Deciphering brain metastasis in epithelial ovarian cancer: multimodal analysis and potential biomarkers

Abstract

Epithelial ovarian cancer (EOC) remains the most lethal gynaecological malignancy in developed countries, with recurrence and drug resistance posing significant clinical challenges. Brain metastases (BM) from epithelial ovarian cancer, once rare, are an increasing phenomenon and are characterised by a dismal prognosis. To explore the molecular underpinnings of BM in EOC, we conducted a multimodal genomics and transcriptomics analysis of matched primary tumour and brain metastases samples from a retrospective cohort. Our findings revealed high genomic concordance between primary tumour (PT) and BM, with alterations in key pathways such as MYC (MYC Proto-Oncogene, bHLH Transcription Factor) targets, extracellular matrix remodelling, and inflammatory signalling characterizing the BM. AFP (Alpha-fetoprotein) and GFAP (Glial Fibrillary Acidic Protein) emerged as potential biomarkers from the primary lesion for BM onset, while network analysis identified MET (MET Proto-Oncogene, Receptor Tyrosine Kinase), GDF15 (Growth Differentiation Factor 15), and S100A9 (S100 Calcium Binding Protein A9) as candidate mediators of tumour-brain crosstalk. These results offer new insights into EOC brain tropism, highlighting potential targets for therapeutic intervention and personalized patient management in the precision oncology era.

Introduction

Epithelial ovarian cancer (EOC) is the deadliest gynaecological malignancy in developed countries. Despite the improvement in the therapeutic strategies, around 80% of patients experience recurrence and ultimately develop drug resistance. The recurrences are frequently confined to the abdominal cavity, where the peritoneum is the most common involved site. Only 8% of the recurrent patients show isolated distant metastases without concomitant peritoneal or nodal disease¹. Brain metastases (BM) from EOC are late-stage, uncommon disease manifestations usually occurring in the context of a widely disseminated disease. The reported incidence of BM in EOC patients ranges from 0.3 to 11% in different series^2–4. However, the incidence of BM is expected to rise in the coming years due to extended survival conferred by targeted therapies, such as poly (ADP-ribose) polymerase inhibitors (PARPi)⁵. While these therapies improve systemic disease control, they may inadvertently increase the likelihood of BM development by allowing time for the seeding and growth of distant metastases. The occurrence of BM typically indicates an adverse prognostic factor, with a median survival of around 8–12 months after their diagnosis⁶.

Treatment options, besides corticosteroids and antiepileptic drugs, include systemic chemotherapy, radiotherapy and surgical resection. However, until now, there is a lack of a common agreement on the therapy^7,8 and the identification of effective biomarkers of brain involvement and prognosis^9–11.

Although previous studies have explored the genomic landscape of matched primary tumours and brain metastases¹², limited data exist for EOC-specific populations, particularly at the gene expression level.

This represents a relevant unmet need, especially in the era of personalized medicine, where precise prognostication is essential for empowering clinicians to make well-informed decisions regarding treatment and patient care from the very beginning of the disease¹³.

By conducting a multimodal characterisation on a retrospective cohort of primary EOC samples (PT) and their matched BM, we established that BM samples harbour peculiar mutational and gene expression features possibly suitable for targeted therapies, determining that PT samples harbour distinct molecular features that predict the onset of BM.

We identified a distinct brain-metastatic transcriptional profile in EOC with BM, characterized by extracellular matrix remodelling, neuronal–glial phenotypic adaptation, metabolic reprogramming, and heightened proliferative signalling. BM exhibited unique gene expression changes, including upregulation, among others of AFP (Alpha-fetoprotein), GFAP (Glial Fibrillary Acidic Protein), and MET (MET Proto-Oncogene, Receptor Tyrosine Kinase), suggesting neuronal-like adaptation and immune evasion. Gene Set Enrichment Analysis (GSEA) highlighted key pathways such as MYC (MYC Proto-Oncogene, BHLH Transcription Factor), E2F (E2F Transcription Factor), and mTORC1 (Mammalian target of rapamycin complex 1) signalling, indicating that EOC cells co-opt brain-specific mechanisms to survive and thrive in the central nervous system (CNS), offering potential biomarkers and therapeutic targets for brain-directed metastasis.

Results

Patients and samples

On a total of 4978 EOC patients treated at Fondazione Policlinico Gemelli IRCCS (FPG) between January 2000 and December 2021, a subset of 111 patients experienced BM (mEOC – 2.2% of total), as previously described (Fig. 1A)⁵. In this time frame ten mEOC (mEOC 1–10) patients (0.9% of mEOC) underwent BM surgical resection. For these patients, formalin-fixed, paraffin-embedded (FFPE) tissue samples were collected from both the primary tumour (mEOC_PT) and the corresponding brain metastases (mEOC_BM). The clinical characteristics of the entire population are reported in the Table 1. The median mEOC patients’ age at the diagnosis of EOC was 54 years. Eight patients were affected by high-grade serous ovarian cancer (HGSOC), one by clear cells ovarian carcinoma (CCOC) and one by endometrioid epithelial ovarian cancer (EEOC). Supplementary Data 1 shows the digitized pathological slides of the PT and the corresponding BM. All patients received first-line chemotherapy with carboplatin and paclitaxel, and all of them were platinum-sensitive (relapse after more than 6 months since the last platinum-based chemotherapy). One patient received maintenance treatment with bevacizumab after the first line of adjuvant chemotherapy. Four patients received platinum-based chemotherapy followed by maintenance therapy at the time of platinum-sensitive relapse (two bevacizumab and two PARPi). The other patients did not receive maintenance treatments in any setting. The median time frame between the primary EOC diagnosis and the BM was 29.5 months. In six cases, BM were located in the cerebrum, and in four cases, in the cerebellum. In Supplementary Data 1, the radiological scans of BM are shown.

Fig. 1. Cohort selection, survival trajectories, and genomic landscape in brain-metastatic EOC.

a Consort Diagram of the study population. b Swimmer plot. Each horizontal bar represents one patient. HGSOC High-grade serous ovarian cancer, CCOC Clear cells ovarian carcinoma, EEOC Endometrioid epithelial ovarian cancer, WT wild type. c Oncoplot illustrates the mutational profiles of patients from mEOC1 to mEOC10, with both primary (P) and metastatic (M) tumour samples represented for each patient. The top annotation provides clinical information, including age at diagnosis, tumour stage and tumour clonality. Tumour Mutational Burden (TMB) and Microsatellite Instability (MSI) are depicted as bar plots for both primary and metastatic tumours. The colour-coded legend below classifies TMB as High, Medium or Low and MSI as Stable. Genomic alterations are categorized into single-nucleotide variants (SNV) and small insertions/deletions (InDels), Copy Number Variations (CNV) and Gene Fusions. SNV/InDel mutations are further annotated based on clinical relevance, with asterisks indicating Tier I and Tier II classifications, while colour coding differentiates mutation types. CNV are stratified into gain/loss and gain/loss supported categories, represented by distinct colours. *Created with Biorender.

Table 1.

Clinical, pathological, and survival characteristics of patients included in the study and matched controls without BM

Characteristics	Cases	Controls	Controls
	mEOC n = 10	other_EOC_PT n = 9	other_EOC_METS n = 9
Mean age, years
	53.7	57	47.4
Histotype
EEOC	1	0	1
HGSOC	8	8	8
CCOC	1	0	0
Dedifferentiated	0	1	0
BRCA status
Wild type	9	4	5
BRCA1 mutated	0	3	3
BRCA2 mutated	1	2	1
2016 FIGO stage
IA	1	0	0
IIIA2	2	0	0
IIIB	0	1	1
IIIC	8	7	8
IVB	0	1	0
Status at follow-up
DOD	9	4	2
AWD	0	3	4
NED	1	2	3

EEOC Endometrioid epithelial ovarian cancer, HGSOC High grade serous ovarian cancer, CCOC Clear cells ovarian carcinoma, DOD Dead of disease, AWD alive with disease, NED No evidence of disease.

BM treatment differed between patients, including surgery, radiotherapy and chemotherapy. Nine out of ten patients died from disease (median overall survival 45 months, min-max: 32–96). The follow-up of the only alive patient is 64 months, and she is currently free of disease (Fig. 1B). Though numerically small, this paired-sample resource is uniquely poised to unravel the molecular evolution of brain-tropic EOC and to validate our multimodal findings in a clinically impactful setting.

To elucidate the gene expression changes underlying EOC progression to and establishment within the brain, we selected different control populations. First, four healthy brain samples (ctrl_BRAIN) were collected. Median age of the ctrl_BRAIN patients was 55 years, two were male and two were female; two samples came from hippocampus, one from the cerebral cortex and one from the cerebellum. Second, two healthy ovary samples (ctrl_OVARY) were collected.

We then selected two distinct EOC populations with comparable characteristics to the mEOC cohort. (Table 1) Specifically, primary EOC samples from patients with long-term survival (>5 years) who did not develop BM (other_EOC_PT). Similar to the mEOC cohort, the mean age of the cohort was 57 years. Eight patients were affected by HGSOC, one by a dedifferentiated EOC. All patients were treated with debulking surgery followed by first-line chemotherapy with carboplatin and paclitaxel. Four patients out of nine received maintenance therapy (one with PARPi and the remaining with bevacizumab). All patients except for one had a platinum-sensitive relapse, median overall survival (OS) was 124 months (min-max: 102-NE).

We next selected non-brain metastatic relapse samples (other_EOC_METS), comprising four lymph node metastases, three bowel metastases, one hepatic metastasis, and one splenic metastasis (Supplementary Data 2). All the samples were collected at the time of surgery performed for a platinum-sensitive relapse. All patients have received previous debulking surgery, carboplatin and paclitaxel chemotherapy and maintenance treatment with bevacizumab (except for two patients). Median OS was not reached (95% Confidence Interval, 110-NE). Further clinical characteristics are in the Table 1.

Healthy lymph nodal, hepatic and intestinal tissues (ctrl_HEALTHY1-3) were used as controls.

Samples from other_EOC_PT and other_EOC_METS populations were not paired. This multi-group design was intended to distinguish transcriptional features unique to brain-tropic EOC from generic metastatic or baseline EOC traits, thus ensuring a clear view of brain-specific adaptations.

DNA genomic profiling

Somatic sequencing on mEOC_PT and matched mEOC_BM showed very high concordance of single nucleotide variants (SNV) (Fig. 1C).

Analysis of SNV and microsatellite regions allowed to determine that the vast majority of mEOC had a monoclonal origin and maintained a stable tumour mutational burden (TMB) and microsatellite instability (MSI) status (Fig. 1C and Supplementary Data 3A uploaded higher resolution image).

A pathogenic TP53 variant was identified in 70% of patients, in all cases shared between PT and BM. PIK3CA was altered in 20% of patients, specifically patients affected by EEOC and CCOC. Other Tier I and Tier II variants resulted to be either patient’s private (ARID1A - AT-rich interaction domain 1A) or PM/BM tumour’s private (RET - Ret Proto-Oncogene, FAT - FAT Atypical Cadherin, TET2 - Tet Methylcytosine Dioxygenase 2). A comprehensive summary of all genes, including full names, deregulation type, GO IDs and related terms, the most representative reference and the assigned functional cluster, is provided in Supplementary Data 4.

Copy number variants (CNV), confirmed 16 chromosomal alterations (6 copy number gains and 10 copy number losses), supported by concordant changes in mRNAs expression (upregulated genes in gains, downregulated in loss) (Supplementary Data 3B). CNV were more frequent in mEOC_BM, potentially reflecting late-arising genomic alterations that promote CNS colonization. Five of ten samples pairs lacked CNV altogether, while mEOC-4 showed losses in B2M (Beta-2-Microglobulin) and MGA (MAX Dimerization Protein), that were not validated at transcriptional level.

To conclude, although SNVs remained highly conserved between mEOC_PT and mEOC_BM, CNV alterations were more prevalent in the metastatic samples. Clonal evolution analysis further supported a monoclonal origin shared across primary and brain lesions.

RNA gene expression analysis

We performed a transcriptomic comparison of mEOC_PT and mEOC_BM, along with well-defined control tissues, to elucidate the gene expression changes underlying EOC progression to and establishment within the brain. As RNAseq platform, we employed a 3’-end RNAseq (or 3’ Digital Gene Expression – 3’DGE) approach, which is specifically suitable for highly degraded samples, allowing to fully capture the transcriptome profiling by quantifying the terminal portion of any mRNA¹⁴.

A specific transcriptional signature defines the EOC brain metastasis

We initially employed 3’DGE and performed a cross-comparative analysis of differentially expressed genes (DEGs) across mEOC_BM, mEOC_PT, ctrl_BRAIN, and other_EOC_METS (Fig. 2A). Additionally, the top five upregulated genes for each comparison (ranked by fold change among significant DEGs) are displayed in the adjacent panels. The full list of DEGs, including Symbol, p-adj, log2FoldChange is displayed in Supplementary Data 5. To establish a reference profile of tumour-specific alterations, we first compared mEOC_BM against ctrl_BRAIN. This analysis identified 2845 upregulated and 1664 downregulated genes in mEOC_BM (Fig. 2A, right). While many of these changes reflect lineage-specific expression, the top five upregulated genes RAB25 (RAB25, Member RAS Oncogene Family), PI3 (Peptidase Inhibitor 3), MISP (Mitotic Spindle Positioning), TMPRSS4 (Transmembrane Serine Protease 4), and UGT2B7 (UDP Glucuronosyltransferase Family 2 Member B7), suggest CNS-specific adaptations. Their selective elevation in BM suggests that EOC cells exploit these pathways to cope with the restrictive CNS microenvironment.

Fig. 2. A specific transcriptional signature defines the EOC brain metastasis.

a Venn diagram illustrating intersected differentially expressed genes (DEGs) uniquely upregulated (red) and downregulated (blue) in epithelial ovarian cancer brain metastases (mEOC_BM) compared with primary ovarian tumours (mEOC_PT), non-CNS metastatic sites (other_EOC_MET), and healthy brain tissue (ctrl_BRAIN). Highlighted are top 5 genes specifically upregulated in each pairwise comparison. b Boxplots depicting normalized gene expression (log2 CPM counts) across the entire cohort (mEOC_BM, mEOC_PT, ctrl_BRAIN, ctrl_OVARY, otherEOC_PT, other_EOC_MET, ctrl_HEALTHY) for selected candidate genes consistently enriched in brain metastases: SQLE, ADSL, LAMC3, ATAD5, and LINC00562. Statistical comparisons were performed using Student’s t test against mEOC_BM, with significance thresholds annotated (*p < 0.05; **p < 0.01). c Dot plots showing Gene Set Enrichment Analysis (GSEA) results for hallmark pathways significantly enriched in mEOC_BM relative to ctrl_BRAIN (left), other_EOC_MET (right, top), and matched mEOC_PT (right, bottom). Pathways are ranked by normalized enrichment score (NES). Dot size reflects the number of contributing genes; dot colour indicates statistical significance (adjusted p-value).

Next, to distinguish brain-specific drivers beyond generic metastatic traits, we compared mEOC_BM to other_EOC_MET (Fig. 2A, left). Among the 164 significantly upregulated genes, AFP exhibited the highest induction, followed by GFAP, ELAVL3 (ELAV Like RNA Binding Protein 3), and HCN2 (Hyperpolarization Activated Cyclic Nucleotide Gated Potassium and Sodium Channel 2). Although GFAP, ELAVL3, and HCN2 could partially stem from surrounding neuronal or glial cells, their recurrent presence with high expression levels suggests that EOC cells may acquire or mimic neuroglial traits to foster survival in the CNS. Meanwhile, AFP, which we later observe in primary EOC as well, may exert immunoregulatory or neurotrophic influences, strengthening the concept of an early CNS-compatible phenotype.

By comparing mEOC_BM with matched mEOC_PT, we discovered 33 upregulated and 115 downregulated genes (Fig. 2A, bottom). Late-stage transcriptomic shifts during CNS colonization were evident, with GFAP emerging as the most differentially expressed gene. Integrating these data, we defined a core brain-metastatic gene signature, SQLE (Squalene Epoxidase), ATAD5 (ATPase Family AAA Domain Containing 5), LAMC3 (Laminin Subunit Gamma 3), ADSL (Adenylosuccinate Lyase), and LINC00562 (Long Intergenic Non-Protein Coding RNA 562), consistently overexpressed in mEOC_BM (all padj < 0.05) (Fig. 2B). Notably, LAMC3, involved in cortical lamination, may facilitate CNS-specific ECM remodelling and metastatic engraftment¹⁵.

To investigate how these molecular changes converge at a pathway level, we performed GSEA on each comparison: mEOC_BM vs. ctrl_BRAIN; mEOC_BM vs. other_EOC_METS; and mEOC_BM vs. mEOC_PT (Fig. 2C).

In the comparison of mEOC_BM with healthy brain tissue, E2F targets, G2M checkpoint, interferon gamma response and oestrogen response late. Moreover, we noted an enrichment in the MYC targets, signalling a transition toward hyperproliferation and heightened transcriptional output. Concomitant upregulation of stress-related/inflammatory programs (TNFα signalling) suggests an immune-evading or pro-inflammatory shift suitable for metastatic colonization.

By contrasting mEOC_BM with other_EOC_METS, marked emphasis on spermatogenesis, mTORC1_SIGNALING, E2F targets, and G2M checkpoint emerged. Finally, comparing mEOC_BM with mEOC_PT confirmed a trend toward persistent activation of MYC, E2F pathways, G2M checkpoint, and spermatogenesis in metastases, highlighting the final transcriptional refinements EOC cells acquire upon CNS invasion. Altogether, these transcriptomic data reveal a brain-tropic phenotype underpinned by ECM remodelling, tumour-glial interaction, metabolic reconfiguration, and heightened proliferative signals.

Early transcriptional signature predicts ovarian cancer predisposition to brain metastasis

To determine whether mEOC_PT harbour transcriptional clues indicating future brain metastatic competence, we employed 3’DGE and assessed DEGs relative to both ctrl_OVARY and other_EOC_PT (Fig. 3A). For each comparison, the adjacent panels show the top five upregulated genes (selected by the highest fold change among significant DEGs) to summarize the strongest early transcriptional drivers.

Fig. 3. Early transcriptional signature predicts ovarian cancer predisposition to brain metastasis.

a Venn diagrams representing the intersection of significantly upregulated (red) and downregulated (blue) genes in primary epithelial ovarian cancer samples predisposed to brain metastasis (mEOC_PT) when compared to healthy ovarian tissue (ctrl_OVARY) and primary ovarian tumours without brain metastatic potential (otherEOC_PT). Highlighted are the top five genes uniquely elevated in each comparison. Statistical comparisons were performed using Student’s t test against mEOC_PT, with significance thresholds annotated (*p < 0.05; **p < 0.01). b Boxplots illustrating normalized gene expression (log2 CPM counts) for the 13 intersected genes (AP1B1, SERINC2, TXNL4A, SLC29A4, AFP, NUDT14, NDUFA4L2, FGFR3, TFCP2L1, TSPAN1, GDF15, S100P, NCCRP1) across mEOC_BM, mEOC_PT, ctrl_BRAIN, ctrl_OVARY, other_EOC_PT, other_EOC_METS. c Dot plots showing Gene Set Enrichment Analysis (GSEA), displaying significant KEGG pathways enriched in mEOC_PT compared with ctrl_OVARY (left) and otherEOC_PT (right). Dot size represents the number of enriched genes, while colour intensity indicates statistical significance (adjusted p-value).

Comparisons of mEOC_PT with ctrl_OVARY revealed upregulation of FXYD3 (FXYD Domain Containing Ion Transport Regulator 3), ESRP1 (Epithelial Splicing Regulatory Protein 1), RAB25, BICDL2 (BICD Family Like Cargo Adaptor 2), and CKMT1B (Creatine Kinase, Mitochondrial 1B) (Fig. 3A, left).

In contrast, comparisons of mEOC_PT with other_EOC_PT identified S100P (S100 Calcium Binding Protein P), NCCRP1 (NCCRP1, F-Box Associated Domain Containing), PIGR (Polymeric Immunoglobulin Receptor), AFP, and DNER (Delta/Notch Like EGF Repeat Containing) as the most highly upregulated (Fig. 3A, right).

The intersection of these comparisons uncovered 13 genes uniquely upregulated in mEOC_PT, AFP, S100P, GDF15 (Growth Differentiation Factor 15), NCCRP1, AP1B1 (Adaptor Related Protein Complex 1 Subunit Beta 1), NUDT14 (Nudix Hydrolase 14), NDUFA4L2 (NDUFA4 Mitochondrial Complex Associated Like 2), FGFR3 (Fibroblast Growth Factor Receptor 3), TFCP2L1 (Transcription Factor CP2 Like 1), TSPAN1 (Tetraspanin 1), SLC29A4 (Solute Carrier Family 29 Member 4), TXNL4A (Thioredoxin Like 4A), and SERINC2 (Serine Incorporator 2), collectively suggesting a “premetastatic” program favouring CNS colonization (Fig. 3B). To contextualize these individual gene findings, we conducted GSEA on mEOC_PT relative to ctrl_OVARY and other_EOC_PT (Fig. 3C). In the mEOC_PT vs. ctrl_OVARY comparison, pathways associated with neuroactive ligand-receptor interaction and ECM interaction were prominently enriched, hinting that certain EOC primaries are already primed for microenvironmental cues critical to metastasis. Among these, the neuroactive ligand–receptor interaction gene set notably emerged, suggesting that EOC cells may engage in neuronal-like signalling or receptor-mediated communication, potentially reflecting early adaptive traits that later contribute to tumour–neural interactions observed in advanced brain metastases.

In the mEOC_PT vs. other_EOC_PT comparison, neuroactive ligand-receptor interaction reappeared, affirming that a subset of primary tumours expresses an ECM-focused, immune-evasive, or neuroactive phenotype absent in non-brain-metastatic EOC. This synergy of ECM remodelling, stress/immune pathways, and neuronal-like interactions may represent a cohesive “premetastatic signature,” enabling EOC cells to more readily seed distant organs such as the brain.

Ligand–receptor interactions predict ovarian cancer adaptation to the brain microenvironment

Building upon our observation of neuroactive ligand–receptor enrichment in mEOC_PT, we next sought to identify membrane-bound or secreted factors potentially relevant for EOC cells’ adaptation to the CNS. Pairwise comparisons of upregulated genes from mEOC_PT vs. other_EOC_PT and mEOC_BM vs. other_EOC_METS identified 15 transcripts consistently elevated in both the primary site and brain lesions (Fig. 4A). Such overlap suggests a recurrent signaling axis that arises early and persists into overt brain colonization.

Fig. 4. Ligand–receptor interactions predict ovarian cancer adaptation to the brain microenvironment.

a Venn diagram depicting upregulated DEGs in mEOC_PT vs. other_EOC_PT and mEOC_BM vs. other_EOC_METS. b Ligand–receptor network of overlapping upregulated genes, highlighting nodes (ligands in green, receptors in blue, dual-function genes in red): MET, GDF15, and S100A9 as key hubs. c Boxplots of MET and S100A9 (log2 CPM counts) illustrating higher expression in mEOC_PT and mEOC_BM relative to other groups.

To further characterize these transcripts, we applied an agnostic filtering strategy: genes were retained if annoted with GO terms for plasma membrane [GO:0005886] or extracellular space [GO:0005615]) and subsequently classified as ligands or receptors using the OmniPathR (v3.14.0) database. Next, we further cross-referenced these candidates with CNS-related GO terms as nervous system development [GO:0007399] and synaptic transmission [GO:0007268]. Through this multi-step filtering, we identified three candidate ligand/receptor genes consistently upregulated across tumour stages: MET, GDF15, and S100A9. Notably, GDF15 was already highlighted among intersected genes in the premetastatic signature (Fig. 3B). The network representation (Fig. 4B) depicts MET, linked here to ECM-modulatory genes such as LRG1 (Leucine Rich Alpha-2-Glycoprotein 1), TIMP3 (TIMP Metallopeptidase Inhibitor 3), DCN (Decorin), and ADAM17 (ADAM Metallopeptidase Domain 17), emphasizes the receptor’s known roles in cell motility and invasion, aligning with earlier GSEA findings implicating ECM receptor interactions in EOC primaries. GDF15, in association with its receptor GFRAL (GDNF Family Receptor Alpha Like), a pairing implicated in stress response and a metabolic reprogramming axis relevant for hypoxic and nutrient-limited CNS niche. We next examined the expression levels of MET, GDF15 and S100A9 across mEOC_PT, mEOC_BM, other_EOC_PT, and other_EOC_METS (Fig. 4C). All genes were significantly upregulated in mEOC_PT and mEOC_BM compared to their respective control groups (DESeq2 padj < 0.05).

Altogether, these results nominate MET, GDF15, and S100A9 as candidate ligand–receptor nodes that underlie EOC progression into the CNS. Their upregulation from primary tumours to brain-metastatic lesions highlights both biomarker potential for predicting CNS risk and novel therapeutic avenues to intercept metastatic dissemination.

Discussion

Here, we provided a multimodal analysis of BM derived from EOC, through a comprehensive genome profiling and a gene expression analysis of matched PT and BM samples. Although BM remain relatively infrequent, their emergence is clinically relevant, reinforcing the need for early predictive markers and actionable targets.

For patients with isolated BM relapse, surgery remains a viable therapeutic option. However, approximately 50% of these patients are deemed ineligible due to concomitant extracranial disease, anatomical inaccessibility, or poor overall clinical status¹⁶. Consequently, the acquisition of matched primary and metastatic tissue is technically and logistically challenging. In this context, our paired-sample cohort, though limited in size, represents a valuable and rare resource for dissecting the molecular evolution of brain-tropic EOC. To the best of our knowledge, this is the largest and only cohort of multimodal analysis performed on matched PT and BM in EOC patients. Previous studies have identified possible genomic signatures of multi-organ diffusion¹⁷. Nevertheless, that study was limited to DNA-only analysis, whereas we identified major patterns primarily at the RNA level.

Our findings demonstrate that EOC PT and matched BM exhibit many similarities in terms of genomic variants and gene expression. Nonetheless, certain pathways associated with cell cycle progression, MYC targets, cellular growth, and ECM interaction shed light on the mechanisms underlying brain colonization and metastatic fitness in the neuronal microenvironment.

It has been previously shown that EOC BRCA mutated patients have a higher incidence of BM (3% vs 0.6% of BRCA-wild type patients)^17,18. However, in our cohort, no patient showed a BRCA mutation, nor other somatic variants that may suggest a germline¹⁹ origin. This, together with the almost superimposable genomic profile, indicates that biological mechanisms of adaptation to CNS, rather than germline mutations, could be related to the BM in the EOC population.

Our results identify potential biomarkers of BM onset. In particular, AFP emerged as upregulated both in primary EOC that metastasized to the brain and in brain healthy tissues. AFP is a protein predominantly expressed during fetal development, with serum levels significantly decreasing after birth as AFP gene expression is downregulated. In adults, AFP is widely recognized as a biomarker for ovarian germ cell tumours. Interestingly, elevated AFP levels have also been reported in at least four distinct neurodegenerative disorders. These disorders, inherited as autosomal recessive traits, are characterized by a combination of cerebellar ataxia, abnormal ocular movements, and peripheral neuropathy. This may represent a potential link between AFP expression and nervous system²⁰, still unexplored in cancer. The consistent overexpression of AFP and additional candidates such as TFCP2L1 and ESRP1 provide a foundation for developing predictive biomarkers and highlight early molecular divergence that may inform risk stratification and therapeutic intervention in EOC.

GFAP emerged as upregulated in mEOC_BM. This is a component of the glial cytoskeleton and a candidate biomarker of glial cell damage. Moreover, consistent with our data, it has been previously described as a serum biomarker of BM in patients with metastatic breast cancer²¹.

In BM, the hallmarks of the MYC target were altered. High expression of hallmarks of MYC targets have already been correlated with higher cell proliferation, higher stage, pathological grade, and worse subtype in a large population of breast cancer²². Moreover, MYC family genes are found overexpressed in more than half of all brain cancers, highlighting its role as one of the most important oncogenes. MYC proteins are involved in brain tumour initiation, maintenance and progression²³. LAMC3, which is crucial for cortical lamination, as shown by laminar disorganization in LAMC3 deficient mice, may also play a role in CNS‑specific extracellular matrix remodelling and metastatic engraftment, given the broader function of laminin γ‑chains in facilitating cell adhesion, migration, and metastasis²⁴. Moreover, it has been previously described in EOC progression²⁵. Crucially, in our data, the intersection of LAMC3 and MYC target genes within the MET signalling network, a pathway recognized for bolstering cell motility, introduces a compelling avenue for future research.

The network analysis prioritized MET, GDF15, and S100A9 as elevated both in mEOC_PT and mEOC_BM. S100A9 is a protein expressed in myeloid-derived cells involved in controlling immunological homoeostasis and cell metabolism. It is considered a tumour promoter, by influencing tumour genesis, progression, and metastasis²⁶. It has been reported that high S100A9 expression was associated with BM and shorter PFS in EGFR-mutated lung cancer patients treated with osimertinib. Moreover, targeting the S100A9–ALDH1A1–RA signalling axis reduced BM in mice models²⁷.

Some strengths and limitations must be acknowledged. We enlightened some of the mechanisms that could explain BM onset and potential druggable targets. First, the sample size in all comparisons was limited, and differential expression analyses did not account for potential confounders such as treatment history or tissue type. Second, no independent validation was performed, and brain regions were not matched between cases and controls, introducing substantial intertumor heterogeneity. Moreover, the patients included in the study received treatment over an extended period, when targeted therapies such as bevacizumab or PARPi were not routinely administered. Notably, PARPi pharmacokinetics demonstrated a possible brain penetration in mice, but with differences between molecules. Substantial functional and genomic alterations in the tumour have been recorded following exposure to these drugs, potentially leading to profound changes in its molecular profile^28–31. This may limit the applicability of our findings. Finally, spatial and temporal heterogeneity were not considered, but they should be acknowledged in future studies to gain a more comprehensive understanding.

In conclusion, this study provides novel insights into the molecular underpinnings of EOC BM, emphasizing the need for further research to validate these findings and explore their translational potential in improving clinical outcomes.

Methods

Patient cohort and samples selection

The present study was conducted following the guidelines of the Declaration of Helsinki (1984) and subsequent revisions. The methods were performed in accordance with relevant guidelines and regulations with the approval of the Ethical Committee of Fondazione Policlinico Universitario Agostino Gemelli IRCCS (protocol ID: METACER-OV, date of approval: 07/04/2022, ID number: 4865). We retrospectively collected clinical data from EOC patients treated at Fondazione Policlinico Gemelli IRCCS between January 2000 and December 2021. Inclusion criteria were as follows: diagnosis of BM confirmed by computed tomography scan, magnetic resonance imaging, and/or positron emission tomography conducted during follow-up, for monitoring treatment response, or in response to symptoms, with no prior brain-directed therapy. Exclusion criteria included a history of other malignancies or CNS diseases unrelated to brain metastases. Written informed consent for the publication of clinical and molecular data, including photos, was obtained from all patients or their legal representatives before they participated in the study. As a control for RNA-seq analysis, we selected comparable EOC populations who did not experience BM. Specifically, we collected the primary EOC formalin-fixed, paraffin-embedded (FFPE) samples of nine long-term survivor patients (>5 years) who did not experience BM (other_EOC_PT); and the FFPE of relapse samples from nine long-term survivor patients (>5 years) patients who experienced relapses at different sites than the brain (other_EOC_METS), specifically four were lymph node metastases, three were bowel metastases, one was a hepatic metastasis, and one a splenic metastasis. Moreover, healthy tissues of the brain obtained by autoptic samples (ctrl_BRAIN 1–4), ovary (ctrl_OVARY 1–2) and tissue corresponding to metastasis other than brain, namely lymph nodes, liver and bowel (ctrl_HEALTHY1-3) were collected. Figure 1A shows the study design.

Nuclei acids extractions and DNA genomic profiling

A comprehensive genomic profiling (CGP) assay was performed using Illumina TruSight Oncology 500 high throughput assay. After a careful evaluation by a dedicated pathologist, only samples showing ≥20% of tumour cellularity were processed. Nucleic acids were extracted using the AllPrep DNA/RNA FFPE Kit (Qiagen, Hilden, Germany). The quantity and quality of extracted DNA and RNA were evaluated using Qubit and Agilent TapeStation, respectively. Libraries were prepared with specific panels (TruSight Oncology 500 High-Throughput Assay, Illumina) and sequencing was performed with the Illumina Novaseq6000 platform using the 2 × 101 bp paired-end sequencing protocol. Raw sequencing data underwent quality control and bioinformatic analysis using Illumina Software TSO500 Dragen solid v2.5 and a pipeline within the Clinical Genomics Workspace software platform from Velsera. An evaluation of SNV, insertions/deletions (indels) and CNV in 523 genes, as well as known and unknown fusions and splicing variants in 55 genes was provided. It also estimates TMB and MSI.

SNV and indels underwent a multistep filtering process characterized by the application of four primary criteria.

Variant annotation was performed using Ensembl Variant Effect Predictor (VEP) (https://www.ensembl.org/info/docs/tools/vep/index.html), accession date: 1st of October 2024. Oncogenicity annotations originated from the OncoKB database and specifically referred to genetic mutations that drive the development of cancers or influence their response to treatments³².

Initially, variants with an allelic frequency exceeding 1%, as reported by the Genome Aggregation Database (GnomAD)³³, were excluded.

Subsequently variants with VAF under 2% were excluded.

In addition, variants with a detrimental impact on the protein, including frameshift, splice donor, stop gained, splice acceptor, start lost, in-frame insertion, and in-frame deletion, were preserved.

The assessment of CNV was conducted considering any level of variations in the number of copies, considering amplifications defined as ≥5 copies and homozygous deletions. The analysis does not include the detection of CNV involving single exons or complete gene deletions.

To enhance the rigour of our CNV calls, we employed transcriptomic data for validation. Specifically, for each cohort, mEOC_PT and mEOC_BM, RNA expression levels were used as an independent confirmation. Within each group, a gene was considered amplified if its transcript abundance exceeded the 75th percentile of that group’s expression distribution, and conversely, a deletion was confirmed if the expression was below the 25th percentile.

Clonal Evolution

To model tumour evolution, we employed Pyclone VI³⁴. This algorithm facilitates the inference of clonal architecture and the tracking of spatial and temporal patterns in tumour evolution. The results were subsequently visualized using R V. 4.3.2 packages including ggplot2, fishplot and clonevol. ClonEvol is a tool for the reconstruction and interpretation of clonal evolution³⁵. It uses the clustering of heterozygous variants identified using other tools as input to infer consensus clonal evolution trees and estimate the cancer cell fraction (also called clonal frequency) of the clones in individual samples.

RNA Library Preparation and Sequencing

Total RNA was micro-dissected and purified from paraffin-embedded tumour tissue with Maxwell® RSC RNA FFPE Kit (Promega), following manufacturer’ instructions and quantified using the Qubit 2.0 Fluorometer (Thermo Fisher Scientific). Libraries were prepared from 250 ng of total RNA using the 3′ Digital Gene Expression (3′DGE) mRNA-seq pipeline (NEGEDIA s.r.l.), which includes rigorous quality-control steps to minimize technical variability. Libraries were sequenced on an Illumina NovaSeq 6000 platform using a 100-cycle, single-end run configuration. Raw sequencing reads were quality-filtered using the BBtools software suite (specifically the bbduk utility) to remove low-quality bases and adapter contaminants. Cleaned reads were aligned to the human reference genome GRCh38 using the STAR aligner (v2.6). Gene-level counts were obtained using HTSeq-count using GENCODE annotations for GRCh38. The resulting raw counts were further filtered to remove lowly expressed genes (threshold: at least one count per million in at least two samples). Normalized expression levels were obtained via counts per million (CPM), and log2 transformations were applied where indicated. We employed a negative binomial model (DESeq2 v.1.46.0) with Benjamini–Hochberg p-adj correction for differential expression analysis.

Intersected Differential Expression analysis and gene annotation

To identify DEGs, we performed multiple pairwise comparisons as indicated in the text. Significant DEGs were defined as those with an absolute log2 fold-change (|log2FC|) ≥ 1 and adjusted p-value (padj) < 0.05. Venn diagrams were used to illustrate overlapping DEGs across specific comparisons. Additionally, the top five upregulated genes (ranked by fold-change among significant DEGs) are shown in the adjacent panels for each comparison. Gene annotation was performed using org.Hs.eg.db (v 3.20.0) and GO.db (v 3.20.0) via the AnnotationDbi interface in R (V 1.68.0). We mapped Ensembl IDs (from DEGs) to gene symbols and GO terms (BP, MF, CC). Subsequently, we refined annotations using a literature-based curation for each gene to tailor functions relevant to ovarian cancer metastasis. For all boxplot visualizations, statistical significance was assessed using Student’s t test comparing mEOC_BM or mEOC_PT to the respective groups, with thresholds indicated as p < 0.05 (*) and p < 0.01 (**).

Pathway Enrichment Analysis

GSEA was performed to identify hallmark pathways enriched in each comparison. Genes were ranked based solely on the log₂ fold change obtained from differential expression analysis. The pre‑ranked gene list was generated by ordering genes in descending order of log₂ fold change. Normalized enrichment scores (NES) and false discovery rate (FDR) values were calculated using the default settings of the GSEA software (Broad Institute). Significant pathways were defined as those with |NES|≥1.5 and FDR ≤ 0.05. Pathway analyses were carried out using the R package clusterProfiler (v4.14.4) in conjunction with the org.Hs.eg.db annotation database (v3.20.0). Hallmark gene sets were retrieved via the msigdbr package (v7.5.1). Additionally, KEGG pathway analysis was performed only for the mEOC_PT vs ctrl_OVARY and mEOC_PT vs other_EOC_PT comparisons to provide further insights into functional enrichments.

Ligand-receptor analysis

To systemically identify potential ligand–receptor genes potentially involved in EOC adaptation to the CNS, we adopted an agnostic filtering approach. We first screened differentially upregulated genes from mEOC_PT and mEOC_BM relative to their respective controls (other_EOC_PT and other_EOC_MET). From this set, we retained only transcripts annotated for membrane-associated or secreted proteins, with GO terms that specify membrane or extracellular localization (plasma membrane [GO:0005886], extracellular space [GO:0005615]). We then employed the OmniPathR (v.3.14.0) database to classify these filtered genes as ligands or receptors, focusing on those with known or putative with known signalling functions. To ensure potential relevance to CNS, we further intersected these candidates with GO terms associated with nervous system development [GO:0007399], synaptic transmission [GO:0007268]. Network visualization was conducted in Cytoscape (v.3.10.3), with nodes representing ligands/receptors and edges denoting curated interaction.

Author contributions

R.T., D.C., A.F., C.M., C.N., “Conceptualization, Writing - Original Draft”; R.T., D.C., S.D., C.N., “Methodology”; L.G., M.K., M.S, L.V., L.M., A.M., G.R., D.M., A.P., “Software”, “Formal Analysis” and “Data Curation”; M.B., F.S., “Project administration”; A.G., R.D., M.M., M.C., M.P., M.C.S. “Investigations”; F.C., V.I., D.C., A.F., C.N., S.D., C.M., G.S., “Supervision” and “Writing–review and editing”. All authors reviewed the manuscript.

Data availability

Due to European General Data Protection Regulation (GDPR) and institutional data protection policies, we are unable to publicly release the raw sequencing data (FASTQ files) because they contain potentially identifiable genomic information. However, to ensure transparency and reproducibility, we have made processed RNA-seq count matrices (CPM-normalized values). For researchers requiring access to the raw sequencing files, such data can be provided upon reasonable request, and the transfer will require the completion of a data transfer agreement with our institution in compliance with GDPR.

Competing interests

D.C. is founder, shareholder, and consultant of NEGEDIA S.r.l. G.S. is consultant for AstraZeneca, MSD and Covidien AG, a Medtronic company. G.S. is member of the speakers bureau/honoraria for Olympus Europa SE & CO. KG, Baxter Healthcare SA, Intuitive Surgical Inc., GlaxoSmithKline S.p.A. G.S. is member of the advisory board/advisory committee of TESARO Bio Italy S.r.l. Johnson & Johnson. A.F. has received grants/research support from AstraZeneca and MSD. A.F. is member of the advisory board of GlaxoSmithKline. A.F. is member of the speakers/honorary board of Covidien (Medtronic), GSK, pharmaand GmbH and AbbVie. A.F. is consultant for Oncoinvent AS. C.M. is consultant/advisory board member for Clovis, Pharmamar, GSK, AstraZeneca, AbbVie, pharmaand GmbH and MSD. C.M. has received travel accommodation from Pharmamar and Roche. C.N. is member of the speakers bureau/honoraria for AstraZeneca, GSK, MSD, GH and Illumina. C.N. is member of the advisory board/advisory committee of AstraZeneca and MSD. C.N. has received grants/research support from AstraZeneca and GH. C.N. has received travel accommodation from AstraZeneca, MSD and Illumina.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-025-01206-y.