Early imaging or blood biomarkers of tumor response are desperately needed to customize antiangiogenic therapy for cancer patients. Anti–vascular endothelial growth factor (VEGF) therapy can “normalize” brain tumor vasculature by decreasing vessel diameter and permeability, and thinning the abnormally thick basement membrane. We hypothesized that the extent of vascular normalization will be predictive of outcome of anti-VEGF therapy in glioblastoma. We used advanced magnetic resonance imaging methods to monitor vascular parameters and treatment response in 31 recurrent glioblastoma patients enrolled in a phase II trial of cediranib, an oral pan-VEGF receptor tyrosine kinase inhibitor. We evaluated the correlation between clinical outcome and magnetic resonance imaging–measured changes in vascular permeability/flow (i.e., Ktrans) and in microvessel volume, and the change of circulating collagen IV levels, all after a single dose of cediranib. Here, we show that evaluation of biomarkers as early as after one day of anti-VEGF therapy with cediranib is predictive of response in patients with recurrent glioblastoma. Changes in Ktrans, microvessel volume, and circulating collagen IV correlated with duration of overall survival and/or progression-free survival (P < 0.05). When we combined these three parameters into a “vascular normalization index,” we found that it closely associated with overall survival (ρ = 0.54; P = 0.004) and progression-free survival (ρ = 0.6; P = 0.001). The vascular normalization index described here should be validated in randomized clinical trials. [Cancer Res 2009;69(13):5296–300]
Several candidate biomarkers have been proposed for antivascular endothelial growth factor (VEGF) therapy with bevacizumab, an anti-VEGF antibody (Genentech). However, there are no validated predictive biomarkers for any anti-VEGF therapy in cancer (1). Identification of effective therapies in a timely fashion is critical for patients with limited survival, such as those with glioblastoma.
Functional magnetic resonance imaging (MRI) “water diffusion mapping” has been proposed as an early imaging biomarker for glioblastoma (2). However, glioblastomas secrete high VEGF levels, which lead to high vascular permeability and vasogenic edema (3, 4). Thus, blocking VEGF in glioblastoma likely leads to decreased vascular permeability, and edema alleviation (5, 6). Unfortunately, this could lead to difficulties in evaluating tumor response or diffusion maps using MRI (7). A recent study has identified a decrease in fluorothymidine uptake measured by positron emission tomography as a potential predictive marker for recurrent glioblastoma patients treated with bevacizumab and CPT-11 (8). Because the decrease in tumor uptake of fluorothymidine might also be confounded by a reduction in vascular permeability induced by bevacizumab, it is not clear if this biomarker reflects a decrease in permeability or proliferation.
Here, we combined imaging with circulating biomarkers. In preclinical models, anti-VEGF agents rapidly but transiently “normalize” the structure and function of glioblastoma vessels—they decrease vessel diameter and permeability, increase tumor oxygenation, and induce a thinning of the abnormally thick basement membrane (9, 10). Phase II trial data showed that cediranib (AstraZeneca, a pan-VEGF receptor tyrosine kinase inhibitor) decreased—as early as after 1 day of treatment—two neuroimaging biomarkers: vessel diameter (transiently) and permeability as assessed by Ktrans (more persistently) in a cohort of 16 consecutive recurrent glioblastoma patients (5). In further preclinical studies, we found that cediranib can improve survival in glioblastoma-bearing mice via edema reduction even in the face of persistent tumor growth (11). In that study, edema alleviation was due to transient normalization of the vasculature, and was associated with a transient increase in plasma collagen IV. We reasoned that increases in circulating collagen IV would reflect basement membrane thinning.
Because preclinical and clinical findings suggest that vascular normalization by cediranib therapy might be beneficial in glioblastoma patients, we hypothesized that changes at day 1 in three biomarkers related to vascular normalization would be predictive of improved survival in recurrent glioblastoma patients. Moreover, we hypothesized that a composite index based on the early changes in biomarkers of vascular normalization would more closely predict the outcome of cediranib treatment.
Materials and Methods
Dynamic Contrast-Enhanced–MRI Methodology
Image acquisition. All patients were scanned (with informed consent) at two baseline time points, typically 3 to 7 d (average, 5.7) and then 1 d before the first treatment, as well as 1 d after the first treatment on a 3 Tesla MRI system (Siemens), for the following key sequences:.
Dynamic contrast-enhanced images. This is a series of acquisitions of a 50.6-mm-thick slab consisting of 20 slices. All scans are 2.9 × 2.0 mm in-plane resolution, with a 2.1 mm slice thickness, 0.4 mm interslice gap, using a fast gradient echo technique (TR, 5.7 ms; TE, 2.73 ms). Data to allow computation of a T1 map of the tissue of interest are initially created using five different flip angles (2°, 5°, 10°, 15°, 30°). Then, the same slab of tissue is sampled with a 10° flip angle every 5.04 s for 252 s (50 time points), and 0.1 mMol/kg of gadopentetate-dimeglumine was injected 52 s after the beginning of the acquisition at 5 cc/s. Imaging time was 4 min and 12 s.
Permeability maps. Dynamic contrast-enhanced–MRI data were processed using custom made software written in Matlab to obtain maps of Ktrans (corresponding roughly to wash-in rates of the contrast agent; ref. 12). Ktrans can be influenced by flow, or by permeability, or both. In high-flow organs such as the brain, flow limitations are not usually a concern, but the blood-brain barrier severely limits permeability unless it is disrupted by disease. Even in such a state, Ktrans does not fully correspond to permeability, but it is related rather to the permeability-surface area product of the capillary bed (in nonflow-limited situations). Based on our previous data (5), Ktrans most likely is strongly related to permeability in this setting, although some flow dependence may be present as well.
Dynamic susceptibility contrast imaging. A 75-mm slab of tissue was imaged using a dual-echo, combined gradient-echo, and spin-echo planar sequence (TE, 34/103); each image had 1.7 mm in-plane resolution and 5 mm through plane resolution (128 × 128 matrix). There was a 2.5 mm interslice gap and 10 slices. We acquired 120 blocks of images—a block every 1.33 s. Gadopentetate-dimeglumine (0.2 mMol/kg) was injected at 5 cc/s after 85 s of imaging. Imaging time was 2 min and 45 s.
Blood volume maps. Relative CBV of smaller vessels (spin-echo images) was calculated using a standard deconvolution technique (13), with CBV corrected for leakage of the contrast agent across the blood-brain barrier (14, 15). These maps are relative and therefore unitless.
Patients continued cediranib monotherapy until progression, and we monitored the duration of PFS using T1-enhancement volume (Supplementary Figs. S1 and S2; Fig. 1).
Measurement of Circulating Collagen IV
Peripheral blood was obtained with informed consent from all patients at baseline and 1 d following the first dose of cediranib. Blood was collected in an EDTA-containing vacutainer, spun down, and plasma was aliquoted, and frozen immediately. We measured the plasma levels of soluble collagen IV using ELISA kits from Exocell, Inc. We measured the plasma levels of Ang1/Ang2 using ELISA kits (R&D Systems) and matrix metalloproteinase 2 using MSD plates (MesoScale-Discovery). Samples were run in duplicate.
We used multivariable Cox regression analyses of changes in three parameters: Ktrans, microvessel CBV, and plasma collagen IV 1 d after the first dose of cediranib. The vascular normalization index (VNI) was obtained as the negative of the linear predictor in the Cox regression with all three variables included, using regression's coefficient estimates. For example, if coefficients in the Cox regression predicting hazard of death were a for Ktrans change (ΔKtrans), b for CBV change (ΔCBV), and c for plasma collagen IV changes (Δcoll IV), then the index was
We expressed the changes as the logarithm of the day 1 to pretreatment level ratios. We present hazard ratio estimates and P values from the Wald test. In addition, we used a recursive partitioning analysis classes—a prognostic factor of outcome used for stratification of glioblastoma patients in clinical studies in which age, extent of resection, performance, and mental status are combined into one discrete variable (16, 17). Furthermore, we adjusted for log-transformed pretreatment tumor volume. The importance of sequential addition to the regression model of Ktrans, collagen IV, and CBV was tested in the analysis of deviance table, using residual deviance statistic and related P values from the likelihood ratio test. (χ2 tests were performed to compare the observed improvement versus adding a random noise.) Associations between biomarker changes and progression-free survival (PFS; at 6 mo) and overall survival (OS; at 1 y) were analyzed using logistic regression with cubic-spline function of a covariate.
Results and Discussion
Cediranib efficacy in recurrent glioblastoma. Thirty of 31 patients enrolled had at least a transient radiographic response to cediranib (Fig. 1). One patient dropped out of the study before receiving cediranib. The median PFS (with 95% confidence intervals) of the 30 recurrent glioblastoma patients who received cediranib was 117 days (88–145; Supplementary Fig. S1). The median OS was 227 days (177–293; Supplementary Fig. S2). The rate of patients alive and progression-free at 6 months was 25.8% (14.7–47.9).
Association between biomarkers and outcome. Several of the parameters measured by us—functionally related to vascular normalization (5, 10, 11)—changed after 1 day of cediranib treatment (Supplementary Table S1). However, only the early changes in Ktrans, microvessel CBV, and plasma collagen IV correlated with survival outcomes (Supplementary Table S2; Fig. 2). Of the 30 patients treated, all three parameters were reliably measured in 28 patients.
A greater reduction in Ktrans after one dose of cediranib was seen in patients with increased PFS (P = 0.0015) and OS (P = 0.0039; Supplementary Fig. S3A and B). Of note, in two previous phase I studies of the pan-VEGF receptor tyrosine kinase inhibitor vatalanib (Novartis) in patients with colorectal cancer and metastatic liver lesions, favorable outcomes correlated with significantly greater reductions in Ki (a parameter related to Ktrans) at day 2 (18). Similarly, in advanced hepatocellular carcinoma patients, a greater reduction in Ktrans at 14 days after sunitinib (a VEGF receptor tyrosine kinase inhibitor; Pfizer) was significantly associated with increased PFS in a phase II study (19).
In addition, a greater increase in the CBV of tumor microvessels after one dose of cediranib was seen in the glioblastoma patients with extended OS (P = 0.0056; Supplementary Fig. S3C).
Finally, a greater increase in collagen IV levels in plasma was detected in patients with extended PFS (P = 0.0010; Supplementary Fig. S3D).
Association between a VNI and outcome. To test if we could more precisely predict outcomes after cediranib, we then combined these three parameters into a composite index. The index for prediction of OS was calculated as
and the index for prediction of PFS as
where x1, x2, and x3 denote log-transformed ratios of day 1 to baseline levels of, Ktrans, collagen IV, and CBV, respectively.
We found that this VNI closely correlated with PFS (Spearman's ρ = 0.60; P = 0.001) and OS (ρ = 0.54; P = 0.004; Figs. 2C–D and 3). Analysis of residual deviance after sequential addition of variables into Cox regression showed that collagen IV is significant, in addition to Ktrans, for prediction of PFS (P = 0.002, likelihood ratio test); similarly, microvessel CBV is significant, in addition to Ktrans, for prediction of OS (P = 0.008; Supplementary Tables S3–5). Thus, collagen IV and microvessel CBV are important components of these regression models: the residual deviance reduction resulting from addition of both collagen IV and microvessel CBV is higher than the deviance reduction for Ktrans alone.
Next, we stratified patients by score quartiles and found that Kaplan-Meier estimates of PFS and OS were gradually increasing according to the quartile group (Fig. 4). The VNI differentiated patients with favorable versus unfavorable outcome after a single dose of cediranib, indicating that evaluating the degree of tumor “vascular normalization” immediately after anti-VEGF treatment may be potentially used to predict better outcome in glioblastoma.
The prognostic recursive partitioning analysis class was 3 for 12 patients (40%), 4 for 12 patients (40%), and 5 for 6 (20%) of patients. Recursive partitioning analysis value showed only a trend for correlation with OS (P = 0.096), potentially due to the small sample size. In multivariate analysis, the associations between PFS and change at day 1 in Ktrans and circulating collagen IV, and VNI, and between OS and change at day 1 in Ktrans and CBV, and VNI were highly significant (Supplementary Fig. S4, Tables S3, S6, and S7; Fig. 3).
Implications. Although PFS evaluated by volumetric measurements of tumor enhancement correlated with OS in this study (Spearman's ρ = 0.69), it cannot be used as a “predictive biomarker.” The ability to identify tumor-specific changes rapidly after treatment may allow tailoring of therapy to those patients most likely to benefit, and early discontinuation of an ineffective therapy in others. In addition, these biomarkers could also be valuable for treatment of benign tumors (e.g., schwannomas) or nonneoplastic diseases (e.g., macular degeneration) characterized by abnormal vessels (20). The biomarker candidates from this hypothesis-generating study need to be validated in large trials as predictive biomarkers and methodology needs to be established for its clinical use in individual glioblastoma patients.
Disclosure of Potential Conflicts of Interest
A.G. Sorensen: Commercial research grants, AstraZeneca, Exelixis, Genentech, Novartis, Schering-Plough, Siemens, and Takeda-Millennium; consultant/advisory board, EPIX Pharmaceuticals, AstraZeneca, Genentech, Takeda-Millennium, Bayer/Schering, Olea Medical, Mitsubishi, Novartis, and Siemens. T.T. Batchelor: Honoraria from speakers bureau, Vertex, Schering-Plough, and Enzon; consultant/advisory board, AstraZeneca, Genentech, and Takeda-Millennium. R.K. Jain: Commercial research grant, AstraZeneca and Dyax; consultant/advisory board, AstraZeneca, Dyax, Millennium, and SynDevRx. The other authors disclosed no potential conflicts of interest.
Note: Supplementary data for this article are available at Cancer Research Online (http://cancerres.aacrjournals.org/).
Grant support: NIH grants P01-CA80124, R01-CA115767, R21-CA117079, K24-CA125440, R01-CA129371, P41-RR014075, M01-RR-01066, and Federal Share/National Cancer Institute Proton Beam Program Income grants, Harvard Clinical and Translational Science Center, Montesi Family Research Fund, and Damon Runyon Foundation.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
We thank G. Gorospe, C. Koppel, T. Benner, M. Foley, O. Wu, and S. Roberge for outstanding technical assistance.