The Limitation of Current Liquid Biopsies Existing blood-based biomarkers for immunotherapy, including cell-free DNA and circulating tumor cells, reflect only tumor-intrinsic properties. Successful prediction and monitoring of immune checkpoint inhibitor (ICI) response requires simultaneous capture of both tumor-derived and immune-system-derived signals, since ICI efficacy depends on an active host immune response rather than tumor biology alone.
Why Extracellular Vesicles Extracellular vesicles (EVs) are nano-sized particles secreted by both tumor cells and immune cells, carrying a sub-transcriptome representative of their cell of origin. Plasma-derived bulk EVs thus reflect both tumor-intrinsic biology and immune cell activity - including contributions from CD4+ and CD8+ T cells, dendritic cells, regulatory T cells, and macrophages - making them uniquely positioned to jointly capture the immune and tumor dynamics that determine ICI outcome.
Study Design and Key Finding Transcriptomic profiling of plasma-derived EVs was performed for 50 metastatic melanoma patients receiving ICI (discovery cohort, microarray-based) and validated in an independent cohort of 30 patients (evRNA-seq). A novel Bayesian deconvolution model was developed to partition EV transcripts into tumor and nontumor sources. EV profiles revealed validated differentially expressed genes and pathways separating responders from nonresponders, and a random forest classifier built from pretreatment EV DEGs predicted ICI response with AUROC 0.737 in independent testing.
EV Isolation and Profiling Plasma EVs were isolated from 1 ml of patient plasma using QIAGEN exoRNeasy Serum/Plasma Midi Kit column isolation, yielding 10-30 ng of EV RNA per ml. Discovery cohort EVs were profiled on Applied Biosystems GeneChip Human Transcriptome Array 2.0 (microarray). Validation cohort EVs were profiled by RNA sequencing (evRNA-seq) on Illumina HiSeq 2000 using the SMARTer Stranded Total RNA-Seq Kit v2. Serial blood samples were collected at pretreatment and on-treatment time points.
Cross-Platform Normalization To enable co-analysis of microarray and RNA-seq data across discovery and validation cohorts, a ComBat batch correction was applied treating the two platforms as two batches, with treatment response included as a biological covariate to preserve true biological variation during batch correction. The resulting joint matrix was quantile normalized and then split back into platform-specific matrices for downstream analysis. This multi-step normalization approach was validated by confirming that platform-specific effects did not confound the main biological findings.
Differential Expression and Classification Differential expression between responders and nonresponders was computed using limma's empirical Bayes linear modeling framework with a 1.5 log fold change and nominal P-value cutoff of 0.1. A gene was considered validated if it met significance criteria in both discovery and validation cohorts independently. For the predictive classifier, 100 randomized train-test partitions (n=41 training, n=30 testing, mixing both platforms in each split) were used with K=5 internal cross-validation to optimize random forest hyperparameters, ensuring that platform effects were not confounded with predictions.
EV-Tumor Concordance In cell line experiments, EV transcriptomes correlated strongly with their cell of origin (average R2 = 0.87). In 9 patients with paired EV and tumor biopsy samples, EV and tumor transcriptomes were closely correlated (average R2 = 0.82). The majority of genes expressed in bulk tumors were also detected in paired EVs, with only a subset of transcripts unique to either compartment.
Immune Pathway Enrichment in EVs Gene set enrichment analysis revealed that EVs were specifically enriched for immune-related gene ontology terms including T cell activation, NK cell activation, MHC class II biosynthesis, B cell receptor signaling, and chemokine-mediated signaling. In contrast, tumor samples were enriched for metabolic and proliferative pathways such as chromatin remodeling, DNA repair, and cell cycle regulation. CIBERSORT deconvolution of patient EVs showed enrichment of neutrophils, NK cells, and CD4+/CD8+ T cells specifically in EVs compared to matched tumors, and a relative depletion of macrophages and mast cells.
Why This Matters for ICI Monitoring The immune cell enrichment in EVs, particularly for CD4+, CD8+ T cells, and NK cells that are central mediators of anti-PD1 responses, supports the hypothesis that bulk plasma-derived EVs simultaneously reflect both tumor biology and immune activation status. This dual information content makes EV transcriptomics more informative for ICI response prediction than cfDNA-based assays that only capture tumor-derived signals.
On-Treatment Differential Pathways Among 258 differentially enriched MSigDB pathways between responders and nonresponders in on-treatment EV profiles from the discovery cohort, 25 were validated in the independent evRNA-seq cohort. Validated pathways included T cell receptor signaling, CTLA4, TGF-beta, SMAD2/3, Notch, TNFR2, and VEGF receptor signaling, all implicated in ICI resistance. Longitudinal analysis showed that T cell receptor pathway activity declined over time in nonresponders but was maintained in responders.
On-Treatment DEGs At the gene level, 43 FDR-corrected DEGs were identified in the discovery cohort on-treatment comparison. Cancer testis antigens MAGEA1 and MAGEA3 - known to be expressed by melanoma cells - were among validated DEGs elevated in nonresponders. KLF10, a TGF-beta pathway member with roles as both a tumor suppressor and a regulator of TH1/TH17 and CD8+ memory T cell formation, was elevated in responders on-treatment and was predicted by Bayesian deconvolution to derive primarily from nontumor sources, suggesting its differential expression reflects immune rather than tumor dynamics.
Pretreatment Predictors and Classifier Performance In pretreatment EV profiles, 26 of 101 differentially enriched pathways validated in the independent cohort, with differences in Notch, TGF-beta, MAPK-related (ERBB4), and keratinization pathways. Thirty-eight nominal DEGs with replicated P-values and log fold changes included CD1A, MAP2K4, TRBV7-2, and IGFL1. A random forest classifier built on pretreatment validated DEGs predicted response versus nonresponse with AUROC 0.784 in internal cross-validation and AUROC 0.737 in independent testing. IGFL1 expression alone showed significant stratification in Kaplan-Meier progression-free survival in both discovery and validation cohorts.
Bayesian EV Deconvolution Model A novel Bayesian probabilistic model was developed to partition observed mixed plasma EV expression into tumor and nontumor components for each gene. The model explicitly accounts for a per-gene packaging coefficient representing differential enrichment or depletion of transcripts during export from tumor cells into tumor-derived EVs. EV transcripts from healthy controls provided the nontumor reference profile. The full model was implemented in Stan using No-U-Turn Hamiltonian Monte Carlo (10,000 iterations, 4 chains) for key genes, with a computationally efficient maximum a posteriori (MAP) version using Sequential Least Squares Programming for genome-wide estimates.
Deconvolution Validation and Biological Insights Deconvolution predictions for known tumor genes (MITF, TYR, MLANA, TP53) and known immune genes (CD8A, CD4, FOXP3, PDCD1) aligned with expected tumor or nontumor assignments. Melanoma-related KEGG pathways (MELANOMA, MELANOGENESIS) had the highest mean estimated tumor fractions, while immune pathways (T cell receptor, antigen processing) had the lowest. Pretreatment DEGs were enriched for tumor-derived transcripts, while on-treatment DEGs had greater nontumor contribution, consistent with on-treatment changes reflecting immune rather than tumor biology.
Mutational Information in evRNA-Seq The validation cohort evRNA-seq data was used for somatic mutation calling against the hg38 reference using BWA and GATK HaplotypeCaller, followed by filtering against the COSMIC somatic mutation database. Responders had significantly higher COSMIC driver somatic mutation fractions than nonresponders (Mann-Whitney P=0.00637). Overall survival Kaplan-Meier curves showed significant stratification by high versus low COSMIC mutational fraction (log-rank P=0.014), while progression-free survival stratification was not significant, suggesting evRNA-seq mutational information is more predictive of long-term than short-term ICI effects.
Deconvolution Model Limitations The Bayesian deconvolution model is limited by three main factors: (1) simplifying assumptions that the EV packaging coefficient is linear and shared between in vitro cell lines and in vivo patient samples; (2) inability to account for immune infiltration within tumors and patient heterogeneity; and (3) limited sample size (fewer than 50 matched tumor-EV pairs). These limitations likely explain why the model sometimes classifies predominantly immune genes such as CD8A into the tumor compartment. Quantitative tumor fraction estimates should be interpreted qualitatively until more experimental validation of EV source specificity is completed.
Performance Gap Compared to Tumor Biopsies The pretreatment EV classifier achieved AUROC 0.737, compared to AUROC 0.8 to 0.9 for classifiers built from bulk tumor biopsy transcriptomic profiles. This gap reflects the mixed biological content of plasma EVs and the analytical complexity of cross-platform normalization. However, the key advantage of EV profiling is its minimally invasive nature, enabling serial sampling throughout treatment to track immune dynamics, which is impossible with repeated tumor biopsies.
Future Development Priorities Cell-type-specific EV selection (e.g., tumor-specific EV enrichment using EpCAM or immune cell EV depletion) and complementary whole-exome sequencing of EV DNA would refine both the transcriptomic and mutational analyses. Longitudinal EV monitoring at multiple on-treatment time points could create dynamic response trajectories, potentially enabling earlier detection of acquired resistance before radiological progression. Integrating evRNA-seq with ctDNA and tumor-biopsy transcriptomics into a multi-modal biomarker platform may achieve clinical predictive performance while preserving the practical advantage of liquid biopsy sampling.