Bevacizumab Reduces Permeability and Concurrent Temozolomide Delivery in a Subset of Patients with Recurrent Glioblastoma

The dependence of tumor growth on angiogenesis has provided a powerful rationale for antiangiogenic approaches to cancer therapy (1). Targeting blood vessels in glioblastoma (GBM) has been a particularly attractive strategy given the characteristic high degree of endothelial proliferation, hyperpermeability, and proangiogenic growth factor expression (e.g., VEGF) associated with this solid tumor (2, 3). The approval of bevacizumab, which neutralizes VEGF, in December 2017 for recurrent glioblastoma (rGBM) represented the first new therapy for this disease in years. However, no study in GBM has shown improved overall survival in unselected patients when combining bevacizumab or other antiangiogenic therapies with concomitant chemotherapy or radiation and the reasons for this are unclear (4–6).

Antiangiogenic therapy is hypothesized to work by pruning tumor blood vessels and by normalizing those remaining; a process whereby the abnormal, inefficient tumor blood vessels are restored to a more efficient, normalized state resulting in improved delivery of concomitant chemotherapy and reversal of hypoxia (7). The extent and duration of vascular normalization is dependent on the dose of antiangiogenic agents, with higher doses causing excessive pruning, increased hypoxia, and a shorter window of normalization (7, 8). Although two retrospective studies suggested that a lower dose of bevacizumab is associated with improved progression-free survival (PFS) and overall survival (OS) in patients with GBM and one prospective study found a trend toward improved PFS, it is unclear what mechanism(s) are driving these results, what the optimal dose of antiangiogenic therapy is, or if there is a time course to vascular changes during which combination therapy may be more efficacious (9–11). Therefore, we sought to assess how the approved dose of bevacizumab modulates tumor vasculature and impacts the delivery of temozolomide in patients with rGBM to gain further insight into the failure of this drug to improve survival.

Patients with rGBM being treated with temozolomide 50 mg/m²/day and bevacizumab 10 mg/kg every 2 weeks were eligible to participate in this study. The dosing for temozolomide was based on the RESCUE study showing benefit of temozolomide rechallenge in a subset of patients (12). Patients with lower grade tumors that had progressed to GBM were eligible. All patients had to have radiographic evidence of tumor recurrence >2 months since completion of their last cycle of temozolomide or other alkylating agent and be >12 weeks from the completion of radiation. All patients signed informed consent and this trial was approved by the Dana-Farber Harvard Cancer Center Institutional Review Board (NCT01987830) in accordance with U.S. Common Rule. Patients underwent a baseline simultaneous PET-MRI scan (3T Siemens TimTrio with a BrainPET insert) after being on daily temozolomide for at least 5 consecutive days to minimize nonspecific temozolomide binding elsewhere in the body. The second PET-MRI visit was performed one day after the first bevacizumab infusion and the third PET-MRI one day prior to the third bevacizumab infusion (Supplementary Fig. S1).

We adapted a previously described MRI protocol that consisted of scout, T2-weighted sampling perfection with application-optimized contrasts using different flip-angle evolution (T2SPACE) imaging, fluid-attenuated inversion recovery (FLAIR) imaging, dynamic contrast enhanced (DCE) MRI, dynamic susceptibility enhanced (DSC) MRI, diffusion tensor imaging (DTI), pre- and post-contrast T1-weighted, and magnetization-prepared rapid gradient-echo (MPRAGE) images (see Supplementary Data for parameter specifics; ref. 13) An 8- or 32-channel radiofrequency receive coil compatible with the PET was used for the study (14).

Temozolomide was labeled with [11C] as described previously (15). Up to 10 mCi of [¹¹C]Temozolomide was administered intravenously as a 30-second bolus shortly after the start of the PET data acquisition. PET data were acquired for 90 minutes. PET images were reconstructed using the Ordinary Poisson Ordered Subset Expectation Maximization (OP-OSEM) 3D algorithm from prompt and random coincidences, normalization, attenuation, and scatter coincidences sinograms using 16 subsets and 4 iterations. The reconstructed volume consisted of 153 slices with 256 × 256 pixels (1.25 × 1.25 × 1.25 mm³).

The DCE-MRI data were processed using in-house custom software written in MATLAB (MathWorks Inc) to obtain maps of K_trans based on the two-parameter Tofts model (16). Population-level arterial input functions were used for DCE analysis. Gradient echo DSC-MRI data were used to calculate macro-vessel cerebral blood flow (CBF) maps using NordicIce (17). To assess median values within the tumor and peritumoral regions, all structural MRI sequences (FLAIR, MPRAGE pre/postcontrast), parameter maps (K_trans, CBF), and PET standardized uptake value (SUV) maps were registered to the T2SPACE MR images using the BRAINSFit module in 3D Slicer. A deep-learning algorithm (DeepNeuro) was used to initially segment contrast-enhancing tumor on the MPRAGE images (excluding regions of necrosis or blood) and abnormal FLAIR hyperintensity on the FLAIR images (18). These regions of interest (ROI) were reviewed and edited as needed (E.R. Gerstner). Median tumor values for K_trans and CBF were calculated for the contrast-enhancing ROI, the FLAIR ROI, or the combination of the two ROIs. One patient had two separate tumors, so vascular changes within each tumor were evaluated separately to assess whether there were heterogeneous responses.

To identify subregions within the tumor where SUV increased or decreased during bevacizumab therapy, the change in SUV for each follow-up visit was compared with the first visit. T2SPACE MR images from all follow-up visits were registered to the first visit using BRAINSFit module in 3D Slicer with the subsequent transformation applied to all other images and parameter maps in the follow-up visit. All images and parameter maps were downsampled to 3 × 3 × 3 mm³ to more closely mimic native resolution for the DSC-MRI and DCE-MRI data. Brain extraction was performed using the Robust Learning-Based Brain Extraction algorithm (19).

Delta SUV maps were created by subtracting the registered SUV map of each follow-up visit from the SUV map of the first visit. Subregions where SUV decreased (or increased) were defined as regions within the union of the contrast-enhancing ROI and the FLAIR ROI where the SUV decrease (or increase) was greater than 1 SD from the mean decrease (or increase) in normal brain. Changes in K_trans and CBF were then assessed to evaluate the vascular changes within these subregions where temozolomide uptake increased or decreased.

The inability of the transfer constant K_trans to provide a separate estimation of flow and capillary permeability represents a major and well-known limitation (20). In short, in tissues with high permeability, K_trans is ultimately limited by tissue perfusion. Similarly, when permeability is low compared with perfusion, Ktrans is independent of perfusion and reflects the capillary permeability surface area product. Therefore, to better understand the impact of perfusion on uptake of temozolomide, the SUV, K_trans, and tumor ROI data were downsampled to the lower resolution DSC space using the mri_robust_register algorithm in FreeSurfer (21). The contrast-enhancing tumor ROIs were split into low-permeability regions and high-permeability regions. This was performed by identifying the “knee point” of the K_trans distribution per patient (also known as the “elbow point”), where the distribution of the whole-tumor K_trans values in ascending order is best approximated by two straight lines (Supplementary Fig. S2). To increase the robustness of our analyses, image voxels with the 5% highest/lowest values were considered to be outliers and removed before analysis. Within these regions where permeability was low (and, thus, K_trans independent from perfusion), we evaluated the influence of perfusion on temozolomide delivery. Specifically, within the low permeability regions, any changes in SUV will be driven by changes in perfusion rather than permeability.

Wilcoxon signed rank test was used to assess for differences in each imaging parameter across visits. Spearman rank correlation was used to test the correlation between temozolomide uptake as measured by SUV and K_trans or SUV and CBF. The false discovery rate was used to adjust for multiple testing resulting in a False Discovery Rate of 0.1515.

Twelve patients with recurrent GBM were enrolled resulting in 23 evaluable PET scans (Table 1).

When looking within the entire enhancing tumor volume, bevacizumab treatment resulted in an initial decrease in the volume of contrast enhancement (11/12 patients), decrease in median tumor K_trans (9/11 patients), and a decrease in median tumor SUV (6/8 patients; Figs. 1 and 2). In comparison, change in median CBF within the entire enhancing tumor was more variable as a result of bevacizumab treatment (Fig. 2).

To understand the impact of permeability and perfusion on temozolomide uptake within the entire contrast-enhancing tumor, we looked at the correlation between median tumor SUV (our marker of temozolomide uptake) and K_trans or CBF at each visit (Fig. 3). Although there was a positive trend at each visit, only the correlation between SUV and K_trans at visit 1 was significant once corrected for multiple testing (ρ = 0.92, P = 0.0013).

Median whole tumor values fail to account for heterogeneity in GBM, so we evaluated regional changes in temozolomide uptake by looking at changes in perfusion and permeability within areas where the change in SUV was greater than 1 SD above or below the distribution within normal brain. Ten visit pairs were evaluable–either baseline to visit 2 (24 hours post first infusion, N = 5) or baseline to visit 3 (4 weeks after first infusion, N = 5). In areas where SUV decreased, K_trans decreased in 7 of 7 patients at the subsequent visit and CBF decreased in 6 of 7 patients (Fig. 4). In areas where SUVs increased, the change in K_trans and CBF was more variable–5 of 7 experienced an increase in K_trans and 3 of 7 experienced an increase in CBF (Fig. 5). There were no statistically significant changes.

Because permeability (as measured by K_trans) seemed to have a strong impact on temozolomide uptake (as measured by SUV), we evaluated the relationship of SUV and CBF in tumor regions where permeability was low (and, thus, temozolomide uptake would be driven by perfusion). The median K_trans value in the low permeability regions was 0.035 (range 0.009–0.224). For comparison, median K_trans value in the high permeability regions was 0.102 (range 0.034–0.706)–similar to data from a prior study in GBM (22). Across all visits and for all evaluable tumor voxels, the median relative size of the low permeability region was 64% (range 39%–83%) of the tumor. In the low permeability regions, macrovessel CBF was positively correlated with SUV although statistical significance was not met once adjusted for multiple testing (Fig. 5).

Bevacizumab modulated tumor vascular structure and function resulting in a change in the delivery of temozolomide in patients with rGBM. Within the entire contrast-enhancing tumor volume as well as in subregions of the tumor where there was a decrease in temozolomide uptake, permeability had a strong influence on drug delivery. However, in subregions where permeability was low, the correlation between increased perfusion and increased temozolomide uptake became stronger, suggesting that in areas where the integrity of the blood—brain barrier (BBB) is not completely compromised, increased perfusion may overcome decreased permeability to impact temozolomide delivery. These results suggest that if perfusion can be improved, drug delivery to areas of the tumor with an inefficient vasculature may be improved. However, because these low permeability regions accounted for only about half of the enhancing tumor volume, the clinical significance of this finding is difficult to determine. As a proof-of-concept study, though, our results suggest that we can reliably measure vascular changes in humans and the impact those changes have on drug delivery–opening up the possibility for noninvasively improving our understanding of future drug combination studies.

Our results may offer one explanation for why, despite the perceived therapeutic potential of antiangiogenic therapies in GBM, several brain tumor studies using bevacizumab or other antiangiogenic therapies have failed to show a survival benefit when combined with chemotherapy (4–6, 23). Antiangiogenic therapy induced vascular remodeling, reverting the abnormal tumor vasculature to a more normal state, is hypothesized to improve drug delivery by providing more uniform delivery of chemotherapy. This concept has been supported in preclinical models but is challenging to confirm in humans (7, 24, 25). Our results suggest that a decrease in permeability appears to be the predominant response of tumor vasculature to bevacizumab therapy at the standard clinical dose for GBM, potentially decreasing penetration of concomitant chemotherapy. Even though temozolomide crosses into the brain via passive diffusion, most patients in our study had a decreases in whole tumor temozolomide uptake when given in conjunction with bevacizumab, suggesting that the diffusion of temozolomide was impaired with restoration of vascular permeability (22). These findings are consistent with a study in patients with lung cancer where an even higher dose of bevacizumab (15 mg/kg) resulted in decreased perfusion and decreased drug penetration as well as additional animal and modeling studies demonstrating that permeability of molecules can decrease as distance from the contrast-enhancing tumor core increases (26–29).

Lower doses of antiangiogenic therapy have been hypothesized to have a more profound normalizing effect on tumor vasculature, but a trial combining low-dose bevacizumab with lomustine failed to show improved progression-free survival over standard, higher dose bevacizumab monotherapy (11, 30). However, although this study was not designed to specifically evaluate patients at first recurrence, a strong trend toward improved PFS was seen in that subgroup treated with combination low-dose bevacizumab plus lomustine and other retrospective studies suggested improved PFS and OS with lower doses of bevacizumab. Thus, although antiangiogenic therapy appears to clinically benefit some patients with rGBM, the duration of response is transient and, at least at the dose and schedules currently in clinical practice, the fraction of tumor where normalization could lead to improved drug delivery may be insufficient to effectively target the tumor. With the small number of patients included in this study, although, we may not have captured patients with more profound examples of normalization since, based on our prior work with cediranib, an oral pan-VEGF receptor tyrosine kinase inhibitor, only 25% of patients experienced vascular normalization with increased perfusion (13).

Our study demonstrates the rapid and heterogenous impact that bevacizumab has on GBM vasculature and decreased permeability seems to be the driving force limiting concomitant drug delivery. Thus, at least in a subset of patients with rGBM, standard dose bevacizumab may impede concomitant chemotherapy. In areas where permeability is low, perfusion may influence drug delivery, but this was only observed in subregions of the entire tumor volume. Future efforts in the development of combination drug strategies should focus on how to balance change in permeability with maximizing tumor perfusion by optimizing dosing of bevacizumab. Using multimodal imaging tools holds great promise to aid our understanding of changes in tumor biology during treatment.

K.E. Emblem reports intellectual property rights at NordicNeuroLab AS in Bergen, Norway. D.G. Duda is an employee/paid consultant for Bayer, Tilos, twoXAR, and BMS, and reports receiving commercial research grants from Bayer, Exelixis, BMS, Leap, and Merrimack. R.K. Jain is an employee/paid consultant for Chugai, holds ownership interest (including patents) at Enlight and SynDevRx, and is an advisory board member/unpaid consultant for Ophthotech, SPARC, SynDevRx, XTuit, Merck, the Boards of Trustees of Tekla Healthcare Investors, Tekla Life Sciences Investors, Tekla Healthcare Opportunities Fund, and Tekla World Healthcare Fund. T. Batchelor is an employee/paid consultant for Genomicare, and reports receiving other commercial research support from UpToDate, Inc., Oakstone Publishing, Oncology Audio Digest, Champions Biotechnology, Amgen, NXDC, Upsher Smith, and Merck. All other authors declare no potential conflicts of interest by the other authors.

Conception and design: E.R. Gerstner, K.E. Emblem, J.M. Hooker, D.G. Duda, R.K. Jain, T. Batchelor

Development of methodology: E.R. Gerstner, K.E. Emblem, K. Chang, J.M. Hooker, D.G. Duda, T. Batchelor

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E.R. Gerstner, Y.-F. Yen, J. Dietrich, S.R. Plotkin, T. Batchelor

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E.R. Gerstner, K.E. Emblem, K. Chang, B. Vakulenko-Lagun, Y.-F. Yen, A.L. Beers, C. Catana, J.M. Hooker, D.G. Duda, J. Kalpathy-Cramer, R.K. Jain, T. Batchelor

Writing, review, and/or revision of the manuscript: E.R. Gerstner, K.E. Emblem, K. Chang, Y.-F. Yen, A.L. Beers, J. Dietrich, S.R. Plotkin, C. Catana, J.M. Hooker, D.G. Duda, B. Rosen, J. Kalpathy-Cramer, R.K. Jain, T. Batchelor

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E.R. Gerstner, A.L. Beers, J. Kalpathy-Cramer, T. Batchelor

Study supervision: E.R. Gerstner, J.M. Hooker, T. Batchelor

This work was supported by K23CA169021 (to E.R. Gerstner), 1R01CA129371 (to T. Batchelor), the European Research Council (ERC) grant 758657, the South-Eastern Norway Regional Health Authority grants 2017073 and 2013069, and the Research Council of Norway grant 261984 (to K.E. Emblem). This work was also supported by grants S10RR023043 and P41RR14075.

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.