Aim: Our previous studies identified three extracellular vesicle derived microRNAs (miR-92a-1-5p, miR-375 and miR-148a-3p) potentially associated with prostate cancer (PCa), particularly in advanced stages such as bone-metastatic PCa. This study aimed to evaluate their clinical diagnostic utility as a panel derived from plasma extracellular vesicles. Methods: A total of 49 treatment-naive participants, including 21 with benign prostatic hyperplasia (BPH), 15 with localized PCa, and 13 with bone-metastatic PCa, were enrolled. Plasma extracellular vesicles were isolated, and absolute quantification of the three miRNAs was performed using digital polymerase chain reaction (PCR). Their individual and combined diagnostic performance for distinguishing PCa from BPH and identifying bone metastasis was assessed. Results: Among the three miRNAs, miR-92a-1-5p and miR-375 showed a statistically significant difference in expression between PCa and BPH. The combined three-miRNA panel achieved an improved area under the curve (AUC) of 0.736, representing an enhancement over single markers and outperforming serum prostate-specific antigen (PSA) (AUC = 0.682) in this cohort. The panel yielded an AUC of 0.766 for detecting bone-metastatic PCa versus BPH, but showed limited ability to discriminate localized from metastatic disease. Conclusion: These preliminary findings suggest that the plasma extracellular vesicle-derived three-miRNA panel has auxiliary diagnostic value for differentiating prostate cancer from benign prostatic hyperplasia, and may serve as a supplementary tool to serum PSA.

Extracellular vesicles (EVs) can transport functional RNA, but the clinical development of EV-based RNA therapeutics has been constrained by the technical demands of ex vivo production and cargo loading. A recently reported strategy uses a plasmid to direct the formation of small interfering RNA (siRNA)-containing small EVs (sEVs) in vivo. Following administration, hepatocytes produce the therapeutic RNA and release it in association with endogenous sEVs, allowing systemic delivery to be achieved without isolating and reloading vesicles outside the body. This commentary examines the biological basis and potential therapeutic relevance of this approach. Particular attention is given to endogenous cargo loading, the use of genetically encoded targeting peptides to alter vesicle distribution, and the experimental issues that remain before clinical translation can be considered. These issues include control of vesicle composition, definition of the delivered dose, formulation-related toxicity, and the effects of repeated administration.

Artificial intelligence (AI) is redefining the interpretation of epigenetic regulation in RNA biology by enabling the integration of complex, multi-layered datasets spanning genomic, epigenomic, and transcriptomic levels. These improvements are relevant for non-coding RNA associated with extracellular vesicles (EV-ncRNA), whose content may offer a minimally invasive readout of the molecular programs defining cancer cell states. In cancer, epigenetic modifications affect RNA editing, epi-transcriptomic changes, alternative splicing, and circular RNA (circRNA) biogenesis, also modulating the EV-ncRNA content. AI models can propose regulatory activity from sequence context and to integrate chromatin accessibility. These advances offer an opportunity to associate regulatory programs with EV-ncRNA signatures, moving beyond descriptive profiling toward biologically interpretable models. AI can then integrate multimodal data and apply these models to deduce specific cancer cell states from EV-ncRNA obtained by liquid biopsy. This perspective outlines a conceptual framework for understanding EV-ncRNA signatures in ovarian cancer (OC) and for exploring their potential roles in biomarker discovery, patient stratification, and dynamic disease monitoring. The major challenges that may delay effective implementation are also discussed, including data heterogeneity, the absence of standardized EV multi-omics datasets, and the limited interpretability of deep learning models.