The Pervasive Presence of Fluctuating Oxygenation in Tumors

Tumor hypoxia is one of the most studied physiologic phenomena in cancer research and has been shown to be prognostically significant in many clinical studies, independent of treatment type (1–3). Since the discovery of diffusion limited tumor hypoxia in the 1960s, most strategies for overcoming or targeting hypoxic areas in tumors have been developed using a paradigm of chronically hypoxic areas in tumors or areas of hypoxia that are oxygen-deficient because they are beyond the typical diffusion limit of O₂. Although many investigators have measured the presence of so-called intermittent or acute hypoxia in preclinical (4–11) and clinical tumors (12), instabilities in tumor oxygenation were presumed to be temporally rare, spatially isolated events attributed to transient episodes of vascular stasis (6, 7, 13–16).

Our group previously showed that vascular stasis was not necessary for tissue to experience transient hypoxia; alternatively, fluctuations in red cell flux could easily cause this effect (17). Based on a series of studies, we have advocated the concept that instabilities in red cell flux are the norm within tumors and that this condition could lead to widespread instability in oxygenation throughout a tumor. Because oxidative stress has been shown to cause a differential cellular response for intermittent versus chronic hypoxia in tumor and endothelial cells (18–25), the pervasive presence of fluctuating oxygenation in tumors has consequential implications for our understanding of tumor progression, stress response, and signal transduction; in all studies, intermittent hypoxia has been shown to increase molecular or physiologic responses in a manner consistent with more malignant tumor phenotypes (18–25).

In this study, we show data directly measuring temporal fluctuations in vascular pO₂ in three rat tumors, characterizing the spatial and temporal differences. We show that O₂ delivery to tumors is constantly fluctuating, resulting in reoxygenation events throughout the tumor.

Dorsal skin flap window chamber. Tumors were grown in the dorsal skin flap window chambers in female Fischer 344 rats (150–175 g). The rats were surgically implanted with titanium window chambers as previously described (26). A small piece of tumor ∼1 mm³ from a donor animal was transplanted onto the fascia at the time of surgery, and glass windows were placed on both sides of the chamber. Three rat tumor lines were studied: R3230Ac mammary adenocarcinoma (n = 6), FSA fibrosarcoma (n = 6; ref. 27), and 9L glioma (n = 6; ref. 28).

At 2 to 4 d after surgery, animals were shipped from Duke University Medical Center to University of Pennsylvania and allowed to acclimate for 4 to 7 days before any measurements were taken. All experimental protocols were approved by Duke University Medical Center and University of Pennsylvania Institutional Animal Care and Use Committees.

Phosphorescence lifetime imaging. Phosphorescence lifetime imaging (PLI) was performed on the tumors in the dorsal window chamber using a previously described imaging set-up (29).

Experimental protocol. Tumors were selected for imaging as they reached a diameter of ∼3 mm. Rats were anesthetized with 50 mg/kg pentobarbital i.p. for imaging; this dose is known to maintain stable heart rate and blood pressure within a range reported for unanesthetized rats (30) over a period of several hours (11) and we have previously shown that blood oxygen content is equivalent to unanesthetized rats using this method (31). Before taking any measurements, the femoral vein was cannulated and the rat was placed in lateral recumbancy on a temperature-controlled heating pad onto the stage with the tumor side of the window chamber facing the camera. The window chamber was secured using a custom-made holder attached directly to the stage so that the glass window laid flat. Once secured and immobile, neither the rat nor the microscope stage were touched to maintain the same location for sequential images on the camera.

0.3 mL of 8 mg/kg phosphor G₂ (32) was then injected i.v. into the animal, and imaging began. The field of view for the image was chosen randomly within the tumor, although care was taken to ensure that the entire field of view was tumor tissue and not normal tissue. Images were taken every 2.5 min for 60 to 90 min before the animal was sacrificed with an overdose of pentobarbital i.v., and a final image was taken. Experiments in which the pO₂ did not drop to near zero after the animal was sacrificed were discarded (n = 2; these experiments are not included in the n = 6 for each tumor type).

Phosphor G₂ was excited using light at 450 nm, which penetrates ∼50 μm into the tumor tissue (32); therefore, PLI images reflect the oxygenation status of vessels near the surface of the tumor.

Image analysis. Each PLI image yielded a 2.2 × 2.85 mm (480 × 752 pixels) map of vascular pO₂ at each time point. The maps of vascular pO₂ were imported into MATLAB, and all further analyses were done in MATLAB. Pixels from the original maps were averaged over every 4 × 4 pixels and a new, averaged map of vascular pO₂ was created (120 × 188 pixels; each pixel is 15.2 × 18.3 μm). This averaged map of vascular pO₂ was used for all subsequent analyses. Analyses are presented in terms of pixel sizes; the light signal from the tissue scatters laterally before it is captured by the PLI camera, which is several centimeters away from the surface of the window. Therefore, the signal from any given pixel contains a contribution from the surrounding pixels.

In an initial exploratory analysis, sequential time images were subtracted, i.e., time point 1 subtracted from time point 2, for all images. The results for each pixel were then plotted as either a positive change >1 mm Hg, a negative change >1 mm Hg, or no change. This analysis was also conducted on the difference images using a threshold of change of 5 mm Hg.

Spatial statistics. To more robustly examine spatial patterns of vascular pO₂, watershed segmentation analysis for spatial change detection was conducted for each experiment. First, the absolute difference between the time-averaged pO₂ for the entire experiment and each individual time point was calculated. The median value of each individual absolute difference image was then subtracted from the image, and watershed segmentation was performed on a pixel by pixel basis using an eight-connected neighborhood. Watershed analysis detects gradients in pO₂ values and segments an image along those gradients; segmented regions can be thought of as pO₂ isobars.

Global Moran's I analysis (33, 34) was performed on all time points for all images to examine if spatial autocorrelations were consistently occurring within an image for disc-shaped spatial ranges with radii of 1, 4, 8, 12, 20, 40, and 100 pixels using a binary weighing function (Supplementary Fig. S1). Global Moran's I analysis was also performed on differences in sequential time images using the same spatial ranges and weighing function as described above.

A two-way ANOVA test was conducted on the resulting Moran's I values to determine if different tumor types had different patterns of spatial autocorrelations. Multiple comparison of the means was done to determine which groups were significantly different using the Bonferroni critical value of P < 0.05.

Spatial autocorrelation within a single image was examined using local Moran's I (35). Each time point was checked for local spatial autocorrelations within disc-shaped regions with radii of 1, 4, 8, 12, 20, and 40 pixels using a binary weighing function. Additionally, subsequent time points were subtracted and changes in pO₂ between two time points were examined for local spatial autocorrelations using local Moran's I.

Finally, the averaged map of vascular pO₂ was subjected to discrete Fourier transform to determine the dominant period of fluctuations for each pixel over time. Results from high frequency fluctuations (<5 min) were discarded as the sampling rate of once every 2.5 min could not resolve these high frequencies.

The median vascular pO₂ for all 9L images was 38 (26–40; all values reported are medians and 25–75%) mm Hg, 30 (26–35) mm Hg for FSA, and 22 (19–23) mm Hg for R3230. There was no clear trend over time of increasing or decreasing median pO₂ of the entire image for all the experiments (P > 0.05). The representative experiments shown had a median pO₂ of 34 (27–41) for 9L, 21 (16–25) for FSA, and 18 (16–21) mm Hg for R3230 (Fig. 1). These values are higher than previously measured tissue pO₂ values in these same tumor lines (11), consistent with the imaging phosphor largely remaining in the vasculature. Perivascular pO₂ measurements have been made previously in the R3230 tumor using recessed tip oxygen microelectrodes. The data presented here match very well with these microelectrode measurements, which were performed independently (36).

Although the median pO₂ for a single experiment generally changed only 1 to 2 mm Hg between any 2.5 min measurement interval, plots of randomly selected individual pixel pO₂ versus time reflect significant spatial and temporal variability (Fig. 1) in vascular pO₂. This spatiotemporal variability strongly suggests that sampling a few spatially or temporally discrete pO₂ measurements may not be representative of tumor oxygenation.

O₂ delivery is constantly fluctuating. Examination of the raw PLI images of a representative set of tumors showed different patterns of pO₂ fluctuations for each tumor line (Fig. 2; same representative set of tumors used as samples throughout manuscript; Supplementary Movies S1–S3). 9L images had different areas of the images with their pO₂ fluctuating independently of each other; in the sample shown in Fig. 2, pO₂ on the left side of the image decreases with time, whereas pO₂ on the right side increases. Both FSA and R3230 seemed to have a distinct spatial pattern of pO₂ at the initial time point and pO₂ fluctuations with time maintained these patterns. The FSA sample has an inverted triangular area at lower pO₂ values than the rest of the image. This area initially shrinks and then increases to almost encompass the entire image by the 60-min time point. The R3230 tumor in Fig. 2 has an oval-shaped outline visibly at a lower pO₂ at the first time point, and over time, this oval-shaped outline remains visibly at a lower pO₂.

These observations of differences in vascular pO₂ pattern between tumor types was reinforced through an exploratory analysis looking at the sign of pO₂ changes >1 mm Hg for each pixel from once point to the next (Supplementary Movies S4–S6). Not a single pixel remained unchanged over three sequential time points in all the experiments, and fewer than 5% of the pixels changed <1 mm Hg between any two subsequent time points. These patterns were also visible in the difference images when examining the sign of pO₂ changes >5 mm Hg for each pixel (Supplementary Movies S7–S9). Although there are 2.5-min intervals where only a small percentage of the image shows changes in vascular pO₂ of >5 mm Hg, many of the 2.5-min intervals show 30% to 50% of the image fluctuating at >5 mm Hg.

Tumor type effects spatial distribution of oxygen. Watershed segmentation (Fig. 3) showed distinctly different oxygen isobar patterns for different tumor types. At most time points, 9L was highly segmented, with regions of very different pO₂ values next to each other. R3230 also showed areas which were highly segmented; however, many of the images contained one or two large spatial pO₂ segments. FSA had a great deal of segmentation, but segments with similar pO₂ were almost always conjoined.

Global Moran's I analysis showed significant differences in patterns of spatial autocorrelation for the different tumor types (Fig. 4A). Generally, spatial autocorrelation decreased as the size of the region examined increased for all tumor types. Both 9L and FSA had several time points, which showed no autocorrelation in I values at the smallest spatial range examined (1 pixel radius; Supplementary Table S1) at this spatial range, suggesting that in these images adjoining pixels had pO₂ values which were independent of each other. Within tumor types, I values for different spatial ranges tended to be significantly different from each other (Supplementary Table S1). Combined with the knowledge that spatial autocorrelation decreased as the area examined increased, this analysis emphasizes the heterogeneity of O₂ delivery in tumors.

Global Moran's I was also used to examine patterns of change in pO₂ between time points (Fig. 4B). There was much less spatial autocorrelation in the change in pO₂ values between time points than within a single time point, although all tumor types had some difference images, which showed a high degree of spatial autocorrelation. Spatial autocorrelation further decreased in the difference images as the spatial range examined increased. However, differences in spatial autocorrelation were evident for different tumor types. 9L and R3230 showed significantly less spatial autocorrelation than FSA at spatial ranges higher than a 1-pixel radius (Supplementary Table S2), and I values for these two tumor types were more tightly clustered than FSA at all spatial ranges. FSA and R3230 tend to have significant decreases in Moran's I values within each tumor type for small increases in spatial ranges (1 and 4 pixel radii for FSA, 1 and 8 pixel radii for R3230), suggesting changes in pO₂ values over time were spatially coordinated only for small regions in these tumor types.

Global Moran's I for an image is an average of normalized local Moran's I values for each pixel within that image. Therefore, an examination of local Moran's I values for an image offers insight into the areas within an image which have the greatest spatial autocorrelation.

The sample images shown are local Moran's I values for a spatial range of an 8-pixel radius (Fig. 5). The 9L tumor had the same general pattern of spatial autocorrelation at all time points, with a large region of spatial autocorrelation on the lower right-hand side. The sample FSA showed an area of high spatial autocorrelation in the inverted triangular region, which increased and decreased in size with time. R3230 showed high spatial autocorrelation at all time points in the oval-shaped outline visible in Fig. 2; this area of high spatial autocorrelation seemed to increase in size with time. Consistent with the global Moran's I results for this spatial range, the sample 9L, FSA, and R3230 tumors had a similar proportion of the image showing high spatial coordination.

Figure 6, which shows local Moran's I values for difference images for the same sample tumors in Fig. 5, offers insight into the spatial autocorrelation of changes in pO₂ between time points. The 9L sample shows that there is almost no spatial coordination in changes between 2.5 min time points. The sample FSA tumor has noticeably more spatial autocorrelation in pO₂ changes, with regions of high spatial autocorrelation appearing first along the top and lower right side of the image, with a few later time points showing a large region of spatial autocorrelation along the left side. Most time points for the sample R3230 difference images showed a few, highly diffuse spots of high autocorrelation, with most of the image reflecting spatial randomness of pO₂ values. There was a recurring region with high spatial pO₂ autocorrelation in the top left corner of the R3230 sample. The trends in spatial autocorrelation seen in Fig. 6 are consistent with global Moran's I values for each tumor type.

For all of the sample tumors shown, the areas of spatial autocorrelation at a single time point are not the same areas which show spatial autocorrelation for pO₂ changes between time points. The distinct pattern of the inverted triangle visible in Fig. 5B for FSA is not apparent in Fig. 6B; similarly, the oval-shaped outline boldly apparent in Fig. 5C for R3230 is almost entirely missing in Fig. 6C.

All tumors fluctuate with slow periodicities. Fourier analysis showed the periodicities with the largest pO₂ fluctuation magnitudes were >10 min for all three tumor types. Generally, the three slowest periodicities (ranging from 10 to >40 minutes) for a pixel have a similar order of magnitude contribution to the overall pO₂ fluctuation magnitude; faster fluctuation periodicities have 10² to 10⁴ lower order of magnitude contribution to the overall pO₂ fluctuation magnitude.

The concept of hypoxia being a source for radioresistance goes back to the seminal work of Thomlinson and Gray (37). In this study, the investigators carefully measured the width of the viable rim of biopsies taken from patients with lung tumors and determined that the maximum width tended to be 180 to 200 μm, regardless of the size of the tumor nodule. They concluded that the viable rim width was consistent with the maximum diffusion distance of oxygen in tumor tissue. Thomlinson and Gray hypothesized that cells that reside near the necrotic edge would be hypoxic and therefore radioresistant. Although this observation spurred a number of clinical trials, it was nearly 30 years before it was definitively shown that hypoxia was a cause for radioresistance in human tumors (38).

Vaupel was among the first investigators to show that human tumors contained hypoxic regions and that the presence of hypoxia was an important prognostic factor in cervix cancer (39). Subsequently, similar reports were published involving a number of human tumors, including head and neck, sarcomas, and prostate cancer (40–42). The work of Thomlinson and Gray alluded to and was based upon the assumption of a steady-state gradient of oxygen from a source vessel, a concept that was further promulgated by the advent of hypoxia marker drugs, which typically exhibit increased binding in tumor cells as a function of their distance from the nearest microvessel. This feature of tumor hypoxia has been termed chronic, implying that cells that located far from a microvessel reside in a constant state of hypoxia.

The notion that tumor cells might be transiently hypoxic as a result of instabilities in tumor blood flow was introduced in the late 1970s in three seminal papers. In one study, Intaglietta and colleagues observed arteriolar vasomotion in feeding vessels of tumors grown in window chambers (43). They reported the sinusoidal behavior of RBC movement in vessels, presumably caused by this vasomotor activity Yamaura and Matsuzawa irradiated tumors growing in window chambers and carefully monitored the location of tumor regrowth (44). They observed that the tumors almost always regrew at the periphery of the tumors and suggested two potential mechanisms in explanation of this localized radioresistance. In one proposal, cells in the tumor periphery are assumed to be better oxygenated, resulting in a higher growth fraction and a higher fraction of relatively resistant cells in S-phase at the time of irradiation. Yamaura and Matsuzawa also proposed that transient hypoxia could be responsible for radioresistance, as vascular stasis had occasionally been observed at the tumor periphery. Brown showed the reemergence of radiobiologically hypoxic cells 24 hours after administration of a hypoxic cytotoxin, Misonidazole, in the EMT6 tumor (45). This study was the first to show that transient hypoxia could be radiobiologically important.

This paper proves our hypothesis that fluctuating vascular oxygenation is a prevalent characteristic of these three tumor lines. The main result of this study is one that has perhaps been intuitively understood but not previously explicitly shown: O₂ delivery to tumor tissue is constantly changing. Previous characterization of hypoxia as perfusion-limited or diffusion-limited are simply the extreme cases of O₂ delivery dominant or O₂ metabolism dominant areas in tumors; most tumor tissue does not distinctly fall into either category and the local pO₂ is heavily influenced by both.

This study shows that up to 50% of tumor vascular pO₂ can change >5 mm Hg within a 2.5-min interval; previous theoretical predictions have calculated that a change in vascular pO₂ of 10 mm Hg can cause a considerable increase (∼30%) in the proportion of severely hypoxic tumor tissue (<3 mm Hg; ref. 46). Although these predictions cannot be directly applied to the data in this study, changes in vascular pO₂ of >5 mm Hg, such as those seen in this study, would certainly be expected to alter the hypoxic fraction of the surrounding tissue. Experimental results have shown that fluctuations in vascular pO₂ result in fluctuations in tissue pO₂ to the maximum oxygen diffusion distance (∼150 μm; ref. 47).

Perhaps the universal characteristic of tumor pO₂ is that the vascular pO₂ fluctuations are occurring at low frequencies or periods on the order of 10s of minutes. This idea is not a new one; the seminal work of Chaplin and colleagues using perfusion markers showed that the injection of dyes 20 minutes apart resulted in marker mismatch in vessels (4). Later studies showed that the significant time scale was at least 15 minutes for perfusion marker mismatch (6), a time scale which has been reinforced with direct measurements of pO₂ fluctuations (5, 9, 11, 12). This study is the first to show that these direct, single-point studies of tumor pO₂ can be generalized to describe the dominant period of pO₂ fluctuations for the entire tumor. A recent publication has shown that pO₂ fluctuations are occurring at dominant periods of 10s of minutes in spontaneous canine tumors (12), suggesting the characteristic behavior of fluctuating pO₂ is not limited to murine tumors or xenografts grown in rodents and may have a common mechanism across all tumors. Studies measuring pO₂ fluctuations in normal tissue (muscle) in rats and mice have not observed significant fluctuations (5, 9).

Although no studies directly measuring fluctuating pO₂ have been done in human tumors, fluctuations in red cell flux have been measured clinically in humans and were shown to occur with a median periodicity of 13 min (range, 4–44 min; ref. 48). Measurements of fluctuating blood flow in different areas within these human tumors also revealed spatial heterogeneity in the fluctuations; spatial heterogeneity in temporal oxygen fluctuations has important implications for optimization of traditional therapies such as radiation. If areas of fluctuating hypoxia were able to be visualized, higher doses of radiation could be delivered to hypoxic areas at a time during which pO₂ values in that area were at the peak of their fluctuations. Further studies examining the spatiotemporal periodicity of tumor oxygenation over a longer timescale (days) relevant to treatment scheduling need to be conducted.

This paper has shown that fluctuations in vascular pO₂ are constantly occurring throughout tumors. Nearly all of the current understanding of tumors and tumor cell adaptation to hypoxia contains the underlying assumption of cellular exposure to chronic hypoxia. However, numerous publications have shown that intermittent hypoxia or hypoxia/reoxygenation events alter the behavior of tumor cells. Graeber and colleagues showed that multiple rounds of hypoxia/reoxygenation selected for apopotosis-resistant tumor cells, with each hypoxic treatment increasing the percentage of p53-deficient cells by 2.4% (20). Another study more broadly showed the potential of intermittent hypoxia to act as a physiologic selective agent for tumor cell mutations: Reynolds and colleagues showed the mutation frequency of tumor cells increased with each cycle of hypoxia exposure (23). Whole-body exposure of tumor-bearing mice to intermittent hypoxia has been shown to result in increased spontaneous metastases in lungs (19) and lymph nodes (18). Recently, a study has shown that fluctuating hypoxia promotes different phenotypes in endothelial and tumor cells than chronic hypoxia, further promoting cells to participate in tumor progression and treatment resistance (21, 49).

The literature also increasingly suggests that the molecular effects of hypoxia will have to be revisited. Studies looking at oxidative stress have shown that cycles of intermittent hypoxia increased HIF-1α protein expression and transactivation function in tumor cells through molecular mechanisms which are distinct from those of chronic hypoxia (25). Activation of HIF-1 was shown to increase with increasing number of hypoxia/reoxygentation cycles (21, 25). A related study has also shown that intermittent hypoxia also increases c-fos mRNA in tumor and endothelial cells and increased activation of activator protein 1, both of which were abolished through the use of an superoxide dismutase mimetic (24). These studies strongly suggest that intermittent hypoxia is more potent in activating gene expression than chronic hypoxia. Additionally, one study has shown that >200 genes are selectively affected by intermittent, but not chronic, hypoxia (22); this presents a multitude of potential therapeutic targets which are currently overlooked due to a misunderstanding of the prevalence and importance of fluctuating oxygenation in tumors. Given the probable widespread presence of fluctuating hypoxia in human tumors, many more studies need to be done under this new paradigm of tumor physiology.

Spatiotemporal analysis reveals some clues regarding the underlying mechanism(s) of fluctuating hypoxia. Fluctuations in red cell flux, possibly due to vasomotion, are known to cause fluctuations in tumor tissue pO₂ (17, 46, 47). This proposition of fluctuations in red cell flux within tumor vessels is consistent with the analysis shown in Fig. 5: while pO₂ for individual pixels show temporal fluctuations throughout the image, the continuous presence over time of the inverted triangle and oval-shaped outline visible in two of the tumors suggest a network of vessels through which red cells are moving. This result is not unexpected; in prior work with microelectrodes, we found significant fluctuations in tissue pO₂ of regions of 200 to 300 μm in diameter in which small networks of five to six microvessels were involved (47). However, in Fig. 6, autocorrelation of changes in pO₂ between time points is generally not occurring in the locations of the networks in Fig. 5. One potential mechanism for these localized areas of change in pO₂ may be vascular intussusception, which has been shown to locally alter blood flow in experimental tumors with a similar timescale to the one measured in our study (50).

No potential conflicts of interest were disclosed.

Grant support: NIH/National Cancer Institute grant CA40355 and Department of Defense grant BC043284.

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.