The Challenge of Predicting Immunotherapy Outcomes: Anti-PD-1 immunotherapy has become standard of care for advanced non-small cell lung cancer (NSCLC), improving five-year overall survival to around 20% in unselected patients and up to 40% in patients with high PD-L1 expression. However, more than 50% of PD-L1-high patients still fail to respond, highlighting the urgent need for better biomarkers.
A Novel Biomarker Source - Urine: This study proposes urine extracellular vesicles (EVs) as a non-invasive liquid biopsy source for identifying predictive proteins. Unlike blood-based tests, urine collection is entirely non-invasive and can capture both human host proteins and bacterial proteins from the gut microbiome that translocate into the bloodstream and are filtered into the urine.
Study Goals: Researchers at Semmelweis University analyzed the urine EV proteome of 33 advanced-stage NSCLC patients treated with anti-PD-1 immunotherapy, stratifying patients by whether they had long progression-free survival (PFS) greater than 6 months or short PFS of 6 months or less. Machine learning was used to validate the predictive power of identified proteins.
Two-Layer Analysis: Uniquely, this study analyzed both host human proteins and bacterial proteins found in urine EVs, as well as correlating urine bacterial signatures with gut microbiome metagenomics data from stool samples. This dual-layer approach provides insight into how the gut microbiome may influence immunotherapy outcomes.
Patient Cohort: 33 advanced-stage NSCLC patients (Stage IIIB/IV) treated between 2019 and 2020 were enrolled. 22 patients had long PFS (over 6 months) and 11 had short PFS. Treatments included nivolumab monotherapy (n=16), pembrolizumab monotherapy (n=10), atezolizumab (n=1), and durvalumab-based chemo-immunotherapy combinations (n=6).
Urine EV Isolation and Proteomics: Urinary EVs were isolated from 900 microliters of urine using EVtrap magnetic beads, which achieve over 95% recovery efficiency. Proteins were digested with Lys-C and trypsin, then analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) on a Q-Exactive HF-X instrument. The database search used both human and bacterial UniProt databases.
Machine Learning Validation: Random Forest (RF) models with stratified five-fold cross-validation were trained on human and bacterial protein abundance data to classify patients by PFS group. Bayesian Additive Regression Trees (BART) provided a non-linear validation model. Gut microbiome composition was determined by shotgun metagenomics on stool samples from 23 of the 33 patients.
Statistical Analysis: Spearman's correlation tests identified proteins whose abundance correlated with PFS duration. Wilcoxon rank-sum tests compared protein abundance between short and long PFS groups. Multivariate Cox proportional hazard regression with clinical confounders (gender, chemotherapy, PD-L1, COPD, BMI) identified independent predictors.
Broad Proteomic Landscape: LC-MS/MS analysis identified 6,183 proteins in urine EV samples, comprising 3,513 human proteins, 2,647 bacterial proteins, 19 fungal proteins, and 4 viral proteins. Of the 3,513 human proteins, 186 showed differential abundance between short and long PFS groups (p less than 0.05).
Top Predictive Human Proteins: From the differentially abundant proteins, multivariate Cox regression identified MPP5, IGKV6-21, NT5E, and KRT27 as significant positive predictors of long PFS (protective factors), while LMAN2, NUTF2, NID1, TNC, IGF1, BCR, GPHN, and PPBP were significant predictors of short PFS (risk factors).
Pathway Enrichment Findings: Proteins associated with long PFS were enriched in general immune function and innate immunity pathways. Proteins linked to short PFS enriched pathways related to the Endosomal/Vacuolar pathway, Complement cascade, COPI-mediated anterograde transport, and nucleoside trisphosphate metabolic processes - suggesting dysregulated vesicle trafficking and immune evasion mechanisms in non-responders.
Machine Learning Performance: The Random Forest model achieved an AUC of 0.89 and 95% accuracy for predicting PFS using human urine proteins. The bacterial protein model achieved an AUC of 0.74. The BART Bayesian model validated the key findings, confirming the reliability of the top biomarker proteins.
Bacterial Proteins as Biomarkers: The study identified 2,647 bacterial proteins in urine EVs, with the fraction of bacterial to total urine EV proteins being significantly higher in patients with long PFS compared to short PFS. This suggests that a healthy gut microbiome state, reflected in higher bacterial EV shedding, may support better immunotherapy outcomes.
Species-Level Patterns: E. coli proteins dominated the long PFS bacterial signature (63% of long PFS-associated bacterial proteins), while short PFS-associated bacterial proteins came from multiple diverse taxonomic groups. E. coli and Enterococcus faecalis urine EV proteins were significantly more abundant in long PFS patients.
Gut-Urine Correlation: Stool shotgun metagenomics data confirmed that gut abundances of E. coli and E. faecalis positively correlated with their urine EV protein abundances, directly linking gut microbiome composition to urinary proteomic signatures. In the gut, Akkermansia, Bacteroides, and Escherichia genera were overrepresented in long PFS patients.
Bacterial Pathway Enrichment: Pathway analysis of long PFS-associated bacterial proteins found significant enrichment in Glycolysis/Gluconeogenesis, Two-component system signaling, ABC transporters, and Aminoacyl-tRNA synthesis pathways - metabolically active bacterial states that may reflect a healthier microbiome capable of supporting immune function.
First Urine Proteome ICI Predictor Study: To the authors' knowledge, this is the first study to characterize the predictive role of both host and bacterial urine proteomes in anti-PD-1-treated NSCLC patients. Existing biomarkers like PD-L1 TPS and TMB fail to capture more than half of potential responders, creating a pressing clinical gap.
Complementary to Existing Biomarkers: Interestingly, PD-L1 expression did not differ significantly between long and short PFS groups in this cohort, highlighting that urine proteomics captures different biological signals. The top protein biomarkers were independent of PD-L1 TPS in Cox regression, suggesting additive predictive value.
Practical Advantages of Urine Testing: Urine collection requires no medical procedure, is repeatable at any timepoint, and can be processed using standard mass spectrometry workflows already available in clinical research settings. The EVtrap bead-based isolation method is highly standardized and achieves consistent results.
Monitoring Treatment Response: Follow-up urine samples at 120 days post-treatment showed that BCAS1 and KRT27 protein levels decreased over time specifically in long PFS patients, suggesting these proteins could serve not only as baseline predictors but also as on-treatment monitoring markers.
Small Cohort Limitations: The study enrolled only 33 patients, which limits statistical power and the ability to control for all clinical variables. The 22:11 ratio of long-to-short PFS patients creates a class imbalance that may affect machine learning model generalizability.
Need for Prospective Validation: All results were obtained from a retrospective cohort at a single institution in Hungary. External validation in independent, prospective cohorts across diverse patient populations, treatment regimens, and collection protocols will be needed before clinical adoption.
Expanding to Other Therapies: This study focused on anti-PD-1 agents. Future work could examine whether urine EV proteomics predicts outcomes for combination immunotherapy, chemotherapy, or targeted therapies in NSCLC, and whether similar signatures apply to other cancer types treated with checkpoint inhibitors.
Integrating Multi-Omics: Combining urine EV proteomics with blood-based proteomics, circulating tumor DNA, or gut microbiome sequencing into integrated multi-omics models may further improve predictive accuracy. The study demonstrates the feasibility of simultaneously profiling host and microbial components from a single urine sample.