Abstract
The purpose of this investigation was to identify pathways by which physical activity, implemented as running in an activity wheel, inhibits carcinogenesis. The focus of this analysis was on 20 plasma biomarkers for glucose homeostasis, inflammation, cytokine function, and endocrine activity, known to be affected by a physically active lifestyle. Plasma for analysis was obtained from previously reported carcinogenesis experiments in which the effects on mammary carcinogenesis, induced by i.p. injection of 1-methyl-1-nitrosurea, of running on a motorized activity wheel or a nonmotorized free wheel were compared with sedentary controls. Wheel running reduced cancer incidence (P = 0.0004) and the number of cancers per animal (P = 0.005). Principal components analysis was used to reduce the 20 plasma biomarkers to a concise index that was significantly different by treatment group assignment (P < 0.0001). Statistical analyses provided evidence that supported the hypothesis of a mediational role of these molecules in accounting for the protective effect of wheel running on the carcinogenic process. In addition, the plasma biomarker index derived from principal components analysis was a good discriminator of treatment group assignment (only 4.5% of animals were misclassified). These findings suggest that the plasma biomarkers evaluated have utility in assessing the breast cancer response to a physical activity intervention. Identification of such biomarkers is critical in clinical studies for which evaluating the effects of physical activity on cancer outcomes (diagnosis, recurrence, or mortality) is not possible. Cancer Prev Res; 3(11); 1484–92. ©2010 AACR.
Introduction
Epidemiologic investigations provide strong evidence that individuals who are in the highest quartile of physical activity have a lower incidence of breast cancer and lower mortality rates than individuals who are in the lowest quartile (1–5), and that greater protection is associated with moderate or higher intensity of physical activity (6). Moreover, the magnitude of protection has been reported to be greater in individuals in the lowest quartile of body mass index and the highest quartile of physical activity, suggesting that physical activity may have effects on breast cancer risk independent of body size. Although the effects of physical activity on circulating levels of factors involved in glucose homeostasis, inflammation, cytokine function, and endocrine activity, and the relationship of these same factors to the risk for developing cancer, have been studied extensively (6), no study to date has examined whether changes in systemic concentrations of these factors modulate carcinogenesis in the same model system. The work reported herein addressed this issue by using data from animal experiments in which a model for breast cancer was combined with a wheel-running model for physical activity (7, 8). The animals in these experiments had similar body weights in the physical activity and sedentary control groups and levels of body fat <20%, which are in the reference range for the rat. The purpose of this work was to explore the role of 20 plasma biomarkers as a mediating step between the physical activity intervention and two cancer end points: incidence and number of cancers per animal. The biological basis for the selection of the set of biomarkers studied was based on evidence, which has been reviewed in detail (9, 10), that their specific misregulation occurs during carcinogenesis and that amelioration of deregulated function would inhibit carcinogenesis.
Materials and Methods
Experimental design
The treatment of animals has been reported in detail in refs. (7, 8). Briefly, female Sprague-Dawley rats were obtained at 20 days of age and housed in solid-bottomed polycarbonate cages. At 21 days of age, rats were injected (i.p.) with 50 mg 1-methyl-1-nitrosurea per kilogram of body weight, as described previously (11). At 28 days of age, 1 week after carcinogen injection, rats were sorted by weight and randomized to one of three groups: (a) motorized activity wheel (n = 22), (b) nonmotorized free wheel (n = 33), or (c) sedentary control (n = 55). The motorized wheel rats ran at a constant speed (40 m/min), whereas the rats assigned to the free wheel self-determined the speed at which they ran. Rats were fed a purified pelleted diet (Research Diet). Food pellets were distributed based on distance ran as a positive reinforcement of running behavior. A computer device attached to the activity wheel monitored distance run, which was recorded daily. At necropsy, rats were skinned and the skin to which mammary gland chains were attached was examined under translucent light for detectable mammary pathologies. All grossly detectable mammary gland pathologies were excised and prepared for histologic classification. Only confirmed mammary carcinomas are reported. The experimental protocols were reviewed and approved by the Institutional Animal Care and Use Committee and conducted according to the committee guidelines.
Blood collection and plasma biomarker analyses
Blood collection.
Following an overnight fast, rats were euthanized over a 3-hour time interval, between 8 and 11 a.m., through inhalation of gaseous carbon dioxide. The sequence in which rats were euthanized was stratified across groups to minimize the likelihood that order effects would masquerade as treatment-associated effects. After the rat lost consciousness, blood was directly obtained from the retro-orbital sinus and gravity fed through heparinized capillary tubes (Fisher Scientific) into EDTA-coated tubes (Becton Dickinson) for plasma. The bleeding procedure took ∼1 min/rat. Thereafter, the unconscious rat was euthanized by cervical dislocation. Plasma was isolated by centrifugation at 1,000 × g for 10 minutes at room temperature.
Enzyme-linked immunosorbent assays.
The enzyme-linked immunosorbent assays (ELISA) utilized a quantitative “sandwich” technique. In each ELISA kit (Diagnostic Systems Laboratory, Cayman Chemicals, and Millipore), a monoclonal antibody specific for insulin-like growth factor-I (IGF-I), corticosterone, insulin, leptin, progesterone, estradiol, C-reactive protein, or growth hormone was precoated on the 96-well microtiter plate. All reagents were brought to room temperature, and assay diluent (100 μL) was added to each well. The standards or unknown sample (100 μL) were then added to the wells in duplicate. After a 2- to 3-hour incubation period, the wells were washed three to four times with the provided wash buffer followed by addition of a second specific streptavidin-horseradish peroxidase–conjugated antibody. The plate was incubated at room temperature for 1 to 2 hours followed by washing and stop solution. Color development and color intensity were measured using an ELISA plate reader (Molecular Devices). A standard curve was prepared by plotting the absorbance versus the concentration of the targets. Values for experimental samples were determined through a four-parameter logistic regression model.
BioPlex assays.
Cytokine measures, including interleukin (IL)-1α, IL-1β, IL-2, IL-4, IL-6, IL-10, granulocyte macrophage colony-stimulating factor, interferon-γ (IFN-γ), and tumor necrosis factor-α were made using a protein multiplex immunoassay system (Rat 9-Bio-Plex Cytokine Array System, Bio-Rad Laboratories), and thyroid hormone 3 and 4 were determined using rat thyroid hormone panel LINCOplex kit (Millipore) following the individual manual provided by the vendors. Briefly, beads conjugated to specific primary capture antibodies were added to wells of a 96-well multiscreen plate, and fluid was uniformly removed from each well using a multiscreen resist vacuum manifold at constant pressure after incubation. Wells were washed with wash buffer, using vacuum removal after each wash. Standards, positive controls, and samples were then added to the respective wells and incubated with beads at room temperature with orbital shaking. After incubation, fluid was removed again as defined above. Wells were washed with wash buffer. Secondary antibody was then added to all wells followed by incubation at room temperature with shaking. After fluid removal and washing, streptavidin-phycoerythrin complex was added to each well and incubated with shaking. Then, the plate was washed with wash buffer. Finally, assay buffer was added to resuspend beads. At all shaking points noted above, 96-well assay plates were covered with sealing tape and aluminum foil to retard photobleaching. In addition, after each wash step, the underside of the plate was blotted with a paper towel. After the above procedures, standards and samples were immediately processed using the Bio-Plex Protein Array System and related Bio-Plex Manager software (Bio-Rad). Standard curves were prepared by plotting absorbance versus the concentration of the targets. Values for experimental samples were determined using a five-parameter logistic regression model.
Statistical methods
Body weight at the time of study termination was compared across groups by using ANOVA. Cancer outcomes were evaluated by Fisher's exact test (incidence), Poisson regression (cancer count), and least squares regression (square root of cancer count). Initial evaluation of biomarkers included testing individually for differences across treatment group (one-way ANOVA), testing for differences by incidence (t test), and testing for association with cancer count (Poisson regression). Because there were 20 biomarkers of interest, we performed a principal components (PC) analysis (PCA) for data reduction to construct a single variable (12). The first PC is the linear combination of the 20 variables that passes through the centroid of the full dataset while minimizing the square of the perpendicular distance of each point to that line; each subsequent PC was constructed in a similar manner, subject to the constraint of being mutually orthogonal (13). The PCA was explored graphically in a multidimensional preference plot, and linear discriminant analysis was used to evaluate the first two PCs as predictors of treatment group (13). The results from analyses in which all 20 biomarkers were evaluated were adjusted to control the expected proportion of errors among the rejected hypotheses; that is, the false discovery rate (14).
The biomarker data were tested for mediation by using the model described in Fig. 1. The mediational model addresses the question of whether the treatment has an indirect effect on the outcome by inducing a change in one or more intervening variables that in turn induce a change in the outcome (15). In this case, the intervening variables (candidates for mediation) were a collection of biomarkers thought to be related to cancer risk. Each marker was also tested individually for potential meditation. Analyses were done in SAS 9.2 (SAS Institute, Inc.).
The mediational model, where τ is the direct effect of exercise on the cancer outcome and τ − τ′ or αβ measure the mediational effect of the variable(s) being tested. We focused on αβ (15) and used Eqs. (B) and (C) to estimate α and β, and finally to test the null hypothesis that the mediational effect, αβ, equals 0. CI values for the mediational effect were estimated by the percentile method in the bootstrapped (resampling with replacement) distribution of the estimate of αβ (22).
The mediational model, where τ is the direct effect of exercise on the cancer outcome and τ − τ′ or αβ measure the mediational effect of the variable(s) being tested. We focused on αβ (15) and used Eqs. (B) and (C) to estimate α and β, and finally to test the null hypothesis that the mediational effect, αβ, equals 0. CI values for the mediational effect were estimated by the percentile method in the bootstrapped (resampling with replacement) distribution of the estimate of αβ (22).
Results
The carcinogenic response observed in the 110 animals (55 sedentary control, 22 motorized wheel, and 33 free wheel) from which plasma was collected was similar to that reported in the parent studies and is shown in Table 1. Both cancer incidence and cancer multiplicity were reduced by physical activity compared with the sedentary control whether rats ran on nonmotorized free wheels or motorized activity wheels. Although the motorized wheel had a greater effect in reducing the carcinogenic response than the free wheel, the difference between the carcinogenic response in the free versus the motorized wheel was not statistically significant. The relative risk for incident cancer was 0.68 overall for physical activity (motorized and free wheel groups combined), 0.68 for the motorized wheel, and 0.69 for running in the free wheel relative to the sedentary control. Also shown in Table 1 are the effects of treatment on final body weights. As expected, sedentary animals were heavier than wheel-running animals; however, average body weights varied <5% among treatment groups. Although these differences were statistically significant, they are not considered clinically relevant (16). The Lee index, a reflection of body fat, was not different among treatment groups (data not shown). Based on dual X-ray absorptiometry analysis, sedentary control animals of similar body weight have 17.0 ± 3% body fat (mean ± SD, n = 10).
Effect of physical activity types on final body weight and carcinogenic response in the mammary gland
Treatment . | Activity type . | n . | Final body weight (g)* . | Cancer incidence, n (%) . | Ave. no. of cancers/rat* . |
---|---|---|---|---|---|
Physically active | Motorized wheel | 22 | 198.1 ± 3.4 | 15 (68) | 1.55 ± 0.29 |
Free wheel | 33 | 183.9 ± 1.9 | 26 (79) | 2.30 ± 0.35 | |
Sedentary control | None | 55 | 202.6 ± 2.3 | 54 (98) | 3.73 ± 0.32 |
Physically active compared with sedentary control | P† | <0.0001 | 0.0004 | 0.005 | |
Motorized wheel compared with sedentary control | P† | 0.25 | 0.0005 | 0.004 | |
Free wheel compared with sedentary control | P† | <0.0001 | 0.004 | 0.0015 |
Treatment . | Activity type . | n . | Final body weight (g)* . | Cancer incidence, n (%) . | Ave. no. of cancers/rat* . |
---|---|---|---|---|---|
Physically active | Motorized wheel | 22 | 198.1 ± 3.4 | 15 (68) | 1.55 ± 0.29 |
Free wheel | 33 | 183.9 ± 1.9 | 26 (79) | 2.30 ± 0.35 | |
Sedentary control | None | 55 | 202.6 ± 2.3 | 54 (98) | 3.73 ± 0.32 |
Physically active compared with sedentary control | P† | <0.0001 | 0.0004 | 0.005 | |
Motorized wheel compared with sedentary control | P† | 0.25 | 0.0005 | 0.004 | |
Free wheel compared with sedentary control | P† | <0.0001 | 0.004 | 0.0015 |
*Values are means ± SEM.
†P values are based on contrasts in a one-way ANOVA (body weight), Fisher's exact test (cancer incidence), and contrasts in a Poisson regression (cancers per rat).
Twenty plasma biomarkers were evaluated in each animal. These biomarkers can be loosely grouped into three families, although most of these molecules have multiple biological activities. The families were designated (a) glucose homeostasis and metabolism, (b) inflammation and cytokines, and (c) sex steroid hormones, as shown in Table 2. Our initial set of analyses examined each marker for difference by treatment group (Table 2), association with cancer incidence (Table 3), and association with cancer count per rat (Table 4). When considered individually, all plasma biomarkers except glucose, C-reactive protein, and estradiol were associated with treatment group. Corticosterone, leptin, IL-2, IL-4, IL-6, IL-10, and IFN-γ were significantly associated with cancer incidence. Glucose, insulin, growth hormone, IGF-I, IL-1α, and progesterone were associated with cancer count, in addition to the same biomarkers associated with cancer incidence (i.e., corticosterone, leptin, IL-2, IL-4, IL-6, IL-10, and IFN-γ). However, as reflected in the slope estimates shown in Table 4, the strength of these relationships was very modest.
Plasma biomarkers by exercise treatment
Variable . | Sedentary control* (n = 55) . | Free wheel* (n = 33) . | Motorized wheel* (n = 22) . | P† . |
---|---|---|---|---|
Glucose | 123.91 ± 5.83 | 131.39 ± 9.11 | 111.23 ± 10.14 | 0.3314 |
Insulin | 1.27 ± 0.08 | 0.99 ± 0.08 | 0.93 ± 0.05 | 0.0117 |
IGF-I | 1,305.20 ± 44.36 | 1,041.00 ± 70.86 | 1,068.73 ± 66.67 | 0.0016 |
T3 | 5.62 ± 0.49 | 3.55 ± 0.39 | 10.81 ± 1.08 | 0.0002 |
T4 | 224.47 ± 10.67 | 185.61 ± 6.64 | 267.00 ± 18.75 | 0.0005 |
GH | 20.23 ± 0.72 | 16.20 ± 0.87 | 17.23 ± 1.38 | 0.0036 |
Cortico | 261.58 ± 6.13 | 545.94 ± 22.35 | 562.50 ± 37.76 | 0.0002 |
CRP | 436.69 ± 11.61 | 418.39 ± 13.97 | 415.45 ± 21.48 | 0.5048 |
IL-1α | 73.36 ± 1.83 | 59.73 ± 1.97 | 56.50 ± 2.49 | 0.0002 |
IL-1β | 137.40 ± 4.01 | 112.55 ± 4.47 | 115.18 ± 4.13 | 0.0002 |
IL-2 | 310.78 ± 8.59 | 413.64 ± 16.79 | 404.45 ± 16.02 | 0.0002 |
IL-4 | 94.80 ± 2.78 | 120.73 ± 4.14 | 123.45 ± 6.82 | 0.0002 |
IL-6 | 4,550.69 ± 132.69 | 5,701.52 ± 208.37 | 5,823.91 ± 261.33 | 0.0002 |
IL-10 | 817.96 ± 26.17 | 1,113.94 ± 54.12 | 1,102.27 ± 36.94 | 0.0002 |
GM-CSF | 22.00 ± 0.64 | 31.61 ± 1.66 | 25.05 ± 1.37 | 0.0002 |
IFN-γ | 821.47 ± 21.75 | 1,098.67 ± 41.33 | 1,058.95 ± 37.05 | 0.0002 |
TNF-α | 433.75 ± 13.28 | 365.42 ± 14.59 | 368.18 ± 11.61 | 0.0007 |
Leptin | 2.11 ± 0.13 | 0.70 ± 0.09 | 1.10 ± 0.13 | 0.0002 |
Progesterone | 16.87 ± 0.74 | 27.42 ± 1.79 | 32.41 ± 2.44 | 0.0002 |
Estradiol | 16.69 ± 1.14 | 14.33 ± 1.32 | 15.52 ± 1.47 | 0.4155 |
PC1 | −1.77 ± 0.16 | 1.83 ± 0.32 | 1.69 ± 0.29 | 0.0002 |
Variable . | Sedentary control* (n = 55) . | Free wheel* (n = 33) . | Motorized wheel* (n = 22) . | P† . |
---|---|---|---|---|
Glucose | 123.91 ± 5.83 | 131.39 ± 9.11 | 111.23 ± 10.14 | 0.3314 |
Insulin | 1.27 ± 0.08 | 0.99 ± 0.08 | 0.93 ± 0.05 | 0.0117 |
IGF-I | 1,305.20 ± 44.36 | 1,041.00 ± 70.86 | 1,068.73 ± 66.67 | 0.0016 |
T3 | 5.62 ± 0.49 | 3.55 ± 0.39 | 10.81 ± 1.08 | 0.0002 |
T4 | 224.47 ± 10.67 | 185.61 ± 6.64 | 267.00 ± 18.75 | 0.0005 |
GH | 20.23 ± 0.72 | 16.20 ± 0.87 | 17.23 ± 1.38 | 0.0036 |
Cortico | 261.58 ± 6.13 | 545.94 ± 22.35 | 562.50 ± 37.76 | 0.0002 |
CRP | 436.69 ± 11.61 | 418.39 ± 13.97 | 415.45 ± 21.48 | 0.5048 |
IL-1α | 73.36 ± 1.83 | 59.73 ± 1.97 | 56.50 ± 2.49 | 0.0002 |
IL-1β | 137.40 ± 4.01 | 112.55 ± 4.47 | 115.18 ± 4.13 | 0.0002 |
IL-2 | 310.78 ± 8.59 | 413.64 ± 16.79 | 404.45 ± 16.02 | 0.0002 |
IL-4 | 94.80 ± 2.78 | 120.73 ± 4.14 | 123.45 ± 6.82 | 0.0002 |
IL-6 | 4,550.69 ± 132.69 | 5,701.52 ± 208.37 | 5,823.91 ± 261.33 | 0.0002 |
IL-10 | 817.96 ± 26.17 | 1,113.94 ± 54.12 | 1,102.27 ± 36.94 | 0.0002 |
GM-CSF | 22.00 ± 0.64 | 31.61 ± 1.66 | 25.05 ± 1.37 | 0.0002 |
IFN-γ | 821.47 ± 21.75 | 1,098.67 ± 41.33 | 1,058.95 ± 37.05 | 0.0002 |
TNF-α | 433.75 ± 13.28 | 365.42 ± 14.59 | 368.18 ± 11.61 | 0.0007 |
Leptin | 2.11 ± 0.13 | 0.70 ± 0.09 | 1.10 ± 0.13 | 0.0002 |
Progesterone | 16.87 ± 0.74 | 27.42 ± 1.79 | 32.41 ± 2.44 | 0.0002 |
Estradiol | 16.69 ± 1.14 | 14.33 ± 1.32 | 15.52 ± 1.47 | 0.4155 |
PC1 | −1.77 ± 0.16 | 1.83 ± 0.32 | 1.69 ± 0.29 | 0.0002 |
NOTE: Families of biomarkers: (a) glucose homeostasis and metabolism: Cortico, GH, glucose, IGF-I, insulin, T3, T4; (b) inflammation and cytokines: CRP, GM-CSF, IFN-γ, IL-2, IL-4, IL-6, IL-10, IL-1α, IL-1β, leptin, TNF-α; and (c) sex steroid hormones: estradiol, progesterone.
Abbreviations: Cortico, corticosterone; CRP, C-reactive protein; GH, growth hormone; GM-CSF, granulocyte macrophage colony-stimulating factor; IFN-γ, interferon-γ; IGF-I, insulin like growth factor-I; IL, interleukin; T3, triiodothyronine; T4, thyroxine; TNF-α, tumor necrosis factor-α; PC1, principal component 1.
*Values are means ± SEM.
†P values for the global F test of any difference between treatments in a one-way ANOVA, controlling the false discovery rate at 5%.
Plasma biomarkers by cancer incidence
Variable . | Cancer-free* rats (n = 15) . | Cancer-bearing* rats (n = 95) . | Difference (95% CI) . | P† . |
---|---|---|---|---|
Glucose | 116.73 ± 9.63 | 124.71 ± 4.98 | −7.97 (−33.97 to 18.03) | 0.5805 |
Insulin | 0.99 ± 0.14 | 1.14 ± 0.05 | −0.15 (−0.44 to 0.14) | 0.4669 |
IGF-I | 968.13 ± 91.90 | 1,211.88 ± 37.31 | −243.75 (−443.47 to −44.03) | 0.0555 |
T3 | 7.26 ± 1.35 | 5.84 ± 0.44 | 1.42 (−1.02 to 3.86) | 0.4669 |
T4 | 222.20 ± 20.32 | 221.18 ± 7.84 | 1.02 (−41.22 to 43.27) | 0.9631 |
GH | 17.15 ± 1.47 | 18.63 ± 0.59 | −1.48 (−4.65 to 1.69) | 0.4747 |
Cortico | 573.53 ± 47.53 | 380.79 ± 17.31 | 192.74 (98.71 to 286.78) | 0.0055 |
CRP | 396.33 ± 26.61 | 431.79 ± 8.63 | −35.46 (−83.27 to 12.35) | 0.3885 |
IL-1α | 60.93 ± 2.64 | 66.68 ± 1.55 | −5.75 (−13.75 to 2.25) | 0.1506 |
IL-1β | 121.67 ± 5.72 | 126.11 ± 3.09 | −4.44 (−20.51 to 11.64) | 0.5851 |
IL-2 | 422.33 ± 13.34 | 350.59 ± 9.48 | 71.74 (23.17 to 120.32) | 0.0007 |
IL-4 | 123.53 ± 5.22 | 105.91 ± 2.86 | 17.63 (2.76 to 32.49) | 0.0242 |
IL-6 | 5,899.13 ± 139.91 | 5,032.39 ± 132.01 | 866.74 (196.97 to 1,536.52) | 0.0007 |
IL-10 | 1,128.60 ± 60.66 | 937.57 ± 27.69 | 191.03 (44.74 to 337.32) | 0.0285 |
GM-CSF | 28.07 ± 2.48 | 25.08 ± 0.78 | 2.98 (−1.38 to 7.34) | 0.4308 |
IFN-γ | 1,122.00 ± 48.00 | 925.31 ± 23.08 | 196.69 (75.39 to 318.00) | 0.0055 |
TNF-α | 388.47 ± 21.65 | 401.98 ± 9.66 | −13.51 (−64.68 to 37.66) | 0.6356 |
Leptin | 0.97 ± 0.20 | 1.57 ± 0.11 | −0.60 (−1.16 to −0.04) | 0.0341 |
Progesterone | 27.40 ± 2.54 | 22.47 ± 1.10 | 4.93 (−0.90 to 10.75) | 0.1735 |
Estradiol | 16.37 ± 2.44 | 15.65 ± 0.79 | 0.71 (−3.65 to 5.07) | 0.8232 |
PC1 | 1.94 ± 0.38 | −0.31 ± 0.23 | 2.24 (1.05 to 3.44) | 0.0007 |
Variable . | Cancer-free* rats (n = 15) . | Cancer-bearing* rats (n = 95) . | Difference (95% CI) . | P† . |
---|---|---|---|---|
Glucose | 116.73 ± 9.63 | 124.71 ± 4.98 | −7.97 (−33.97 to 18.03) | 0.5805 |
Insulin | 0.99 ± 0.14 | 1.14 ± 0.05 | −0.15 (−0.44 to 0.14) | 0.4669 |
IGF-I | 968.13 ± 91.90 | 1,211.88 ± 37.31 | −243.75 (−443.47 to −44.03) | 0.0555 |
T3 | 7.26 ± 1.35 | 5.84 ± 0.44 | 1.42 (−1.02 to 3.86) | 0.4669 |
T4 | 222.20 ± 20.32 | 221.18 ± 7.84 | 1.02 (−41.22 to 43.27) | 0.9631 |
GH | 17.15 ± 1.47 | 18.63 ± 0.59 | −1.48 (−4.65 to 1.69) | 0.4747 |
Cortico | 573.53 ± 47.53 | 380.79 ± 17.31 | 192.74 (98.71 to 286.78) | 0.0055 |
CRP | 396.33 ± 26.61 | 431.79 ± 8.63 | −35.46 (−83.27 to 12.35) | 0.3885 |
IL-1α | 60.93 ± 2.64 | 66.68 ± 1.55 | −5.75 (−13.75 to 2.25) | 0.1506 |
IL-1β | 121.67 ± 5.72 | 126.11 ± 3.09 | −4.44 (−20.51 to 11.64) | 0.5851 |
IL-2 | 422.33 ± 13.34 | 350.59 ± 9.48 | 71.74 (23.17 to 120.32) | 0.0007 |
IL-4 | 123.53 ± 5.22 | 105.91 ± 2.86 | 17.63 (2.76 to 32.49) | 0.0242 |
IL-6 | 5,899.13 ± 139.91 | 5,032.39 ± 132.01 | 866.74 (196.97 to 1,536.52) | 0.0007 |
IL-10 | 1,128.60 ± 60.66 | 937.57 ± 27.69 | 191.03 (44.74 to 337.32) | 0.0285 |
GM-CSF | 28.07 ± 2.48 | 25.08 ± 0.78 | 2.98 (−1.38 to 7.34) | 0.4308 |
IFN-γ | 1,122.00 ± 48.00 | 925.31 ± 23.08 | 196.69 (75.39 to 318.00) | 0.0055 |
TNF-α | 388.47 ± 21.65 | 401.98 ± 9.66 | −13.51 (−64.68 to 37.66) | 0.6356 |
Leptin | 0.97 ± 0.20 | 1.57 ± 0.11 | −0.60 (−1.16 to −0.04) | 0.0341 |
Progesterone | 27.40 ± 2.54 | 22.47 ± 1.10 | 4.93 (−0.90 to 10.75) | 0.1735 |
Estradiol | 16.37 ± 2.44 | 15.65 ± 0.79 | 0.71 (−3.65 to 5.07) | 0.8232 |
PC1 | 1.94 ± 0.38 | −0.31 ± 0.23 | 2.24 (1.05 to 3.44) | 0.0007 |
*Values are means ± SEM. Abbreviations: see Table 2.
†P values for two-group t test, controlling the false discovery rate at 5%.
Plasma biomarkers by cancer count
Variable . | Slope ± SEM* . | P† . |
---|---|---|
Glucose | 0.0031 ± 0.0012 | 0.0141 |
Insulin | 0.3051 ± 0.0951 | 0.0027 |
IGF-I | 0.0008 ± 0.0002 | 0.0003 |
T3 | −0.0228 ± 0.0136 | 0.1157 |
T4 | −0.0009 ± 0.0008 | 0.3012 |
GH | 0.0225 ± 0.0094 | 0.0296 |
Cortico | −0.0019 ± 0.0004 | 0.0003 |
CRP | −0.0005 ± 0.0006 | 0.4949 |
IL-1α | 0.0084 ± 0.0038 | 0.0435 |
IL-1β | 0.0002 ± 0.0019 | 0.9310 |
IL-2 | −0.0027 ± 0.0006 | 0.0003 |
IL-4 | −0.0047 ± 0.0021 | 0.0435 |
IL-6 | −0.0002 ± 0.0000 | 0.0003 |
IL-10 | −0.0007 ± 0.0002 | 0.0019 |
GM-CSF | −0.0153 ± 0.0076 | 0.0631 |
IFN-γ | −0.0013 ± 0.0003 | 0.0003 |
TNF-α | 0.0004 ± 0.0006 | 0.4949 |
Leptin | 0.2037 ± 0.0510 | 0.0003 |
Progesterone | −0.0242 ± 0.0060 | 0.0003 |
Estradiol | 0.0121 ± 0.0070 | 0.1105 |
PC1 | −0.1497 ± 0.0262 | 0.0003 |
Variable . | Slope ± SEM* . | P† . |
---|---|---|
Glucose | 0.0031 ± 0.0012 | 0.0141 |
Insulin | 0.3051 ± 0.0951 | 0.0027 |
IGF-I | 0.0008 ± 0.0002 | 0.0003 |
T3 | −0.0228 ± 0.0136 | 0.1157 |
T4 | −0.0009 ± 0.0008 | 0.3012 |
GH | 0.0225 ± 0.0094 | 0.0296 |
Cortico | −0.0019 ± 0.0004 | 0.0003 |
CRP | −0.0005 ± 0.0006 | 0.4949 |
IL-1α | 0.0084 ± 0.0038 | 0.0435 |
IL-1β | 0.0002 ± 0.0019 | 0.9310 |
IL-2 | −0.0027 ± 0.0006 | 0.0003 |
IL-4 | −0.0047 ± 0.0021 | 0.0435 |
IL-6 | −0.0002 ± 0.0000 | 0.0003 |
IL-10 | −0.0007 ± 0.0002 | 0.0019 |
GM-CSF | −0.0153 ± 0.0076 | 0.0631 |
IFN-γ | −0.0013 ± 0.0003 | 0.0003 |
TNF-α | 0.0004 ± 0.0006 | 0.4949 |
Leptin | 0.2037 ± 0.0510 | 0.0003 |
Progesterone | −0.0242 ± 0.0060 | 0.0003 |
Estradiol | 0.0121 ± 0.0070 | 0.1105 |
PC1 | −0.1497 ± 0.0262 | 0.0003 |
*Values are estimates ± SEM. Abbreviations: see Table 2.
†P value based on a χ2 test for slope = 0 in a Poisson regression of cancer count on individual markers, controlling the false discovery rate at 5%.
Although grouping biomarkers by their currently understood functions has inherent value, there remains much to be learned about these functions, and as noted above, many biomarkers are associated with multiple activities. For this reason, the biomarker data were subjected to PCA. This unsupervised approach to data mining had two particular benefits: The generation of PC vectors reduced the dimensionality of the data for analysis, as discussed in Materials and Methods, and the generation of multidimensional scaling plots, which provide a graphical representation of scaled loading scores, permitted visual inspection of the clustering of biomarkers (Fig. 2). Figure 2A shows how an animal's assignment to being a sedentary control versus a wheel runner corresponded to the relationships among biomarkers determined by PCA. Figure 2B shows how the presence or absence of cancer corresponded to the relationships among biomarkers determined by PCA, and Fig. 2C displays the correspondence of each animal's cancer count with the relationships among biomarkers.
Multidimensional preference plots of first two PCs of 20 biomarkers. A, their relationship to treatment; B, their relationship to incidence; C, their relationship to cancer count. The vectors are useful in interpreting the relationships among the biomarkers as well as their relationship to the individual PC scores. Each animal is represented by a small open circle labeled for treatment (A: SC, MW, FW), cancer incidence (B: Y, N), or cancer count (C: n = 0, 1, 2, 3, 4, and 5 = 5+). Longer vectors show variables that better fit the two-dimensional model. A larger component of them is in the plane of the plot. In contrast, shorter vectors show variables that do not fit the two-dimensional model as well. They tend to be located farther from the plot in the direction of a third axis perpendicular to the page; hence, their projection into the plot is shorter. The results show in A, for example, that IL-6, IL-2, and IFN-γ are positively correlated with both PCs and are expressed more in exercising animals than sedentary controls. Leptin, IGF-I, and insulin, on the other hand, are negatively correlated with PC1; positively correlated with PC2; and expressed more in sedentary controls than in exercisers. In B, cancer-free animals are clustered more or less to the right of 0 along PC1, and express IL-2, IL-6, IL-10, IFN-γ, corticosterone, and progesterone. The animals labeled N in B are also labeled N in C, whereas those with any cancer appear in C with the count.
Multidimensional preference plots of first two PCs of 20 biomarkers. A, their relationship to treatment; B, their relationship to incidence; C, their relationship to cancer count. The vectors are useful in interpreting the relationships among the biomarkers as well as their relationship to the individual PC scores. Each animal is represented by a small open circle labeled for treatment (A: SC, MW, FW), cancer incidence (B: Y, N), or cancer count (C: n = 0, 1, 2, 3, 4, and 5 = 5+). Longer vectors show variables that better fit the two-dimensional model. A larger component of them is in the plane of the plot. In contrast, shorter vectors show variables that do not fit the two-dimensional model as well. They tend to be located farther from the plot in the direction of a third axis perpendicular to the page; hence, their projection into the plot is shorter. The results show in A, for example, that IL-6, IL-2, and IFN-γ are positively correlated with both PCs and are expressed more in exercising animals than sedentary controls. Leptin, IGF-I, and insulin, on the other hand, are negatively correlated with PC1; positively correlated with PC2; and expressed more in sedentary controls than in exercisers. In B, cancer-free animals are clustered more or less to the right of 0 along PC1, and express IL-2, IL-6, IL-10, IFN-γ, corticosterone, and progesterone. The animals labeled N in B are also labeled N in C, whereas those with any cancer appear in C with the count.
The relationships among the biomarkers, elucidated by PCA, are interesting in their own right. PC1 explained 26% of the variability in the biomarkers, whereas PC2 explained another 16%. The vectors in Fig. 2 were useful in interpreting the relationships among the biomarkers as well as their relationship to the individual animal's PC scores. Each animal is represented by a small open circle labeled for treatment (A: sedentary control, motorized wheel, free wheel), cancer incidence (B: Y, N), and cancer count (C: n = 0, 1, 2, 3, 4, and 5 = 5+ cancers per rat). The relative lengths of the vectors (arrows) in these figures measure a third dimension, perpendicular to the page, that captures the degree of association of each variable with the PCs on the axes. The correlation was weaker the shorter the vector. Thus, for example, in Fig. 2A, IL-6, IL-2, and IFN-γ were positively correlated with both PCs and expressed more in wheel-running animals than sedentary controls. Leptin and insulin, on the other hand, were negatively correlated with PC1, positively correlated with PC2, and expressed more in sedentary controls than in exercisers. Figure 2A shows nearly complete separation of treatment groups along the first PC. Only two sedentary control animals were placed among the physical activity group, whereas three physical activity animals were placed among the sedentary control animals; the misclassification rate was 4.5% using a linear discriminant function. Of particular interest was that the unsupervised clustering of biomarkers was similar to that assigned at the outset of analyses and that the clusters were in general, biologically plausible. Figure 2B relabeled the animals as cancer bearing (yes) and cancer free (no). Figure 2C relabeled the animals by cancer count (n = 0, 1, 2, 3, 4, and 5 = 5+). There was poorer discrimination between the presence or absence of cancer and cancer number, although the majority of the cancer-free animals clustered to the right of the zero point on the first PC and below the zero point on the second PC, a pattern that was most easily detected in Fig. 2B. The cancer-free rats tend to have higher plasma levels of corticosterone and progesterone.
The data presented in Fig. 2 supported a further exploratory analysis, the test for a mediational effect of the plasma biomarkers as described in Fig. 1. The estimated mediational effect, , was significantly different from 0 in the test for PC1 [ = 0.38; 95% confidence interval (95% CI), 0.02–0.70] and for one individual marker, IGF-I ( = 0.16; 95% CI, 0.07-0.31). The models were evaluated using cancer count as the outcome measure because there was only one cancer-free sedentary control animal, making logistic regression unstable. The statistical results are summarized in Table 5.
Mediational model estimates
Intervening variable . | Parameter estimates (95% CI)* . | ||
---|---|---|---|
τ . | τ′ . | αβ . | |
PC1 | −0.62 (−0.85 to −0.39) | −0.24 (−0.60 to 0.10) | 0.38 (0.02 to 0.70) |
IGF-I | −0.47 (−0.71 to −0.23) | 0.16 (0.07 to 0.31) |
Intervening variable . | Parameter estimates (95% CI)* . | ||
---|---|---|---|
τ . | τ′ . | αβ . | |
PC1 | −0.62 (−0.85 to −0.39) | −0.24 (−0.60 to 0.10) | 0.38 (0.02 to 0.70) |
IGF-I | −0.47 (−0.71 to −0.23) | 0.16 (0.07 to 0.31) |
Discussion
In a recent series of papers (7, 8, 10), our laboratory has reported on the effects of a wheel-running intervention on the postinitiation phase of experimentally induced breast cancer and identified candidate intracellular pathways that may, in part, account for the cancer-inhibitory activity associated with wheel running. Thus, the focus of those experiments was on cell autonomous mechanisms that regulate cell proliferation, apoptosis, and angiogenesis. However, as discussed in ref. (17), it is equally important to evaluate what are referred to as host-related factors that are also undoubtedly exerting effects on the carcinogenic process, perhaps through these cell-autonomous mechanisms. Clearly, these two levels of mechanistic inquiry, cell autonomous and host related, are not only interactive but also context dependent; that is, the manner in which host-related factors exert effects on a particular population of cells depends on the signaling pathways misregulated within those cells. A good example of this context dependence is illustrated by the work reported in ref. (18) in which transplanted cancer cells differentially responded to dietary restriction depending on the characterized mutations known to occur in those cells. With this in mind, we decided to use plasma available from a population of animals that were either sedentary controls or given access to an activity wheel, and in which the effect of being physically active on the carcinogenic response was similar to that associated with a physically active lifestyle in human populations. As shown in Table 1, physical activity was very protective against the development of mammary cancer in these animals. To limit complexity, the animals in this study were nonobese because obesity is a host-related factor associated with both the carcinogenic process and cancer prognosis (17). Thus, we modeled the situation of individuals with a normal body mass index who are sedentary or physically active (1). The plasma biomarker data were interrogated with four goals in mind: (a) to determine if the biomarkers assessed responded to the physical activity intervention, (b) to determine how those biomarkers were associated with the presence or absence of cancer in the population of animals studied irrespective of the treatment group to which they were assigned, (c) to determine how those biomarkers were associated with the intensity of the carcinogenic response measured as the number of cancers per animal, and (d) to assess the feasibility of reducing the dimensionality of the 20 biomarkers to a single score.
The 20 biomarkers were grouped into three families of host-related effector molecules based on a dominant process in which they participate; however, this was done recognizing that the majority of these molecules exert pleiotropic effects that are context dependent and often interrelated. As shown in Table 2, the majority of plasma biomarkers were affected by wheel running. Only fasting glucose, C-reactive protein, and plasma estradiol were not significantly affected by treatment group. These same data were then subjected to PCA, which is a useful method for data reduction. The technique is based on the relationships among the features of interest, in this case 20 biomarkers, without reference to the outcome variables, and without imposing assumptions on the data. The most useful property is that the first PC is the linear combination of the features with the highest variance among all possible linear combinations. In this sample of data from 110 animals, PC1 alone provided excellent discrimination between sedentary controls versus wheel runners. Only 4.5% of the 110 animals were misclassified. The discrimination is shown in Fig. 2A and applies to animals with a wide range of physical activity exposures because animals ran different distances (between 2 and 7 km/d) and at either a specified intensity (40 m/min) in the motorized wheel or at self-determined intensity in the free wheel. This finding lends support to the concept that a linear combination of biomarkers can be constructed to identify individuals with a physically active lifestyle. Such a tool could be invaluable for population-based studies as well as the clinic-based intervention currently under consideration to evaluate the effects of physical activity on cancer risk (19). Additionally, we used the first two PCs to produce multidimensional preference plots for graphic exploration of the relationships among the biomarkers and between the biomarkers and outcomes. Examination of the unsupervised clustering of biomarkers as shown in the multidimensional scaling plots presented in Fig. 2 not only provided evidence for biomarker clustering consistent with the initial groupings but also suggests relationships among the families.
We next applied the same sequence of analyses to interrogate the relationship between the 20 plasma biomarkers and the presence or absence of cancer in an animal. This was done both for cancer incidence (Table 3; Fig. 2B) and cancer count per animal (Table 4; Fig. 2C), which reflects the intensity of the carcinogenic response in an animal, recognizing that each cancer in this model is considered clonal in its origin (20). A smaller subset of biomarkers was associated with the presence or absence of cancer (specifically corticosterone, leptin, IL-2, IL-4, IL-6, IL-10, and IFN-γ) and with cancer count (specifically corticosterone, leptin, IL-2, IL-4, IL-6, IL-10, IFN-γ, glucose, insulin, growth hormone, IGF-I, and IL-1α) than with being physically active; however, there was notable overlap between the lists. Figure 2 (B and C) shows how cancer incidence (yes/no) and cancer count related to the first two PCs of the biomarker data. The first PC from that analysis was associated with the presence or absence of cancer (P < 0.007) and the number of cancers per rat (P < 0.003); however, Fig. 2 (B and C) shows that discriminatory power is much poorer than for treatment group. Nonetheless, the strength of association between the first PC of the biomarkers and the two outcomes investigated, cancer incidence and cancer count, as well as treatment group, supported a third analysis, the test for mediation.
Identification of the mechanisms of a disease process is an important step toward prevention and treatment. As it was initially defined (15), mediation is a causal model of the process by which a treatment induces an outcome. The treatment is a causal factor for the intervening variables, which in turn are causal factors in the outcome. The causal pathway results in a statistically significant relationship between the treatment and the outcome. In the case of complete mediation, the statistical significance of the treatment is lost when the intervening variables are added to the regression: The full effect of treatment is accounted for by the effect of treatment on the mediator and the effect of the mediator on the outcome (αβ). Mathematically, this result is indistinguishable from confounding. Conceptually, however, mediation and confounding are very different; the defining element for mediation is causality (21).
In a carefully designed experiment, randomization and temporal precedence can rule out alternative explanations for the observed effect and give compelling support to demonstrations of the links in the causal pathway. In the absence of a strong design, the most that can be said when the required statistical significance is observed, as in Table 5, is that the data are consistent with the mediational model; there could be other models that fit the data as well or better. The single randomization in our experiment gives strong evidence of causality for the total effect of wheel running on the cancer outcome. This is the path labeled τ in Fig. 1. In addition, we can make a reasonably strong claim for a causal relationship for wheel running on many of the 20 biomarkers and PC1 in particular, although the experiment was not powered a priori to detect effects on these measures. Had this been the primary aim of the experiment, the 20 biomarkers would have been designated as primary outcomes in addition to the cancer outcomes, and the type I error rate would have been controlled accordingly. To establish causality from the intervening variable(s) to the outcome, the path labeled β in Fig. 1, the researcher must be able to manipulate the intervening variable. In this study, that would have entailed an additional randomization after the first. That is, animals within each treatment arm would have been further randomized to high or low levels of the intervening variable(s), as in a split plot design. By removing this design element, which would be difficult if not impossible to implement, we can conclude the association but not the causality between the intervening variable(s) and the outcome. In short, our data are consistent with the mediational model. This needs confirmation in a study that is adequately powered and designed specifically to show causality on the β path.
Considering the results of all three analytic methods, ANOVA, PCA, and mediational analysis, a brief discussion of the various biomarkers is warranted despite their detailed discussion in a number of reviews, including refs. (9, 10, 17).
The data strongly support the involvement of the biomarkers that we assigned to glucose homeostasis and metabolism in the mediation process, particularly when it is recognized that IGF-I alone also gave a significant result in the mediation analysis (Table 5). Of the biomarkers included in this family, T3 and T4 seemed to be minimally informative, an observation supported by the distinct clustering of these molecules away from other family members. The inflammatory and cytokine family members were also markedly affected by wheel running; however, they are in general less well understood and the clustering patterns generate numerous hypotheses about pro- versus anti-inflammatory effects of physical activity and how they relate to cancer risk and cancer prognosis. It is, however, notable that in this nonobese animal model, IL-6 and C-reactive protein clustered in different groups and that C-reactive protein was unaffected by wheel running. On the other hand, IL-6 was increased in wheel-running animals. As discussed in ref. (10), IL-6 has many activities. These data seem to point to an effect on IL-6 that is independent of its induction of C-reactive protein synthesis in the liver, which has been reported in obesity-associated chronic inflammation. Relative to the sex steroid family, it was surprising to us that progesterone levels were markedly higher in wheel-running animals, but that little effect of exercise was observed on estradiol. Although the latter observation is consistent with other reports of premenopausal models for breast cancer, the progesterone effect could conceivably relate to the position of progesterone in steroid biosynthesis in which it can serve as a precursor for corticosterone and other steroids, noting the marked elevation of plasma corticosterone in wheel-running animals and that this biomarker was also strongly associated with cancer outcomes, although it was not significant when tested as a single mediator of the wheel-running effect.
Limitations, translational opportunities, and future directions
Much of the literature on mechanism has focused on specific pathways, as if there is one biological effect that physical activity has that accounts for its effect on breast cancer, taking a Popperian approach of competing the various hypothesized mechanisms against one another to identify “the mechanism.” Our approach, in using PCA of multiple biomarkers, is a significant departure in causal theory from the “single key mechanism” perspective. Despite the limitations of PCA noted below, the fact that PC1 seemed to have a mediating effect that was stronger than any single biomarker suggests that physical activity may affect breast cancer through multiple pathways. As such, this successful PCA-based approach to data analysis has the potential to alter the theoretical understanding and approach to mechanistic studies of physical activity.
Although the assumption of multivariate normality is important for valid hypothesis testing on the characteristic roots of the covariance matrix, this was not our intention. We examined the within-group PCA for evidence that the variance-covariance matrix differed in sedentary versus exercising animals, and found none. No formal statistical testing was done, although more than half the biomarkers were symmetric, if not arguably normally distributed. Because even when PCA is used strictly for data reduction, and the use of ranks is recommended if data are badly skewed (12), we compared the results presented herein with those obtained by working with rank-transformed data and found negligible differences.
Because of the strength of the relationships between physical activity and cancer risk and prognosis, discussions are ongoing about the rationale for and design of a clinical trial to investigate the effects of a physical activity intervention on either breast cancer risk in women at elevated risk for the disease or on disease-free interval in breast cancer survivors (19). The markers identified as potential risk factors are consistent with previous research and epidemiologic studies. However, this was a secondary analysis on data from other studies not designed specifically to test the mediational hypothesis. These results are all hypothesis generating and need to be reproduced. The combination of 20 markers into a single score is appealing due to its parsimony; the simple linear combination could be refined (to require fewer factors, a more sophisticated function, or both) to produce a reliable indicator for physical activity in a population of free-living adults, and possibly a marker for risk of disease.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
Acknowledgments
We thank John McGinley for his assistance in the preparation of the manuscript, and Vanessa Fitzgerald, Elizabeth Neil, Denise Rush, and Jennifer Sells for their excellent technical assistance.
Grant Support: USPHS grants U54-CA116847 and R01-CA100693 from the National Cancer Institute.
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.