2.1.

Orthotopic Murine Model of PDAC

This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal experiments were approved by the Massachusetts General Hospital Institutional Animal Care and Use Committee (IACUC). We used an orthotopic xenograft mouse model for PDAC established in our laboratory.³² For tumor implantation, a small left abdominal flank incision was made to exteriorize the pancreas of athymic Swiss male Nu/Nu mice (20 to 25 g and 6 to 8 weeks old; Cox Breeding Laboratories, Cambridge, Massachusetts). Subsequently, a suspension of 10⁶ AsPC1 human PDAC cells in 30 μL phosphate-buffered solution (PBS, without ${\mathrm{Ca}}^{2+}/{\mathrm{Mg}}^{2+}$), containing 50% Matrigel™ Basement membrane matrix (BD Biosciences, Franklin Lakes, New Jersey) to prevent leakage of the tumor cells from the injection site was injected into the body of the pancreas. Any growth factors present in the Matrigel can potentially help “jump-start” tumor vascularization. Because Matrigel was used in all implantations, any potential angiogenic effects were captured in the untreated controls and uniformly influenced all tumors in the study. We allowed the tumor to grow for 10 days prior to the start of therapy. The terminal half-life for growth factors is typically just a few days. Hence, these components are presumably cleared and metabolized prior to day 10 when we commenced therapy.

2.2.

Treatments of Orthotopic PDAC Mice

Ten days after the injection of AsPC1 cells, animals were randomized into the following groups: (1) no treatment; (2) a single dose of $0.05\mathrm{mg}/\mathrm{kg}$ body weight cabozantinib (XL184) for antiangiogenic therapy; (3) a single dose of $0.25\mathrm{mg}/\mathrm{kg}$ body weight liposomal benzoporphyrin derivative (BPD) for antivascular near infrared (NIR) photodynamic therapy (PDT; $75\mathrm{J}·{\mathrm{cm}}^{-2}$); and, (4) a single dose of $0.25\mathrm{mg}/\mathrm{kg}$ body weight BPD and $0.05\mathrm{mg}/\mathrm{kg}$ body weight cabozantinib coencapsulated in a liposome for combined antiangiogenic treatment and antivascular NIR PDT. All injections were done intravenously in 200 μL sterile PBS. Mice were illuminated 90 min after BPD injection. The 690-nm light was delivered at an irradiance of $150\mathrm{mW}·{\mathrm{cm}}^{-2}$. Tumor excision and TMV imaging can be applied at any time point, including within a day of antivascular agent administration, and we chose established time points used in the seminal publication demonstrating the clinical efficacy of bevacizumab ($\sim 2\text{weeks}$ post-treatment)³³ that reflects the most common protocol for MVD quantification. Note that, here, we did not attempt to distinguish primary effects of the treatments on the microvasculature versus secondary effects resulting from inhibition of cancer cell signaling pathways and the killing of cancer cells. It is virtually impossible to do so unequivocally due to the tight interdependence of cancer cells and microvasculature,²⁹ however, it is known that direct effects on the microvasculature are best captured at early time points.^9,11

2.3.

Staining Protocol for Serial Tumor Sections

An overview of the immunofluorescence staining protocol is shown [Fig. 2(a)]. Orthotopic primary tumors were excised 2 weeks post-treatment, embedded in optimal cutting temperature compound and frozen at $-80°\mathrm{C}$. A cryotome was used to cut the entire tumor into 388, 20-μm-thick cryosections. Approximately one-quarter of the sections, 105 sections (one section/70 μm), were (1) fixed in a precooled mixture of $1:1$ acetone:methanol for 15 min at $-20°\mathrm{C}$, (2) air dried for 30 min at room temperature, and (3) washed three times, for 5 min each, in PBS with gentle agitation. A Pap Pen was used to encircle the tissue specimens for mounting small-volume blocking and staining solutions. A blocking solution (Dako Protein Block Reagent) was applied for 30 min at room temperature, followed by application of the immunostains, at $\sim 5\mu \text{g}/\mathrm{mL}$ monoclonal antibody (MAb) each diluted in a background reducing buffer (Dako Antibody Diluent, background reducing) for 2 h at room temperature in a humidifying chamber. Finally, the slides were washed again three times, for 5 min each, in PBS with gentle agitation, mounted (Invitrogen SlowFade Gold with 4′,6-diamidino-2-phenylindole, DAPI) with a coverslip and sealed with nail polish.

Fig. 2

Methods overview. (a) Work flow for tissue fixation and immunofluorescence staining. (b) Simplified image processing outline for identifying the tumor boundary and intratumoral microvessels. The microvessel content is normalized by the cell count within the tumor (using cell nuclei staining) and multiplied by the tumor volume to calculate the tumor microvessel volume (${\mathrm{mm}}^3$) for an image mosaic of an entire tumor cross-section. The overall TMV is calculated by interpolation and summation of microvessel volume over a periodic subsampling of the entire tumor.

2.4.

Confocal Fluorescence Imaging Microscopy

Confocal fluorescence imaging was performed using a commercial confocal microscope (Olympus FluoView 1000) with a $10\times$ 0.4 numerical aperture (NA) or a $20\times$ 0.75 NA objective. Excitation of DAPI, anti-human cytokeratin 8 (clone LP3K, R&D Systems, Minneapolis, Minnesota) MAb-Alexa Fluor 488 conjugates, and anti-mouse PECAM-1 (CD31; clone 390, Millipore, ) MAb-Alexa Fluor 568 conjugates was carried out using 405-, 488- and 559-nm lasers, respectively. Lasers were scanned sequentially to reduce channel crosstalk. Mosaic images of entire tumor cross-sections were collected and stitched together using the Olympus FluoView software. This method requires a manual procedure to define tumor boundaries for the mosaic. We developed a human cytoskeletal protein (cytokeratin 8) stain to demark the intracellular space of the PDAC cells, which has dual selectivity for the epithelial cancer cells because it does not react with mouse proteins and because cytokeratins are highly expressed in many epithelial cancer cells.³⁴ CD31 (endothelial cell) and DAPI (cell nuclei) staining may be applied to quantify MVD (microvessel area is normalized by the number of cells per field).

2.5.

Slide-Scanning Wide-Field Fluorescence Imaging Microscopy

In order to efficiently image the 105 sections from an entire, orthotopic PDAC tumor, we applied a commercial slide-scanning fluorescence imaging system (Hamamatsu NanoZoomer 2.0-RS) with a $20\times$ 0.75 NA objective. The tumor sections were imaged in $40\times$ mode ($0.23\mu \text{m}·{\text{pixel}}^{-1}$).

2.6.

Image Analyses

All analyses were performed using custom MATLAB (Mathworks, Natick, Massachusetts) routines for batch processing [Fig. 2(b)]. For display and all analyses, the channels were autothresholded to reject background signal using a 10% peak intensity threshold for each channel. Wavelet transforms and image correlation coefficients among the color channels were used to eliminate bubbles and extraneous, reflective objects from the analyses. Current microvessel segmentation methods were incapable of extracting pertinent image features from image mosaics of entire tumor sections, which contain gradient backgrounds and occasional artifacts normally avoided when manually choosing individual, “representative” fields. Therefore, we developed wavelet multiresolution decomposition analyses³⁵ for morphology-based microvessel and cancer cell segmentation. Here, MVD values (unitless) were calculated from whole slices, within viable tumor tissue only, and averaged over slices from the entire tumor rather than a more complex “hot spot” identification and calculation, which is difficult to define objectively. TMV was calculated by multiplying MVD with the viable tumor volume [Fig. 2(b)] in each slice and then summing TMV over the whole tumor by interpolation. Thus, TMV represents the total tumor endothelial cell content in units of volume, an absolute metric, rather than the fractional volume of microvessels.

2.7.

Mathematical Model for Defining Tissue Volume Subsampling to Maximize Throughput Achieving Statistical Significance

The determination of a biomarker, such as tumor TMV, from only partial sampling of an organ or tumor leads to uncertainty in the true value. We estimated this uncertainty by using the spatial autocorrelation function of the biomarker to characterize the fluctuations between neighboring slices. These fluctuations are then used to quantify the overall noise introduced from volume subsampling. Assume an organ or tumor with $N$ equally spaced sections. $f_i$ is the biomarker level contained in slice $i$ and is assumed to be stationary and ergodic. We model the difference between neighboring slices as

It is found that $A=G(1)$. Using the stationarity conditions

2.8.

Statistical Analysis

Treatment response data (tumor weight, MVD, and TMV) among multiple treatment groups were analyzed using a Kruskal–Wallis one-way analysis of variance (ANOVA). The Mann–Whitney $U$ test was used to report $P$-values for specific comparisons that are not statistically significant ($P>0.05$) according to the Kruskal–Wallis ANOVA.

3.1.

Imaging MVD and Total Microvasculature in Whole Tumors

We monitored tumor microvasculature utilizing custom immunofluorescence stains and image analyses in tandem with tissue sectioning and multiarea microscopy to quantitatively image microvessels systematically throughout entire tumors [Fig. 3(a)]. The immunostains identify microvessels (endothelial cells) and epithelial cancer cells in subcellular resolution image mosaics of full tumor cross-sections [Figs. 3(b)–3(d)]. The cancer cell stain, using an anticytokeratin MAb, excludes acellular regions of necrosis and defines the boundary of viable tumor cells in serial sections to enable objective, automated segmentation of the intratumoral microvessels [Figs. 3(b), 3(e), and 4]. This identification of the tumor margin is critical as tumors are often poorly vascularized in comparison to the surrounding tissue²⁹ [Fig. 3(e)]. In the absence of a clear visualization of the tumor margin, the dense peritumoral vasculature can confound the quantification of intratumoral microvessels. Furthermore, MVD measurements often rely on capturing changes in foci of endothelial cell proliferation, or hot spots, in a few high-magnification fields²⁶ such that MVD statistics can suffer from interobserver variability as well as tumor heterogeneity.³⁰ In contrast, TMV imaging estimates the total vessel volume from serial, multifield mosaics of entire tumor cross-sections for automated and objective assessments. Note that this critical identification of the tumor boundary also excludes the possibility of simply pulverizing whole tumors for biochemical assays (e.g., Western blots) because the peritumoral microvasculature obscures the quantification of intratumoral vasculature.

Fig. 3

Three-dimensional (3-D) tumor MVD and TMV imaging. (a) A 3-D view of select slices from a dataset of 105 PDAC tumor sections. (b) A single tumor cross-section imaged with 1.2 μm $x\text{-}y$ sampling illustrates selective immunostaining of mouse endothelial cells (magenta) and human epithelial cancer cells (green) to discriminate the tumor margin. Scale bar, 1 mm. (c and d) Enlarged views of the region marked in b, demonstrating visualization of (c) 4′,6-Diamidino-2-phenylindole (DAPI) staining of cell nuclei and (d) immunostaining of individual microvessels and cancer cells at the tumor margin. Scale bars, 250 μm. (e) The cancer cell immunostain defines the tumor boundary and identifies peritumoral (gray scale) versus intratumoral microvessels (magenta). Scale bar, 1 mm.

Fig. 4

Wavelet-based segmentation of microvessels. Wavelet spatial frequency bandpass filtering enables automated selection of intratumoral microvessels based on morphology. This is advantageous for high-throughput extraction of immunostains from gradient backgrounds and with variable contrast over many tissue sections, as was the case here. Here, intratumoral microvessels (magenta outline) identified by software are overlaid on the raw fluorescence image of mouse endothelial cells (gray scale). (a) Selection of intratumoral microvessels using intensity thresholding fails in high-throughput applications. (b) In contrast, wavelet bandpass filtering is more robust and identifies a greater number of intratumoral microvessels.

The anti-human cytokeratin MAb for human cancer cells applied here is widely applicable to mouse xenograft models of epithelial cancers. This immunostain is selective for human cells and for cytokeratin proteins highly expressed in carcinomas, which provides a dual selectivity for clear demarcation of the tumor boundary. Prior clinical reports have demonstrated cytokeratin staining to image and classify epithelial cancers³⁶ and micrometastases.³⁷ Other MAbs-targeting biomarkers frequently overexpressed by cancer cells (e.g., epidermal growth factor receptors, folate receptors, and epithelial cell adhesion molecule) are also being used successfully to image and quantify tumor margins³⁸ and ultra-rare circulating tumor cells in the clinic.³⁹ Note that some tumor types pose difficulties for specific cancer cell staining and careful consideration will be required to design stains for other applications.

3.2.

Mathematical Model for Maximizing Vascular Biomarker Quantification Throughput Achieving Statistical Significance

For maximal statistical power, imaging the full organ or tumor microvascular network is ideal but prohibitive. For ample, processing an entire 1-cm-diameter tissue specimen requires 170 h and 4 terabytes (TB) of data acquisition (using a Hamamatsu NanoZoomer 2.0-RS fluorescence imaging slide-scanner). However, spatial correlations of the microvasculature can greatly reduce the sampling burden. To guide efficient studies, we developed a theoretical model (Methods) to calculate the minimum number of tissue slices (tissue subsampling) needed to achieve an accurate estimate for a single organ or tumor and for an aggregate group of organs or tumors receiving a treatment regimen. The model is tunable to biological model-specific vascular heterogeneity. The appropriate subsampling can be determined from metrics of vasculature heterogeneity taken from control tissue specimens in order to balance the tradeoff between statistical confidence and experimental throughput. To validate the model in our exemplary tumor TMV biomarker imaging application, we analyzed an entire orthotopic PDAC tumor (8 mm in size) in steps of $\sim 70\mu \text{m}$ to calculate the “true” tumor TMV. We next calculated tumor TMV accuracy for varying subsamplings of the full tumor [e.g., decreasing numbers of ordered serial sections; Figs. 5(a)–5(c)] and found good agreement between empirical and theoretical confidence intervals [$R^2=0.63$, $n=58$ tumor subsampling sizes, no fitting or free parameters; Fig. 5(c)]. The mathematical model is also applicable to models of nonmalignant disease, where the microvascular biomarker heterogeneity of organ or tumor samples can be determined and used to guide the choice of an optimal tissue subsampling frequency.

Fig. 5

Mathematical model for achieving maximal throughput with statistical significance. (a) Schematic of an increasing number of ordered tumor slices used to estimate tumor TMV. (b) Autocorrelation of the tumor microvasculature along the depth of an orthotopic PDAC tumor. The autocorrelation decay rate determines the optimal tumor subsampling. (c) Convergence of tumor TMV imaging accuracy as more of the tumor volume is imaged. The points are tumor TMV, $\pm 95\%$ confidence interval, generated by repeating the TMV calculation on permuted phase shifts of the ordered slices. Empirical confidence intervals are shown for comparison to theoretical confidence intervals generated by the mathematical model. (d) Exemplary relationship between the minimum detectable difference in tumor TMV, ${\delta }_{\mathrm{bm}}$, and imaging an increasing number of tumors, with individual tumor samplings of 3 (0.8% of the tumor volume) or 10 (2.6% of the tumor volume) slices per tumor.

We then applied the mathematical model to determine appropriate tissue subsamplings and sample sizes (number of independent tissue specimens or tumors in this case) to resolve a desired antivascular effect size. The autocorrelation analysis characterizes vascular biomarker heterogeneity and defines the subsampling distance containing nonredundant information [Fig. 5(b)]. We define the nonredundant subsampling for a single-tissue specimen as half the decay length, ${(2e)}^{-1}$, of the autocorrelation function. Here, the decay length for tumor TMV in the PDAC model is 450 μm [Fig. 5(b)], and tumor TMV accuracy reaches 85% confidence at a sampling period of 225 μm [8% of the total tumor volume, Fig. 5(c)]. Because a typical experiment includes multiple mice, it is not necessary to image at the subsampling required for a single tissue specimen; i.e., confidence improves with more specimens.

As a practical guide for subsampling, a statistical calculator is provided (Appendix) that incorporates inter- and intratissue specimen vascular heterogeneity, desired effect size resolution, and sample size. The output of the calculation for the PDAC model imaged here is provided for illustrative purposes [Fig. 5(d)]. In this case, 2.6% of the PDAC tumor volume (10 slices) must be imaged per tumor for a sample size of five mice ($\sim 50$ sections total) to achieve 95% confidence ($P<0.05$), with a resolving power (minimum detectable difference) of a 20% change in tumor TMV. Thus investigators may plan experiments according to the desired statistical power, effect size resolution, and biomarker spatial heterogeneity. In contrast to periodic tumor subsampling, full 3-D reconstructions would require $\sim 38$ times ($100\%/2.6\%$) more labor and data storage ($\sim 1900$ sections totally). The requirement to image only a few percent of the tumor, in order to achieve sufficient statistical power to report on the whole tumor, is good news as it accelerates analysis while reducing cost. This offers promise for developing high-throughput assays.

3.3.

Efficient Quantification of Antivascular Therapeutic Response

To demonstrate the strengths of full tumor subsampling, we performed a comparative study of MVD versus TMV in an orthotopic mouse model of PDAC subjected to different modes of treatment, (Fig. 6) resulting in either minimal or significant tumor reduction [Fig. 7(a)]. Here, MVD (dimensionless) is the mean value across the tumor depth, whereas TMV (${\mathrm{mm}}^3$) accumulates with depth [Fig. 2(b)]. In a first experiment, we treated PDAC mice with cabozantinib (XL184), a small molecule, multireceptor tyrosine kinase (RTK) inhibitor with potent activity against MET (the receptor for hepatocyte growth factor) and vascular endothelial growth factor receptor 2 (VEGFR2), as well as other RTKs important for tumor pathobiology.⁴⁰ Cabozantinib inhibits tubule formation, with no evidence of direct cytotoxicity.⁴⁰ Thus, cabozantinib is classified as an antiangiogenic agent and its effects on nonvascular cells do not directly cause cell death.⁴⁰ McDonald et al. have elegantly demonstrated that cabozantinib is a potent antiangiogenic agent with the added benefit of reducing a prominent escape mechanism—cancer cell migration, metastasis and growth along remaining, functional vessels as well as via lymphatic routes.^4,41,42 Cabozantinib is currently in phase III clinical trials based upon its induction of antiangiogenic (through VEGFR2 inhibition) and anti-invasive (through MET inhibition) effects in preclinical models⁴¹ and in pilot clinical trials.⁴³

Fig. 6

Schematic of the treatment modalities. A simplified schematic of cell surface multireceptor tyrosine kinase (RTK) signaling triggering by growth factor (labeled ‘g’) binding and downstream, intracellular signaling cascades involving adapter proteins and other mediators (labeled with asterisks). Cabozantinib binding to RTKs leads to decreased angiogenic, proliferation, and survival signaling. Benzoporphyrin derivative (BPD)-photodynamic therapy (PDT) primarily damages mitochondria (labeled ‘m’) and its toxicity is due, in part, to subsequent apoptotic signaling cascades.

Fig. 7

TMV is more sensitive than MVD imaging for reporting antivascular response. TMV imaging reports reductions in microvasculature following cabozantinib treatment, PDT, and their combination in orthotopic PDAC tumors while, in contrast, MVD imaging under-reports vascular reduction. (a) Changes in orthotopic PDAC tumor weight 2 weeks post-treatment with the monotherapies and combined PDT and cabozantinib from the data shown in (c–f) below. (b) Corresponding percent changes 2 weeks post-treatment in tumor MVD and TMV relative to no-treatment control mice. (c) The peritumoral MVD is larger than the intratumoral MVD, and the peritumoral MVD is unaffected by treatment, illustrating the importance of identifying intratumoral microvessels. (d) Intratumoral MVD shows no statistically significant changes post-treatment compared to untreated, control mice. (e) Tumor TMV is reduced for all cases and resolves statistically significant antivascular affects. (f) Significant tumor weight reduction is apparent for the treatment groups involving PDT. The results are $\text{mean}\pm \text{s.e.m.}$, $n\ge 16$ mice (a and f) or $n\ge 7$ mice (b–e) per treatment group. Asterisks denote statistically significant differences compared to control, untreated PDAC mice, or among the indicated treatment groups, *$P<0.05$ **$P<0.01$ ***$P<0.001$ [Kruskal–Wallis one-way analysis of variance (ANOVA) test]. In (b) and (e), note that the reason for the lower $P$-value for the cabozantinib alone group [$n=7$ mice; versus the “combo” group]—compared to the PDT alone group ($n=18$ mice; versus the “combo” group)—is that the cabozantinib group has fewer mice.

We then performed a classic antivascular treatment, PDT, on a separate group of PDAC mice using BPD activated by NIR light. PDT is a clinically utilized modality that can be utilized to directly damage vasculature. It is used routinely in the clinic for the targeted closure of neovessels in age-related macular degeneration,^44,45 which is characterized by choroidal neomicrovascularization, while sparing sensitive neighboring tissues in the eye. PDT is also approved for primary and adjuvant treatment of certain cancers.^45,46 Depending on the time of illumination following BPD administration, photodynamic action can selectively destroy vasculature, cancer cells or both.^47–49 Here, we purposefully induced both vascular and cancer cell destruction in order to elicit a change in both overall tumor volume and microvasculature. This experimental arm tests the ability of MVD and TMV to capture antivascular response in cases where the tumor shrinks. A third group of mice received combined PDT and cabozantinib to probe for complementary effects.

The single dose of cabozantinib did not affect the primary tumor burden [Fig. 7(a)], however, a trend was observed in reduced tumor MVD ($P=0.19$, $n=7$ mice, 14 entire tumor cross-section image mosaics; Mann–Whitney $U$ test) and reduced tumor TMV ($P=0.15$, $n=7$ mice, 14 image mosaics; Mann–Whitney $U$ test) compared to control, untreated PDAC mice [$n=8$ mice, 16 image mosaics; Fig. 7(b)]. As expected, both MVD and TMV are valid metrics of antivascular activity in cases where the tumor volume is not significantly impacted. However, upon a 36% reduction in tumor weight following PDT [$P<0.001$, $n=8$ mice; Kruskal–Wallis analysis of variance, ANOVA; Fig. 7(a)], tumor MVD remains unchanged while tumor TMV indicates a 39% reduction in the tumor microvasculature ($P=0.13$, $n=18$ mice, 36 image mosaics; Mann–Whitney $U$ test). The final group of mice received a combination of PDT and cabozantinib, which reduced the primary tumor burden by 52% [$P<0.001$, $n=10$ mice; Kruskal–Wallis ANOVA; Fig. 7(a)], an enhanced tumor reduction of 17% over PDT alone ($P<0.01$; Kruskal–Wallis ANOVA). Both tumor MVD and TMV imaging indicated an antivascular effect [Fig. 7(b)] with a reduction in MVD ($P=0.2$, $n=10$ mice, 20 image mosaics; Mann–Whitney $U$ test] and in TMV [$P<0.001$, $n=10$ mice, 20 image mosaics; Kruskal–Wallis ANOVA]. However, only tumor TMV imaging was able to resolve an antivascular effect of statistical significance and also revealed a statistically significant increase in antivascular activity over PDT alone ($P<0.05$; Kruskal–Wallis ANOVA). The data used to calculate percent changes in MVD, TMV, and tumor weight are presented in Figs. 7(c)–7(f). Collectively, TMV reported antivascular activity in all cases while MVD failed to detect reductions in the tumor vasculature when the tumor shrank in size, which tends to compact the residual microvessel content. Furthermore, TMV measurements resulted in greater statistical power than MVD measurements. These data demonstrate that TMV imaging is more suitable than MVD measurement for quantifying antivascular efficacy.