Circulating extracellular vesicle isomiR signatures predict therapy response in patients with multiple myeloma

Abstract

Multiple myeloma (MM) is a plasma cell neoplasm characterized by high inter- and intra-patient clonal heterogeneity, leading to high variability in therapeutic responses. Minimally invasive biomarkers that predict response may help personalize treatment decisions. IsoSeek, a single-nucleotide resolution small RNA sequencing method can profile thousands of microRNAs (miRNAs) and their variants (isomiRs) from patient plasma-purified extracellular vesicles (EVs). Machine learning-generated miRNA/isomiR classifiers accurately predict therapeutic response in relapsed/refractory MM (RRMM) patients receiving daratumumab-containing regimens, achieving an area-under-the-curve of 0.98 (95% confidence interval [CI]:0.94-1.00). A classifier signature with the plasma cell-selective miR-148-3p, predicts durable response (≥6 months), progression-free (hazard ratio [HR]: 33.09, 95% CI: 4.2-262, p < 0.001), and overall survival (HR: 3.81, 95% CI: 1.05-13.99, p < 0.05). Targetome analysis connects the prognostic classifier to established MM drug targets BCL2 and MYC suggesting biological relevance. Thus, EV-isomiR sequencing in MM patients offers a tumor-naïve alternative to an invasive bone-marrow biopsy for predicting treatment outcome.

Keywords: extracellular vesicles; isomiR modelling; liquid biopsy; miRNAs; multiple myeloma; personalized therapy; response prediction.

Graphical abstract

Figure 1

Schematic workflow overview and MM pEV characterization (A) Simplified overview of the patients included in the different models. (B) Schematic overview of the workflow for plasma extracellular vesicles (pEVs) isolation, followed by EV characterization and miRNA sequencing and subsequent analyses. (C) Particle concentration of plasma EVs from healthy donors and MM patients with active disease or clinical response using Exoid. Data are shown as the mean ± standard error of the mean (SEM) (n = 3). (D) Deconvolution (CIBERSORTx) of mRNA sequencing data from EVs from 5 patients with active MM and 5 patients responding to treatment, using LM22 single-cell sequencing dataset as reference. No clear difference can be observed between the two groups, and almost all immune cell types are represented in both. (E) Schematic overview of the model building procedure using a cross-validated approach and validation in never-seen datasets. Figure was partially created with BioRender.com.

Figure 2

A miRNA network in plasma EVs distinguish patients with active MM from healthy controls (A) Schematic overview of the cohorts used for model building (cohort 1), model validation (cohort 1), and external validation (cohort 2), including sample sizes. (B) ROC curves of the classic miRNA disease detection model. In blue, ROC curve of the validation set, and in yellow, ROC curve of the external validation. AUC and confidence intervals are shown in the table. The model validates with a high AUC in both validation sets. Below are the confusion matrix of the two validations sets (internal on the left, external on the right), for the classic miRNA model. Both matrices show a high true positive rate and a low false negative rate. (C) ROC curves of the isomiR disease detection model. In blue, ROC curve of the validation set, and in yellow, ROC curve of the external validation. AUC and confidence intervals are shown in the table. The model achieves a high AUC in the internal validation set but not in the external set. Below are the confusion matrix of the two validations sets (internal on the left, external on the right), for the isomiR model. Only the internal validation set shows a high true positive rate and a low false negative rate. (D) Reduced network of the miRNA-gene-targets in the classical miRNA diagnosis model, which includes MYC and BCL2. (E) Top: visualization of the RPM levels of three of the top differentially expressed mRNAs (SDC1, JCHAIN, and UAP1) in patients with MM AD compared to healthy individuals. Data are shown as the mean ± standard error of the mean (SEM) (n = 5 per group). Asterisks (∗) denote statistically significant differences: ∗∗∗∗p < 0.0001. as determined by t test. Bottom: PanglaoDB Augmented 2021 enriched cell types in the AD samples compared to healthy individuals, with plasma cells showing the highest significant enrichment.

Figure 3

pEV-IsomiRs for on-treatment response assessment in patients with multiple myeloma (A) Schematic overview of the cohorts used for model building (cohort 1), model validation (cohort 1), and prospective validation (cohort 2), including sample sizes. (B) ROC curves of the classic miRNA response assessment model. In blue, ROC curve of the validation set, and in yellow, ROC curve of the prospective validation set. AUC and confidence intervals are shown in the table. The model validates with a good AUC in the retrospective data (AUC: 0.89) and worse in the prospective dataset (AUC: 0.77). Below are the confusion matrix of the two validations sets (retrospective on the left, prospective on the right), for the classic miRNA model. (C) ROC curves of the isomiR response assessment model. In blue, ROC curve of the retrospective validation set, and in yellow, ROC curve of the prospective validation set. AUC and confidence intervals are shown in the table. The model achieves a high AUC in both sets (AUC: 0.90 in the retrospective and AUC: 0.98 in the prospective). Below are the confusion matrix of the two validations sets (retrospective on the left, prospective on the right), for the isomiR model. (D) Reduced network of the miRNA-gene-targets in the isomiR response assessment model. (E) Top: visualization of the RPM levels of three of the top differentially expressed mRNAs (MZB1, JCHAIN, and SDC1) in MM patients with active disease, compared to patients with MM responding to treatment (responders). Data are shown as the mean ± standard error of the mean (SEM) (n = 5 per group). Asterisks (∗) denote statistically significant differences: ∗∗∗∗p < 0.0001. as determined by t test. Bottom: PanglaoDB Augmented 2021 enriched cell types in the active disease samples compared to responders, with plasma cells showing the highest significant enrichment.

Figure 4

pEV-IsomiRs allow response monitoring in individual MM patients over time (A–D) pEV-isomiR model predictions over time (blue lines, each blue dot represents a measurement) compared to M-protein levels (red lines, each red dot represents a measurement) for four individual MM patients, showing that the model closely tracks the M-protein metric, demonstrating its robustness. Although some misclassifications are present (indicated by the arrows), the model levels in those samples were consistent with the status of the patients in the future months. M-protein relative level (right axis) at each time point was calculated as the ratio at that time point as compared to M protein level at the beginning of treatment.

Figure 5

Pre-treatment pEV-isomiR model forecast durable response, PFS and OS (A) Schematic overview of the patient group generation (left) and the cohorts used for model building (cohort 1) and model validation (cohort 2), including sample sizes. (B) (Left) ROC curve of the isomiR durable response prediction model. The model achieved a high AUC of 0.84 in the validation set. (Right) Confusion matrix of the validation set for the isomiR model and sensitivity, specificity, PPV, and NPV of the model. (C) (Left) Reduced network of the miRNA-gene-targets in the isomiR durable response model. (Right) Expression of Has-Mir-148-P1-3p over different cell types, including plasma cells. All samples with a tissue/cell type annotation in IsomiRDB were included, and number of samples in each category is summarized below each violin plot (n). Asterisks (∗) denote statistically significant differences: ∗∗∗∗ p < 0.0001. as determined by t test. (D) Progression-free survival (PFS) of the RRMM patients (n = 29). The survival curve was computed using the Kaplan-Meier method. The hazard ratio (HR) calculated by Cox-regression is 33.09 (CI: 4.2–262, p < 0.0001). (E) Overall survival (OS) analysis of the RRMM patients (n = 29). The survival curve was computed using the Kaplan-Meier method. The HR calculated by Cox-regression is 3.81 (CI: 1.05–13.99, p < 0.05).