic regions and promoters. For each feature type, the difference between intrinsic and extrinsic subtypes was statistically significant, with introns showing the largest effect ( Fig. 6 b). To identify intrinsic and extrinsic DhMRs, we selected 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5). Unsupervised hierarchical clustering of these DhMRs separated the two subtypes ( Fig. 6 c). Significant differences were found between the two subtypes ( Supplementary Figure S6a–f ). Compared with the extrinsic subtype, 360 DhMRs were significantly up-regulated in the intrinsic subtype and 324 DhMRs were significantly down-regulated (p < 0.05, |log 2 FoldChange| ≥ 0.5) ( Supplementary Table S6 , Fig. 6 d). Most up-regulated and down-regulated DhMRs were distributed in intronic, intergenic, and promoter regions ( Supplementary Figure S6c and d ). We compared the genome-wide 5hmC distribution in tissues, which mirrored the plasma pattern ( Supplementary Table S7 , Supplementary Figure S7a–h ). Between intrinsic and extrinsic subtypes, 100 DhMGs were significantly up-regulated ( Figs. 6 e) and 57 DhMGs were significantly down-regulated ( Fig. 6 f). We conducted pathway enrichment analysis of 100 up-regulated and 57 down-regulated DhMGs ( Supplementary Table S8 ). GO functional enrichment showed that extrinsic-subtype DhMGs were enriched in interleukin-12 regulation and production, and angiogenesis pathways ( Fig. 6 g). Compared with intrinsic genes, extrinsic genes were enriched in signalling pathways related to immune response, amyloidosis, nervous system development, and DNA damage ( Fig. 6 h). Then 31 genes that were significantly enriched in the signalling pathway were extracted for unsupervised cluster analysis. It was found that these 31 differential genes extracted could also effectively separate the two lesion types ( Fig. 6 i). Fig. 6 5hmC feature marker to distinguish different types and mechanisms of adenomyosis . (a). PCA differentiation between Control, intrinsic and extrinsic tissue (Control in green, extrinsic in blue, intrinsic in pink). (b). Characteristics of 5hmC distribution in tissue gDNA of Control, intrinsic and extrinsic tissue in patients with adenomyosis (N = 35) (controls in green, intrinsic in blue, extrinsic in pink). (c). Heatmap of 21 controls, 20 intrinsic and 15 extrinsic patients, based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) or higher (controls in green, intrinsic patients in blue, extrinsic patients in pink). (d). Volcano map (intrinsic patients versus extrinsic patients). HMR significantly changed (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are indicated in pink (up) or blue (down). The black dots represent The HMR with no difference. (e, f). Biomarkers in plasma overlap the biomarkers in tissue (plasma in red, tissue in green). (g, h). Down 57 markers (g) and up 100 markers (h) in GO enrichment bar plot (∗p = 0.005–0.05, ∗ ∗p = 0.0005–0.005). (i). Based on the differences between 31 genes (p < 0.05 & |log 2 FoldChange| ≥ 0.5) in heatmaps of 20 intrinsic patients and 15 extrinsic patients (pink in intrinsic patients, blue in extrinsic patients). 5hmC feature marker to distinguish different types and mechanisms of adenomyosis . (a). PCA differentiation between Control, intrinsic and extrinsic tissue (Control in green, extrinsic in blue, intrinsic in pink). (b). Characteristics of 5hmC distribution in tissue gDNA of Control, intrinsic and extrinsic tissue in patients with adenomyosis (N = 35) (controls in green, intrinsic in blue, extrinsic in pink). (c). Heatmap of 21 controls, 20 intrinsic and 15 extrinsic patients, based on top 200 DhMRs (p < 0.05 & |log 2 FoldChange| ≥ 0.5) or higher (controls in green, intrinsic patients in blue, extrinsic patients in pink). (d). Volcano map (intrinsic patients versus extrinsic patients). HMR significantly changed (p < 0.05 & |log 2 FoldChange| ≥ 0.5) are indicated in pink (up) or blue (down). The black dots represent The HMR with no difference. (e, f). Biomarkers in plasma overlap the biomarkers in tissue (plasma in red, tissue in green). (g, h). Down 57 markers (g) and up 100 markers (h) in GO enrichment bar plot (∗p = 0.005–0.05, ∗ ∗p = 0.0005–0.005). (i). Based on the differences between 31 genes (p < 0.05 & |log 2 FoldChange| ≥ 0.5) in heatmaps of 20 intrinsic patients and 15 extrinsic patients (pink in intrinsic patients, blue in extrinsic patients). PPI network analysis revealed that APP acts as a central hub among the 31 differentially methylated 5hmC genes, with the highest connectivity score ( Fig. 7 a and b).Leveraging the PPI network constructed from the STRING database, we identified several key genes interacting with APP , including Amyloid Beta A4 Precursor Protein-Binding Family B Member 1 ( APBB1 ), Beta-Secretase 2 ( BACE2 ), Presenilin-1 ( PSEN1 ), Apolipoprotein E ( APOE ), Sortilin-Related Receptor 1 ( SORL1 ), Nicastrin ( NCSTN ), Clusterin Alpha Chain ( CLU ), Amyloid Beta A4 Precursor Protein-Binding Family A Member ( APBA1 ), Tumour Necrosis Factor Receptor Superfamily Member 21 ( TNFRSF21 ), and NEDD8-Activating Enzyme E1 ( NAE1 ) ( Fig. 7 c). Furthermore, Gene Ontology enrichment analysis revealed that these APP -related genes were significantly over-represented within pathways driving tissue fibrogenesis and amyloidosis. We therefore examined whether APP modulates fibrosis in adenomyosis ( Fig. 7 d). DFO administration reduced APP expression and collagen deposition in mouse uterine sections ( Supplementary Figure S2a–c and Supplementary Figure S8a–c ). We immunostained APP in 26 formalin-fixed, paraffin-embedded uterine sections from patients with intrinsic adenomyosis, extrinsic adenomyosis, or control hysterectomies. After discarding substandard sections, 21 extrinsic, 17 intrinsic, and 21 control sections were available for analysis. Histological scoring was conducted on the ectopic and eutopic endometrium of the extrinsic and intrinsic types, respectively, as well as on the endometrium of the control group. APP expression was higher in patients with extrinsic adenomyosis than in those with intrinsic adenomyosis, with median H-scores of 7.2 (IQR, 4.0–11.3) and 4.0 (IQR, 1.7–7.0), respectively (p < 0.001). APP expression in the adenomyosis group was elevated compared to the control group, with median scores of 2.0 (range, 0–4.0) for the ectopic endometrium and 1.7 (range, 0–5.0) for the eutopic endometrium, p < 0.001; no significant difference was observed between the control group and the eutopic endometrium, p = 0.522 ( Fig. 7 e and f). Fig. 7 Potential relevance of APP in the classification of adenomyosis. ( a). PPI network of 5hmC differential genes in the GO signalling pathway in the STRING database. (b). GO enrichment histogram of APP -related genes in PPI network (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (c). Functional protein–protein interaction (PPI) network of APP in the STRING database. (d). GO enrichment histogram of APP -related genes in PPI network (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (e). Protein expression of APP markers in normal and present endometrium, intrinsic and extrinsic endometrium by IHC. The expression of APP was observed at 10× (10×) and 20× (20×) magnifications. (f). Immunohistochemical score (H-Score) showed the expression of APP . Potential relevance of APP in the classification of adenomyosis. ( a). PPI network of 5hmC differential genes in the GO signalling pathway in the STRING database. (b). GO enrichment histogram of APP -related genes in PPI network (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (c). Functional protein–protein interaction (PPI) network of APP in the STRING database. (d). GO enrichment histogram of APP -related genes in PPI network (∗p = 0.005–0.05, ∗∗p = 0.0005–0.005). (e). Protein expression of APP markers in normal and present endometrium, intrinsic and extrinsic endometrium by IHC. The expression of APP was observed at 10× (10×) and 20× (20×) magnifications. (f). Immunohistochemical score (H-Score) showed the expression of APP . In the immortalised endometrial epithelial cell line, we overexpressed APP and analysed mRNA and protein levels ( Fig. 8 a and b). The Transwell migration assay ( Fig. 8 c) showed that APP overexpression enhanced cellular migration to the lower chamber, with an increase in migrating cells compared to the control group (p < 0.05) ( Fig. 8 d). Scratch assays ( Fig. 8 e) showed enhanced wound healing in cells overexpressing APP. In these experiments, cell lines overexpressing APP exhibited significantly enhanced wound healing capacity in endometrial epithelial cells. We conducted a comparative analysis of RNA sequencing data from the APP -overexpressing cell line ( OE-APP ) and the parental cell line (NC). PCA revealed significant differences in gene expression profiles between the OE-APP and NC groups, clearly distinguishing the OE-APP group from the NC group ( Fig. 9 a). Further studies revealed that there were 116 differentially expressed genes (DEGs) between the two groups, among which 75 were upregulated and 41 were downregulated (p < 0.05 & |log 2 FoldChange| ≥ 0.5) ( Fig. 9 b). We performed unsupervised clustering using the top 100 DEGs, which further confirmed the distinct gene expression patterns between the NC and OE-APP groups ( Fig. 9 c). We then examined the potential effects of APP overexpression on transcription factors. The results showed that 11 transcription factors were up-regulated and 16 were down-regulated in the OE-APP group ( Fig. 9 d and e). To visualise the distribution of these DEGs, we analysed the relationship between log2 fold changes and -log10 p-values. Notably, the transcription factor HOXD9 (Homoeobox D9) exhibited significant upregulation ( Fig. 9 f). Fig. 8 Effect of overexpression of APP on cell migration ability. The mRNA expression level of (a). APP gene in the OE-APP group was significantly higher than that in the NC group, and the data were standardised using GAPDH as the internal reference, with ∗∗∗∗ indicating A very significant difference compared with the NC group (p < 0.0001). (b). Western blot analysis showed that the expression level of APP protein in the OE-APP group was higher than that in the NC group, and GAPDH was used as the internal reference protein. (c). Transwell migration assay (∗p < 0.05). (d). The number of migrating cells in OE-APP group was significantly higher than that in NC group, ∗ indicating a significant difference compared with NC group (p < 0.05). (e). Scratch test showed that compared with the NC group, the OE-APP group had enhanced cell migration ability at 0 h, 12 h and 24 h, and the degree of wound healing was more obvious. The scale is 100 microns. (∗p < 0.05, ∗∗∗∗p < 0.0001, The significant difference was evaluated by the two-tailed t-tests). Fig. 9 Identification of APP transcription factors. (a). PCA maps showed differences at transcriptome level between the OE-APP group and the NC group. Red is the OE-APP group, blue is the NC group. (b). Volcano maps showed the distribution of DEGs between the OE-APP group and the NC group (p < 0.05 & |log 2 FoldChange| ≥ 0.5). Red dots indicate up-regulated genes, blue dots indicate down-regulated genes, and grey dots indicate genes with no significant change. (c). Heat maps showed unsupervised cluster analysis of differentially expressed genes between the OE-APP group and the NC group. (d). Venn diagram shows the number of upregulated transcription factors (TFS) in the OE-APP group, with blue representing OE-APP s and red representing the transcription factor database. (e). Venn diagram shows the number of down-regulated transcription factors (TFS) in the OE-APP group, with blue representing OE-APP s and red representing the transcription factor database. (f). The scatter plot shows the change and significance of the log2-fold differential expression of transcription factors (expressed as negative logarithms of p-values). The colour ranges from blue (low significance) to red (high significance), and the size of the dots indicates a log2-fold change in the gene. Effect of overexpression of APP on cell migration ability. The mRNA expression level of (a). APP gene in the OE-APP group was significantly higher than that in the NC group, and the data were standardised using GAPDH as the internal reference, with ∗∗∗∗ indicating A very significant difference compared with the NC group (p < 0.0001). (b). Western blot analysis showed that the expression level of APP protein in the OE-APP group was higher than that in the NC group, and GAPDH was used as the internal reference protein. (c). Transwell migration assay (∗p < 0.05). (d). The number of migrating cells in OE-APP group was significantly higher than that in NC group, ∗ indicating a significant difference compared with NC group (p < 0.05). (e). Scratch test showed that compared with the NC group, the OE-APP group had enhanced cell migration ability at 0 h, 12 h and 24 h, and the degree of wound healing was more obvious. The scale is 100 microns. (∗p < 0.05, ∗∗∗∗p < 0.0001, The significant difference was evaluated by the two-tailed t-tests). Identification of APP transcription factors. (a). PCA maps showed differences at transcriptome level between the OE-APP group and the NC group. Red is the OE-APP group, blue is the NC group. (b). Volcano maps showed the distribution of DEGs between the OE-APP group and the NC group (p < 0.05 & |log 2 FoldChange| ≥ 0.5). Red dots indicate up-regulated genes, blue dots indicate down-regulated genes, and grey dots indicate genes with no significant change. (c). Heat maps showed unsupervised cluster analysis of differentially expressed genes between the OE-APP group and the NC group. (d). Venn diagram shows the number of upregulated transcription factors (TFS) in the OE-APP group, with blue representing OE-APP s and red representing the transcription factor database. (e). Venn diagram shows the number of down-regulated transcription factors (TFS) in the OE-APP group, with blue representing OE-APP s and red representing the transcription factor database. (f). The scatter plot shows the change and significance of the log2-fold differential expression of transcription factors (expressed as negative logarithms of p-values). The colour ranges from blue (low significance) to red (high significance), and the size of the dots indicates a log2-fold change in the gene. In the follow-up study, we quantitatively assessed HOXD9 expression levels in both the APP overexpression group (n = 3) and the control group (n = 3) using RT-qPCR and Western blot techniques. The results showed a significant upregulation of HOXD9 expression in cell lines that overexpressed APP ( Fig. 10 a and b). Furthermore, immunohistochemical staining for HOXD9 was conducted on tissue slices from more than 20 patients diagnosed with adenomyosis. Immunohistochemical scores were calculated to assess the correlation between APP and HOXD9 expression in the same patient samples ( Fig. 10 c). Statistical analysis further showed that the expression levels of APP and HOXD9 in the pathological tissues of these patients were consistent ( Fig. 10 d). Correlation analysis revealed a significant positive correlation between APP and HOXD9 (r = 0.6963, p = 0.0007) ( Fig. 10 e). To further investigate the role of HOXD9 in cell migration mediated by APP overexpression, we employed a Transwell migration assay to assess changes in migratory capacity following HOXD9 knockdown. The experimental results showed that silencing HOXD9 significantly decreased the number of cells migrating through the lower chamber (p < 0.01) ( Fig. 11 a), which was further supported by statistical analysis ( Fig. 11 b). Next, the effect of HOXD9 knockdown in cell migration was further validated in a cell scratch assay, using endometrial epithelial cells from the APP-overexpressing group ( Fig. 11 c). Furthermore, rescue experiments yielded concordant trends ( Supplementary Figure S9a–c ). Subsequently, we validated these findings in vivo. Mice were treated with DFO to down-regulate APP expression; endometrial tissue was then harvested and immunostained for HOXD9 . Observation under 10× and 20× objectives revealed that reduced APP expression was accompanied by a marked decrease in HOXD9 levels ( Fig. 12 a and b). Fig. 10 Positive Correlation between APP and HOXD9 Gene Expression Levels Revealed by RNA Sequencing Analysis. (a). RNA sequencing results showed that the expression level of HOXD9 gene in OE-APP group was significantly higher than that in NC group (p < 0.05). (b). Western blot analysis showed that the expression levels of APP and HOXD9 proteins in the OE-APP group were higher than those in the NC group, and GAPDH was used as the internal reference protein. (c). Immunohistochemical staining showed the expression of APP and HOXD9 in the OE-APP group. The left picture is the expression of APP , and the right picture is the expression of HOXD9 . Above is a low magnification (×10) and below is a high magnification (X20), where enhanced positive signals of APP and HOXD9 can be observed in the OE-APP group. (d). Immunohistochemical H-scores for APP and HOXD9 in control versus adenomyosis tissues. Data are presented as box-and-whisker plots (median ± IQR), ∗∗∗∗p < 0.0001 versus control. (e). The scatter plot shows a positive correlation between APP and HOXD9 expression levels (r = 0.6963, p = 0.0007) (The significant difference was evaluated by the two-tailed t-tests, The relationship of APP and HOXD9 was evaluated by correlation analysis). Fig. 11 Role of HOXD9 in APP Overexpression-Mediated Cell Migration. (a). Transwell migration experiment revealed the difference of cell migration ability in the control group (NC group), the APP overexpression group ( OE-APP group) and the APP overexpression group simultaneously knocked down HOXD9 group, with A scale of 100 micron. (b). The number of migrating cells in the OE-APP group was significantly higher than that in the NC group (p < 0.0001), and the number of migrating cells in the OE-APP + si-HOXD9 group was significantly lower than that in the OE-APP group (p < 0.01). (c). Representative phase-contrast images of the scratch-wound assay at 0, 12 and 24 h after wounding (scale bar = 200 μm). A confluent monolayer of the indicated groups (NC, OE-APP, OE-APP + si-HOXD9) was wounded with a sterile 200 μL pipette tip, washed three times with PBS to remove debris and incubated in serum-free medium at 37 °C, 5% CO 2 . Wound width was measured in triplicate wells using ImageJ and expressed as percentage closure relative to 0 h. Quantification (right panel) shows mean ± SD from three independent experiments. Two-way repeated-measures ANOVA followed by Bonferroni post-test: ∗∗p < 0.01, ∗∗∗∗p < 0.0001 versus NC at the same time point. Fig. 12 Expression changes of HOXD9 in a mouse model of adenomyosis inhibited by APP .(a).Immunohistochemical staining showed the expression of HOXD9 in the control group (AM + PBS) and the inhibitor group ( APP + DFO). The left picture shows the AM + PBS group, and the right picture shows the APP + DFO group. The upper figure is a low-magnification image (×10), and the lower figure is a high-magnification image (×20). In the APP + DFO group, a decrease in the positive signal of HOXD9 can be observed. (b). The immunohistochemical h score of HOXD9 was compared with that of the inhibitor added. The data are presented in the form of box plot (median ± IQR), ∗∗∗∗ compared with the control group, p < 0.0001. Positive Correlation between APP and HOXD9 Gene Expression Levels Revealed by RNA Sequencing Analysis. (a). RNA sequencing results showed that the expression level of HOXD9 gene in OE-APP group was significantly higher than that in NC group (p < 0.05). (b). Western blot analysis showed that the expression levels of APP and HOXD9 proteins in the OE-APP group were higher than those in the NC group, and GAPDH was used as the internal reference protein. (c). Immunohistochemical staining showed the expression of APP and HOXD9 in the OE-APP group. The left picture is the expression of APP , and the right picture is the expression of HOXD9 . Above is a low magnification (×10) and below is a high magnification (X20), where enhanced positive signals of APP and HOXD9 can be observed in the OE-APP group. (d). Immunohistochemical H-scores for APP and HOXD9 in control versus adenomyosis tissues. Data are presented as box-and-whisker plots (median ± IQR), ∗∗∗∗p < 0.0001 versus control. (e). The scatter plot shows a positive correlation between APP and HOXD9 expression levels (r = 0.6963, p = 0.0007) (The significant difference was evaluated by the two-tailed t-tests, The relationship of APP and HOXD9 was evaluated by correlation analysis). Role of HOXD9 in APP Overexpression-Mediated Cell Migration. (a). Transwell migration experiment revealed the difference of cell migration ability in the control group (NC group), the APP overexpression group ( OE-APP group) and the APP overexpression group simultaneously knocked down HOXD9 group, with A scale of 100 micron. (b). The number of migrating cells in the OE-APP group was significantly higher than that in the NC group (p < 0.0001), and the number of migrating cells in the OE-APP + si-HOXD9 group was significantly lower than that in the OE-APP group (p < 0.01). (c). Representative phase-contrast images of the scratch-wound assay at 0, 12 and 24 h after wounding (scale bar = 200 μm). A confluent monolayer of the indicated groups (NC, OE-APP, OE-APP + si-HOXD9) was wounded with a sterile 200 μL pipette tip, washed three times with PBS to remove debris and incubated in serum-free medium at 37 °C, 5% CO 2 . Wound width was measured in triplicate wells using ImageJ and expressed as percentage closure relative to 0 h. Quantification (right panel) shows mean ± SD from three independent experiments. Two-way repeated-measures ANOVA followed by Bonferroni post-test: ∗∗p < 0.01, ∗∗∗∗p < 0.0001 versus NC at the same time point. Expression changes of HOXD9 in a mouse model of adenomyosis inhibited by APP .(a).Immunohistochemical staining showed the expression of HOXD9 in the control group (AM + PBS) and the inhibitor group ( APP + DFO). The left picture shows the AM + PBS group, and the right picture shows the APP + DFO group. The upper figure is a low-magnification image (×10), and the lower figure is a high-magnification image (×20). In the APP + DFO group, a decrease in the positive signal of HOXD9 can be observed. (b). The immunohistochemical h score of HOXD9 was compared with that of the inhibitor added. The data are presented in the form of box plot (median ± IQR), ∗∗∗∗ compared with the control group, p < 0.0001.

At present, clinical diagnosis and classification of adenomyosis remain limited, with a lack of specific molecular markers and an urgent need for non-invasive diagnostic techniques. In response to this unmet clinical need, we developed a classification method based on liquid biopsy and aimed to establish a diagnostic and subtyping model for adenomyosis using 5hmC profiles in plasma cfDNA, which were obtained via 5hmC-Seal sequencing. In our study cohort, plasma cfDNA 5hmC levels were globally reduced in adenomyosis patients compared to healthy controls. Differential analysis identified 1500 DhMRs (966 hyper- and 544 hypo-hydroxymethylated loci). GO functional enrichment analysis revealed that genes with elevated 5hmC were significantly enriched in biological processes including positive regulation of cell migration, protein phosphorylation, and MAPK cascade regulation, whereas genes with reduced 5hmC were associated with phosphatidylinositol 3-kinase (PI3K) signalling, hormone response, and haemostasis. Adenomyosis is recognised as an inflammatory disorder characterised by immune dysregulation, aberrant innervation, and tissue fibrosis. 46 , 47 cfDNA originates not only from lesional tissues but also reflects disease-induced microenvironmental changes 48 , 49 ; this microenvironment comprises matrix components, extracellular matrix, inflammatory mediators, and immune cells, all of which are intimately linked to disease initiation, progression, and therapeutic response. 15 , 50 Notably, oestrogen signalling via MAPK and PI3K/AKT/mTOR pathways contributes to adenomyosis pathogenesis, 51 and dysregulated cell migration represents a critical mechanism in disease progression. 52 , 53 , 54 These 5hmC marker genes and DhMGs may contribute to adenomyosis and other gynaecological conditions through modulation of these pathological processes. Protein–protein interaction network analysis further identified hub genes including JAK2, CDC42, MAPK1, EGFR, HIF1A, and ITGB1 , 55 , 56 , 57 , 58 , 59 , 60 which are implicated in the molecular mechanisms and clinical phenotypes of adenomyosis. To establish a diagnostic model, we first intersected differentially hydroxymethylated genes derived from plasma cfDNA with those from adenomyosis lesion tissues. This identified a total of 141 consistent differential genes that originated from disease lesions and were detectable in plasma. Pathway enrichment analysis confirmed that these 141 genes were significantly associated with key biological processes relevant to adenomyosis. 61 , 62 , 63 , 64 , 65 , 66 , 67 Using supervised machine learning, we further selected 10 optimal 5hmC markers from these 141 genes to distinguish adenomyosis patients from controls in both the training and validation cohorts. A logistic regression CV model established by ten 5hmC markers outperformed existing clinical serological indicators CA125 and CA199, 15 , 68 with the sensitivity of 0.88 and specificity of 0.87 (AUC = 0.91). Our results demonstrate that 5hmC markers are effective non-invasive diagnostic biomarkers for adenomyosis. Using 5hmC markers, we classified adenomyosis patients into intrinsic and extrinsic subtypes, which formed discrete clusters. Pathway enrichment analysis of DEGs that are lesion-associated revealed distinct functional alterations between intrinsic and extrinsic subtypes. The biological significance of enriched pathways is tightly correlated with adenomyosis pathogenesis. To dissect subtype-specific mechanisms and identify biomarkers, we selected nine candidate signature genes: THBS1, APP, NCOA2 (tissue fibrosis/hormonal regulation); CD47, CD226, CD36, NFKB1 (inflammation); and EPHA4, AUTS2 (neurodevelopment). Their association with adenomyosis is well-documented. 46 , 47 , 69 , 70 , 71 PPI network analysis identified APP as a central hub, with the highest connectivity score. Previous studies have shown that APP is involved in the proliferation and differentiation of neurons, synapse formation and activity of neuromuscular junction. While, high expression of APP can lead to early onset of Alzheimer's disease, suggesting the relationship between APP levels and the disease. 72 Additionally, APP -derived peptide fragments can bind to the APBB1/TIP60 acetyltransferase complex to promote transcriptional activation. 73 Accordingly, we focused on the critical role of APP in the subclassification of adenomyosis. Using PPI network analysis and GO enrichment, we found that APP was closely associated with distinct subtypes of adenomyosis. Immunohistochemical results further demonstrated that APP expression was significantly different between extrinsic and intrinsic adenomyosis. These molecular differences may underlie the heterogeneous clinical manifestations and therapeutic responses observed across subtypes, and the underlying biological mechanisms warrant further investigation. Given the prominent tissue fibrosis in the extrinsic subtype, we performed intervention studies using DFO, a pharmacological inhibitor of APP , and verified that APP is functionally involved in the regulation of fibrosis in adenomyosis. 74 , 75 Collectively, our findings indicate that APP serves as a reliable diagnostic biomarker for subtype stratification and a promising therapeutic target in adenomyosis. Furthermore, this study deepens our understanding of the regulatory roles of APP in endometrial homoeostasis and disease pathogenesis, providing a theoretical foundation for mechanism-based therapeutic interventions in endometrial disorders. Previous studies have shown APP promotes cell migration and invasion and is associated with fibrosis in multiple diseases. APP is a type I transmembrane protein whose intracellular structural domain can interact with articulin to activate downstream intracellular signalling molecules, thereby affecting the physiological functions of neural and immune cells and influencing cell migration and invasion. APP overexpression enhances, whereas its silencing suppresses, migration and invasion of human breast cancer cells. 76 , 77 APP silencing inhibit the development of nasopharyngeal carcinoma by suppressing the cell viability, migration, and invasion, and by inhibiting the EMT process through the down-regulation of the MAPK signalling pathway. 78 APP promote the proliferation and migration of prostate cancer cells by regulating the expression of metalloproteinases and EMT-related genes. 79 To test whether this mechanism operates in endometrial cells, we examined the effect of APP overexpression on endometrial epithelial cell migration. APP overexpression significantly enhanced migratory capacity, as demonstrated by Transwell and wound healing assays. These results indicate that APP -driven cell migration contributes to adenomyosis progression, linking APP dysregulation directly to disease pathogenesis. To further verify the mechanism of APP , we detected the changes of genes after overexpression of APP by RNA-seq and found that HOXD9 , which is linked to cell migration, was significantly up-regulated. These results were validated by PCR and Western blotting. Immunohistochemical staining confirmed the correlation between APP and HOXD9.HOXD9 belongs to the homoeobox gene family and functions as a transcription factor regulating coding genes and non-coding RNAs. It induces EMT and maintains invasive and migratory potential, and promotes colorectal cancer cell proliferation, invasion and migration in vitro and in vivo. 80 , 81 Furthermore, HOXD9 overexpression enhances migration and invasion, whereas its silencing suppresses proliferation, migration, and EMT in thyroid and hepatocellular carcinomas, and reduces tumorigenicity and metastasis in vivo. 80 , 82 Consistently, previous studies have linked HOXD9 to tumour invasion and migration. 83 , 84 , 85 Functional rescue experiments showed that HOXD9 knockdown following APP overexpression markedly suppressed endometrial epithelial cell migration, whereas HOXD9 overexpression after APP inhibition enhanced this phenotype. Thus, APP regulates endometrial epithelial cell migration through HOXD9 . Furthermore, our findings support HOXD9 as a key transcription factor in APP-related pathophysiology. Clinically, APP expression correlates with fibrosis severity in adenomyosis lesions. We propose that APP promotes adenomyosis development by upregulating HOXD9 to enhance endometrial epithelial cell migration and subsequent fibrosis. Subtype-specific APP expression may underlie distinct clinical manifestations between adenomyosis subtypes. In summary, we propose the following working model: APP upregulation drives HOXD9 nuclear entry to enhance endometrial epithelial cell migration and fibrosis, thereby promoting adenomyosis development. Differential APP expression across subtypes may dictate the magnitude of this migratory-fibrotic response, giving rise to clinical heterogeneity. Targeting the APP–HOXD9 axis may attenuate both cell migration and tissue fibrosis, offering a potential therapeutic strategy to relieve symptoms and halt disease progression. The direct transcriptional targets of HOXD9 are currently being validated, and these findings will be reported elsewhere. We generated a plasma cfDNA 5hmC ten-marker signature that distinguished adenomyosis from controls with an AUC of 0.91 (sensitivity 0.88, specificity 0.87), outperforming CA-125, CA19-9, 15 and trans-vaginal ultrasonography, and enabling molecular stratification of intrinsic versus extrinsic subtypes. Mechanistically, we found the APP–HOXD9 axis as a dual driver of endometrial cell migration and collagen-dependent fibrotic outgrowth; pharmacological blockade of APP concurrently suppressed both processes in vivo, providing an immediately testable, disease-modifying therapeutic strategy. Our study has several strengths. First, the discovery cohort (n = 97) comprised matched tissue–plasma pairs, enabling direct alignment of circulating 5hmC signals with lesion biology. Second, genome-wide 5hmC profiling was performed by 5hmC-Seal on both cfDNA and tissue-derived genomic DNA from matched samples, enabling direct comparison of epigenetic signatures between plasma and lesion while circumventing leucocyte contamination inherent to array-based methylation assays. Third, the cross-validated ten-marker signature exceeds the diagnostic accuracy of current biomarkers and clinical guidelines, and reliably subtypes intrinsic versus extrinsic adenomyosis. Fourth, multi-omics integration, primary cell assays, and orthotopic mouse models functionally validate APP as a subtype-specific driver; targeting APP simultaneously inhibits migration and fibrosis, offering a first-in-class intervention. Finally, raw sequencing data, analytical code, and model coefficients are publicly available to ensure reproducibility and facilitate external validation. Several limitations also need to be noted. First, the single-centre, retrospective design with modest sample size requires prospective, multi-centre confirmation. Second, the exclusively Chinese cohort limits generalisability to other ethnicities. Third, subtypes were not represented. Fourth, the temporal discrepancy between adenomyosis cases (2020–2021) and healthy plasma controls (2016–2019), although mitigated by rigorous age and menstrual phase matching, may introduce unmeasured confounders. Fifth, adenomyosis frequently coexists with uterine fibroids, endometriosis, and ovarian endometriomas; whilst stringent exclusion criteria (fibroids >5 cm) and tissue–plasma intersection enhance specificity, residual confounding from comorbidities cannot be fully excluded. Finally, the absence of healthy tissue biopsies precluded direct tissue-to-tissue comparison between healthy and diseased endometrium, necessitating benign myometrial specimens as tissue controls. To address these limitations, we have initiated a prospective, multicentre, pan-Asian/European trial (target n ≥ 1000) integrating standardised 3-T MRI protocols, single-nucleus 5hmC sequencing, and federated learning to externally validate, recalibrate, and ultimately translate the signature into routine clinical practice.

Data access and verification: LZ and X-TH had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: LZ, X-TH, X-RG. Acquisition of data: X-TH, LZ, H-YC. Analysis and interpretation of data: LZ, LC, H-YG, JL. Drafting of the manuscript: LZ, X-TH. Critical revision of the manuscript for important intellectual content: JL, H-YC, C-LS. Statistical analysis: LZ, LC. Study supervision: X-TH, X-RG. All authors read and approved the final version of the manuscript.

Adenomyosis is a common disease in women of childbearing age, with a prevalence of 5%–70%. 1 , 2 It significantly affects women's health and quality of life. 3 Timely diagnosis and intervention may be associated with improved clinical outcomes, including reduced hysterectomy rates. 4 , 5 At present, surgical pathology remains the gold standard for the diagnosis of adenomyosis. However, it is invasive and requires a considerable number of tissue biopsies, which are often difficult to obtain in advanced-stage patients, necessitating repeated, time-consuming and painful procedures. Therefore, the diagnosis of adenomyosis is increasingly dependent on non-operative methods such as clinical symptoms, imaging, and serological examination. However, the symptoms of adenomyosis are non-specific and about 30% of patients are asymptomatic. 6 Studies have shown that the sensitivity and specificity of transvaginal ultrasound (TVS) and MRI for diagnosing adenomyosis are highly variable., 7 , 8 , 9 , 10 , 11 and the sensitivity is as low as 12% when examined by radiologists and non-gynaecologists. 12 , 13 Although carbohydrate antigen 125 (CA125) and carbohydrate antigen 199 (CA199) have a potential role in the diagnosis of adenomyosis, they lack specificity in distinguishing adenomyosis from other diseases. 14 , 15 Adenomyosis is a heterogeneous disease. Previous studies have classified adenomyosis into four subtypes according to the position of adenomyosis lesions and junctional zone structural changes on MRI. 16 Subtype I (intrinsic) adenomyosis consists of patients whose MRI revealed that the adenomyosis had developed in direct connection to the thickened junctional zone and that healthy muscular structures were preserved outside the adenomyosis. Subtype II (extrinsic) adenomyosis consists of those patients whose MRI revealed that adenomyosis was located in the outer layer of the uterus, that the junctional zone was preserved, and that healthy muscular structures were preserved in between the adenomyosis and the junctional zone. Subtype III (intramural) adenomyosis consists of patients whose MRI showed adenomyosis residing without any geographic relationship to the junctional zone or the serosa. Patients who had an MRI that did not meet any of the categorisation criteria mentioned earlier were assigned to subtype IV (indeterminate) adenomyosis. 13 Distinct subtypes may underlie disparate pathogenic pathways, and are associated with variations in clinical presentation and treatment response. 16 , 17 , 18 , 19 However, molecular biomarkers distinguishing adenomyosis subtypes remain poorly characterised. Due to the limitations of current diagnostic and classification methods for adenomyosis, a non-invasive, accurate and early serological approach is urgently needed. Liquid biopsy has gained increasing attention among clinicians and patients owing to its non-invasive nature, sampling convenience, and capacity for dynamic monitoring. 20 , 21 , 22 , 23 Meanwhile, accumulating evidence indicates that circulating cell-free DNA (cfDNA) derived from plasma is associated with disease initiation and progression, and exhibits promising diagnostic and predictive performance relative to conventional clinical indicators. 24 , 25 , 26 , 27 , 28 5hmC (5-hydroxymethylcytosine) is an important epigenetic marker closely related not only to organ development, including that of the brain, 29 but also to the onset and progression of human diseases such as neurodegenerative diseases 30 , 31 and cancers. 32 , 33 In the human genome, 5-methylcytosines (5mCs) are dynamic and reversible, 34 and can be oxidised into 5-hydroxymethylcytosines (5hmCs) through the Ten–Eleven Translocation (TET) enzymes in an active DNA-demethylation process. 34 , 35 , 36 , 37 In addition, recent studies indicate that 5hmC enrichment in gene-body regions promotes gene transcription 35 and exhibits a tissue-specific distribution pattern. 38 , 39 , 40 Therefore, we hypothesised that 5hmC profiles might have potential value in adenomyosis diagnosis and classification. In this study, we collected plasma from 46 healthy persons, 51 patients with adenomyosis (26 with matched tissue), together with 21 normal uterine tissues. We used the 5hmC-Seal technique 41 to obtain genome-wide 5hmC profiles of plasma cfDNA and tissue genomic DNA. Our results revealed that patients with adenomyosis and healthy individuals had distinct 5hmC profiles, and that 5hmC markers selected by bioinformatics and machine learning algorithms may serve as potential diagnostic tools for adenomyosis. Meanwhile, there were significant differences in the distribution of 5hmC were observed among diverse types of patients with adenomyosis. Amyloid precursor protein ( APP ) expression was additionally examined in tissues from 38 patients with adenomyosis (21 extrinsic, 17 intrinsic) and 21 controls. The role of APP , one of the screened 5hmC markers, in adenomyosis development was also explored using endometrial epithelial cell lines and tissue samples from 20 affected patients. The results suggest that APP may be associated with adenomyosis progression, possibly by promoting epithelial cell migration.

The authors declare that they have no conflict of interests.