CXCL13 is a predictive biomarker in idiopathic multicentric Castleman disease

Abstract

Idiopathic multicentric Castleman disease (iMCD) is a rare and poorly-understood cytokine storm-driven inflammatory disorder. Interleukin-6 (IL-6) is a known disease driver in some patients, but anti-IL-6 therapy with siltuximab is not effective in all patients, and biomarkers indicating success at an early time point following treatment initiation are lacking. Here we show, by comparison of levels of 1,178 proteins in sera of healthy participants (N = 42), patients with iMCD (N = 88), and with related diseases (N = 60), a comprehensive landscape of candidate disease mediators and predictors of siltuximab response. C-X-C Motif Chemokine Ligand-13 (CXCL13) is identified and validated as the protein most prominently up-regulated in iMCD. Early and significant decrease in CXCL13 levels clearly distinguishes siltuximab responders from non-responders; a 17% reduction by day 8 following siltuximab therapy initiation is predictive of response at later time points. Our study thus suggests that CXCL13 is a predictive biomarker of response to siltuximab in iMCD.

Conflict of interest statement

D.C.F. has received research funding for the ACCELERATE registry from EUSA Pharma and consulting fees from EUSA Pharma, and Pfizer provides a study drug with no associated research funding for the clinical trial of sirolimus (NCT03933904). D.C.F. has two provisional patents pending related to the diagnosis and treatment of iMCD, including one related to CXCL13 as a biomarker in iMCD. A.B.O., L.K., and P.B. are employed by and/or receive equity ownership from Medidata Solutions, a subsidiary of 3DS. A.F. has received honoraria and research support from Janssen Pharmaceuticals and EUSA Pharma. F.v.R. provides consultancy for EUSA Pharma, Takeda, Sanofi Genzyme, Adicet Bio, Kite Pharma, and Karyopharm Therapeutics. The remaining authors declare no competing interests.

© 2022. The Author(s).

Figures

Fig. 1. Differential expression of proteins between iMCD and healthy controls demonstrates increased CXCL13 across cohorts.

a Volcano plot visualization of differentially expressed proteins between iMCD (n = 88) and healthy donors (n = 42). To detect differences in the proteome, linear regression models comparing iMCD to healthy, adjusted for age and sex, were run on each analyte. Results were adjusted by Benjamini & Hochberg method with alpha <0.05. We identified 251 upregulated proteins and 118 downregulated proteins in iMCD relative to healthy. Significant proteins that are >2-fold upregulated are labeled. Exact P values are listed in Supplementary Data 1 and the source data file. b The three most upregulated proteins (log2 transformed) in iMCD (n = 88) relative to healthy controls (n = 42) are visualized by boxplot. c To validate the proteomic changes identified against healthy individuals, we compared samples obtained from an independent validation iMCD cohort (n = 23) to the expected upper limit in a healthy population using an orthogonal platform. The median analyte level in iMCD was compared to the 97.5th percentile of the expected healthy range using a one-sided Mann–Whitney U-test with adjustment by Benjamini & Hochberg, alpha <0.05. A one-sided volcano plot of the 40 overlapping proteins quantified is shown. Proteins with median levels confirmed to be significantly elevated in iMCD over the 97.5th percentile of the expected healthy range are labeled. Exact P values are listed in Supplementary Table 1 and the source data file. d Box plots of the absolute concentrations (pg/mL) of the four proteins with the largest fold-change in iMCD (n = 23) compared to the healthy range are shown. The healthy donor range represents the 2.5th to 97.5th percentiles for healthy donors according to the RBM Human Discovery Map v1.0 platform. CCL18 C-C motif chemokine ligand 18, CCL21 C-C motif ligand 21, CK-MB creatine kinase-MB, CXCL13 C-X-C motif chemokine ligand 13, EDAR tumor necrosis factor receptor superfamily member EDAR, HGF hepatocyte growth factor, IgE immunoglobulin E, IGFBP-1 insulin-like growth factor-binding protein 1, IL-1F6 interleukin-36 alpha, IL-5 Ra interleukin-5 receptor subunit alpha, MIP1a macrophage inflammatory protein-1 alpha, MMP7 matrix metallopeptidase 7, NPS-PLA2 phospholipase A2, PBEF pre-B-cell colony-enhancing factor, PUR8 adenylosuccinate lyase, SAA serum amyloid A, VEGF vascular endothelial growth factor, ULN upper limits of normal. Box plots include a center line (median), box limits (upper and lower quartiles), whiskers (1.5x interquartile range), and all data points. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Source data are provided as a Source Data file.

Fig. 2. Serum CXCL13 differentially expressed in iMCD compared to clinicopathologically overlapping diseases.

a Spearman rank correlation test of log2 CXCL13 against platelets in iMCD (n = 88) at the time of sample draw. There was no discernable relationship (R = −0.015, p = 0.89). b Box plots showing the log2 normalized relative fluorescent unit (RFU) concentrations of CXCL13 in iMCD patients with hyaline vascular (n = 26), mixed (n = 40), and plasmacytic (n = 21) histopathology. Analysis of variance followed by a two-tailed post hoc t-test with Bonferroni adjustment was used to test for differences according to histopathological subtype There was no difference between histopathological subtypes (hyaline vascular vs mixed histopathology, p = 0.06; plasmacytic vs hyaline vascular, p = 0.92; and plasmacytic vs mixed, p = 0.07. c Box plots showing the log2 normalized RFU concentrations of CXCL13 in iMCD compared to the related disorders and healthy donors. iMCD (n = 88) differed from each HHV8-MCD (n = 20), RA (n = 20), HL (n = 20), and healthy (n = 42). Statistical significance of disease status was determined with a linear regression model with age, sex, and CRP as covariates, and a two-tailed Wald test was used to test pairwise differences with iMCD, with multiple comparisons Bonferroni corrected (RA: p = 5.6 × 10−4; HL: p = 0.03; HHV8-MCD: p = 0.003; healthy: p = 3.3 × 10−7). d Pearson correlation test demonstrating a strong linear correlation between log2(CXCL13) as measured by the SomaSCAN multiplex DNA-aptamer assay and by an ELISA immunoassay (n = 66) R = 0.93; 95% confidence interval: 0.89–0.96, p = 1.7 × 10−29). Box plots include a center line (median), box limits (upper and lower quartiles), whiskers (1.5x interquartile range), and all data points. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Source data are provided as a Source Data file.

Fig. 3. CXCL13 expression in the lymph node demonstrates a difference between iMCD-TAFRO and reactive but not iMCD-NOS and reactive.

Representative germinal centers from a iMCD-TAFRO, b iMCD-NOS, c RA, and d reactive lymph nodes, bars represent 200 µm. e Comparison of positive pixel intensity in iMCD (n = 17) with reactive (n = 17; p = 0.07) and RA (n = 9; p = 0.45). The error bar represents the standard error of the mean. f Comparison of the number of germinal centers in iMCD (n = 17), RA (n = 9), and reactive (n = 17) lymph nodes. iMCD lymph nodes have significantly more germinal centers than RA (p = 0.002) but not reactive (p = 0.596). g Comparison of positive pixel intensity between iMCD-TAFRO (n = 9) and reactive (n = 17) and between iMCD-NOS (n = 8) and reactive. iMCD-TAFRO demonstrates significantly more CXCL13 expression than reactive (p = 0.003), but there is no difference between iMCD-NOS and reactive (p = 0.67). The error bar represents the standard error of the mean. Statistical differences were determined by a two-tailed non-parametric normal scores test. Box plots include a center line (median), box limits (upper and lower quartiles), whiskers (1.5x interquartile range), and all data points. Source data are provided as a Source Data file.

Fig. 4. CXCL13 identified and validated as a potential early predictor of siltuximab response.

a For each protein, a model was fit to assess the significance of the interaction of time point and response in patients in the phase II siltuximab trial; coefficient estimates at day 8 siltuximab treatment are plotted against those at day 64 of treatment and colored by significance. Eight proteins were significantly decreased in siltuximab responders compared to non-responders at both time points, including the labeled proteins CXCL13, immunoglobulin A (IgA), agouti-related protein (ART), CD5 antigen-like (CD5L), neuropilin-1 (NRP1), interleukin-18-binding protein (IL18Bpa), and two unlabeled proteins, tumor necrosis factor receptor superfamily member 17 (BCMA) and beta-2-microglobulin (b2M). NPS-PLA2 was significantly decreased in responders on day 8 but was not significant by day 64. b Log2 fold-change and c mean (95% confidence interval) of CXCL13 at day 8 and day 64 in responders (n = 17), non-responders (n = 32), and placebo (n = 24) patients of the siltuximab phase II trial. d Coefficient estimates of overlapping proteins quantified in an independent cohort (phase I siltuximab trial) at day 22/29 of siltuximab treatment plotted against those at day 43 and colored by significance. CXCL13 is the only protein identified in both cohorts that was significant at both time points. Resistin was significant at day 43 only. e Log2 fold-change and f mean (95% confidence interval) of CXCL13 at day 22/29 and day 43 in responders (n = 10) and non-responders (n = 13) in the phase I siltuximab trial. Box plots include a center line (median), box limits (upper and lower quartiles), whiskers (1.5x interquartile range), and all data points. Source data are provided as a Source Data file.

Fig. 5. Decreased serum CXCL13 early in siltuximab treatment is predictive of future response.

a Receiver operating characteristic (ROC) curve for the prediction of siltuximab response based on serum levels of CXCL13 in the training cohort (n = 48; 31 non-responders, 17 responders) (AUC = 0.86 [CI:0.753–0.966])) and the test cohort (n = 20; 11 non-responders, 9 responders) (AUC = 1.0) b Waterfall plot of response prediction values by patient colored by true response status in the training cohort. c Classification by logistic regression in the training cohort. Responders plotted along the top horizontal bar and non-responders along the bottom horizontal bar. The x-axis represents the percent change in CXCL13 from the baseline. The logistic regression curve is plotted in blue, with a vertical line drawn at a 17% reduction. By day 8, a 17% reduction in CXCL13 has an 82% recall, 82% true positive rate, 23% false positive rate, 79% accuracy, and 67% precision in response prediction in the training cohort. d Classification by logistic regression in the test cohort. By day 22/29, 17% reduction in CXCL13 has a 100% recall, 100% true positive rate, 18% false positive rate, 90% accuracy, and 92% precision, and e waterfall plot of response prediction values by patient colored by true response status in the test cohort. Source data are provided as a Source Data file.