| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Imaging, Diagnosis, Prognosis |
Authors' Affiliations: 1 Prediction Sciences, La Jolla, California and 2 Department of Pathology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
Requests for reprints: Steven P. Linke, Prediction Sciences, 9404 Genesee Avenue, Suite 210, La Jolla, CA 92037. Phone: 858-404-0454; Fax: 858-777-3614; E-mail: slinke{at}predict.net.
 |
Abstract |
|---|
 Top Abstract Materials and MethodsResultsDiscussionReferences |
|---|
Experimental Design: This was a retrospective study of 324 stage I to III female breast cancer patients treated with tamoxifen for whom standard clinicopathologic data and tumor tissue microarrays were available. Nine molecular markers were studied by semiquantitative immunohistochemistry and/or fluorescence in situ hybridization. Cox proportional hazards analysis was used to determine the contributions of each variable to disease-specific and overall survival, and machine learning was used to produce a model to predict patient outcome.
Results: On a univariate basis, the following features were significantly associated with worse survival: high pathologic tumor or nodal class, histologic grade, epidermal growth factor receptor, ERBB2, MYC, or TP53; absent estrogen receptor (ER) or progesterone receptor; and low BCL2. CCND1 and CDKN1B did not reach statistical significance. On a multivariate basis, nodal class, ER, and MYC were statistically significant as independent factors for survival. However, the benefit of ER-positive status was moderated by BCL2, ERBB2, and progesterone receptor. BCL2 and TP53 also interacted as an independent risk factor. A kernel partial least squares polynomial model was developed with an area under the receiver operating characteristic curve of 0.90.
Conclusions: Our data show the predictive value of BCL2, ERBB2, MYC, and TP53 in addition to the standard hormone receptors and clinicopathologic features, and they show the importance of conditional interpretation of certain molecular markers. Our multimarker predictive model performed significantly better than standard guidelines.
Roughly 75% of breast cancers are positive for the hormone-based estrogen receptor (ER) and/or progesterone receptor (PGR; ref. 2). Binding of estrogen to ER causes its phosphorylation and dimerization followed by transcription of a variety of genes, including secreted growth and angiogenic factors, as well as PGR (3). Because the ER pathway can contribute to tumor growth, most hormone receptor–positive patients are treated with an inhibitory endocrine therapy either as an adjuvant to radiation and/or surgery in early-stage disease or as the primary treatment in more advanced disease. The most common endocrine therapy has been the selective ER modulator tamoxifen. It has been in use for >20 years and demonstrably prolongs survival (4).
In breast tissue, tamoxifen competes with estrogen for binding to ER, thereby reducing proliferation through inhibition of the ER pathway. However, 
40% of hormone receptor–positive patients fail to respond to tamoxifen (5–7). Tamoxifen resistance can arise through several direct mechanisms, such as defects in the ER pathway, cross-talk between the ER pathway and other growth factor pathways (6, 8, 9), and deregulated expression of ER coregulators (8, 10). Cancers may also be resistant to tamoxifen through indirect mechanisms, such as activation of aggressive pathways that enable growth, invasion, and/or metastasis.
Early identification of breast cancer patients with tamoxifen-resistant cancers is important, as the patients may be more responsive to alternative or additional therapies, such as other endocrine therapies, targeted agents, or cytotoxic drugs. There are several current standards that assign patients to risk categories based solely on hormone receptor status and/or clinicopathologic features. These include the NIH guidelines (11), the St. Gallen guidelines (12), and the Nottingham Prognostic Index (NPI; ref. 13). These standards provide three general levels of guidance regarding the recommendation of cytotoxic chemotherapy: 1, low risk, chemotherapy not recommended; 2, intermediate risk, chemotherapy encouraged but optional; 3, high risk, chemotherapy strongly recommended. Unfortunately, these standards are relatively inaccurate at classifying patients, and they produce ambiguous results for a large percentage.
In the past several years, several genomic studies have reported subsets of predictive/prognostic genes based on differential expression (e.g., refs. 14–16). Although such molecular assays show promise, numerous questions have been raised recently about the experimental and statistical methodologies used in the studies, creating concerns about their reproducibility and true superiority over the current standards (e.g., refs. 17–22).
In this study, we analyzed data on the standard hormone receptors (ER and PGR) as well as the growth factor receptors epidermal growth factor receptor (EGFR) and ERBB2. In addition, we studied the tumor suppressors CDKN1B and TP53, the antiapoptotic factor BCL2, the proliferation marker CCND1, and the MYC oncogene. Our final model included the standard markers ER, PGR, and ERBB2 as well as BCL2, MYC, and TP53. Although some studies have indicated that the latter markers have prognostic significance, no consensus has been reached on their utility. Here, we show the importance of conditional interpretation of certain markers on others due to their interdependency. In combination with the application of advanced informatics, this enabled the development of a predictive model that performed significantly better than the current standards.
|
Materials and Methods |
|---|
|
TopAbstractMaterials and MethodsResultsDiscussionReferences |
|---|
Immunohistochemistry. Mouse monoclonal antibodies (clone; epitope, if applicable; dilution) against ER (1D5; NH2 terminus; 1:1,000), PGR (1A6; A/B region; 1:600), BCL2 (124; 1:1), CDKN1B (SX53G8; 1:1,000), EGFR (EGFR.113; extracellular domain; 1:20), and TP53 (DO-7; NH2 terminus; 1:1) were used for immunohistochemical analysis. All antibodies were obtained from DAKO (Glostrup, Denmark) except PGR and EGFR, which were obtained from Novocastra (Newcastle-upon-Tyne, United Kingdom). The HercepTest kit (DAKO) was used for ERBB2.
Tissue microarrays constructed from formalin-fixed, paraffin-embedded primary tumor samples (25) were stained with a standard immunoperoxidase immunohistochemistry protocol (24). Tumors with known positivity were used as positive controls, and the primary antibodies were eliminated for negative controls. Most of the markers were scored for both intensity (on a scale of 0-3) and the estimated percentage of positively staining cells in 10% increments. Final scores on a 0 to 3 scale (0, none; 1, weak; 2, moderate; 3, strong) were determined from a combination of these attributes for these markers. For the statistical analyses, tumors were considered positive for ER or PGR when staining was evident in 
10% of cells. BCL2 was considered "high" when the final score was 3 (strong). Tumors were considered positive for EGFR when the final score was 1 to 3 (weak to strong). ERBB2 staining was scored only on the intensity scale of 0 to 3, and it was considered "positive" when the score was >0. This was based on our previous observation that 1+ ERBB2 staining on tissue microarrays is associated with HER-2 amplification in a high fraction of cases in our laboratory (26).
Fluorescence in situ hybridization. CCND1, ERBB2, and MYC gene amplifications were determined as described previously (23). Briefly, the tissue microarrays were proteolyzed, deparaffinized, dehydrated, and denatured. They were then subjected to standard dual-label fluorescence in situ hybridization with Spectrum Orange–labeled gene-specific probes and Spectrum Green–labeled centromere probe controls from chromosomes 11 (CCND1), 17 (ERBB2), and 8 (MYC) from Vysis, Inc. (Downers Grove, IL). The nuclei were counterstained with 4',6-diamidino-2-phenylindole in antifade solution and examined by indirect fluorescence microscopy. A gene was considered amplified if the ratio of its signal number to that of the corresponding centromere was 2.
Cox proportional hazards analysis. Survival measures were defined as the proportions of patients who were still alive for a defined number of months after diagnosis. For overall survival, death from any cause was included. For disease-specific survival, patients who died due to a cause other than cancer were censored. All variables were studied with Cox proportional hazards analysis.
Univariate analyses were conducted on clinicopathologic features and molecular markers, and conditional analyses were conducted on marker pairs with plausible biological interactions. In some cases, the features were also examined as categorical data: pT1-2 versus pT3-4, pN0-1 versus pN2-3, TP53%, and age. For TP53% and age, a limited number of thresholds were examined based on the data distribution and values reported in previous publications to find a good threshold for categorical analysis of hazard ratios (HR). To control the false discovery rate for multiple tests, q-values were calculated from the corresponding Ps from all tests, including tests not shown in the final tables (27–29). Setting the maximum false discovery rate at <5% (q < 0.05), a corresponding P of 0.047 was determined, and only those features with P < 0.047 were considered significant.
The multivariate model was constructed from the significant results of the univariate and conditional analyses. The features were assessed with the Wald statistic, and nonsignificant ones were sequentially removed to produce the final model.
Analyses were conducted with the computer software MATLAB version R13 (The Mathworks, Inc., Natick, MA) and R (30) with the Survival package (31) and the q-value package (27–29).
Machine learning-based modeling and analysis. A machine learning-based approach was used to combine the variables identified in the multivariate Cox proportional hazards model. Because of the form of the variables in the Cox model, all of the data were coded in a binary format. To classify patients, a model was trained using the kernel partial least squares (KPLS) method with a third-order polynomial kernel and the maximum number of latent factors equal to the number of variables. This allowed up to third-order interactions between features in Mercer kernel space with a limited number of model variables (32). The model was trained in MATLAB R13 using elements from SPIDER version 1.5 (33).
We trained the multimarker KPLS model and evaluated the results based on 5-year overall survival. Patients were excluded if they were censored within 5 years due to loss of follow-up or were negative for both ER and PGR or if any of the constructed features were indeterminate due to missing data. Given its relatively low incidence, patients with unknown MYC amplification status were assumed to be MYC normal. A total of 123 patients (89 with and 34 without 5-year overall survival) met these requirements (the modeling set).
We evaluated the utility of the multimarker model using leave-one-out cross-validation. For each patient in the modeling set, a model was trained using all of the data, except for the subject of interest. After parameterization was completed, the model was applied to the patient data that had not been used during model training. This process was repeated until a validated prognosis score was obtained for each patient in the modeling set.
Receiver operating characteristic (ROC) curves were generated for the multimarker leave-one-out cross-validation results, the NPI, and the 2005 St. Gallen guidelines from the 123-patient modeling set based on 5-year overall survival. Continuous raw scores were used for both the multimarker model and the NPI, whereas three discrete risk categories were used for the St. Gallen guidelines. Bootstrapping (10,000x) was used to obtain a robust, smooth estimate of the mean ROC curve, and the area under the mean ROC curve was calculated for each prognostic index (34).
For Kaplan-Meier survival analysis of the multimarker model, 201 patients were analyzed. A final multimarker KPLS model was constructed using all 123 patients in the modeling set, and this final model was used to score 78 additional patients who previously have been censored due to loss of follow-up. For the patients in the modeling set, the leave-one-out cross-validated scores were used. Patients with multimarker scores greater than a threshold corresponding to 81% sensitivity for <5-year overall survival were assigned to a "bad" prognosis group, whereas the others were assigned to a "good" prognosis group.
|
Results |
|---|
|
TopAbstractMaterials and MethodsResultsDiscussionReferences |
|---|
|
|
In univariate Cox proportional hazards analysis (Table 2), the lack of ER or PGR or the presence of EGFR, ERBB2, or amplified MYC were all significantly associated with shorter survival. Low BCL2 (all scores below the maximum of 3) also was significantly associated with shorter survival. TP53 was significantly associated with worse outcome when the staining intensity was moderate to strong (TP53 high intensity). A high percentage (70%) of cells staining positively for TP53 was not significant for overall survival and was only marginally significant for disease-specific survival in this independent univariate analysis. However, it became more significant when considering interactions with BCL2. The HRs for disease-specific survival and overall survival were not statistically different in all cases when the covariate remained statistically significant for overall survival, although only the Ps are shown for overall survival. CCND1 and CDKN1B were not significant in this analysis, and the number of EGFR-positive patients was insufficient to assess its contribution.
Conditional analysis of molecular markers to determine interactions. Next, analyses were conducted on marker pairs with plausible biological interactions Several of the molecular markers exhibited dependencies relative to other markers, which increased their predictive values. For example, disease-specific survival and overall survival of ER-positive patients who were PGR negative and who had low BCL2 scores were not statistically different than ER-negative patients. However, if either PGR was present or the BCL2 score was high, the ER-positive patients had significantly better outcome. Independent of this observation, ER-positive patients who were both PGR positive and BCL2 high experienced even better outcome (Table 3; Fig. 1A). Similarly, ERBB2 positivity was significantly associated with worse outcome in ER-positive patients but not in ER-negative patients, although the number of ER-negative/ERBB2-positive patients was relatively small (Table 3; Fig. 1B).
|
|
Not surprisingly, values for TP53 intensity (i.e., the percentage of cells staining positively for TP53) and TP53 score correlated with each other in individual patients. Although any amount of TP53 staining typically is indicative of the presence of a mutant form, we observed a sudden and significant decrease in survival in patients with the highest intensity/overall score compared with those with weak or moderate values (e.g., 5-year disease-specific survival was 82-86% when TP53 intensity was 0-2 and only 53% when the intensity was 3). Based on analysis of all TP53 staining variables, we determined that 70% of cells staining positively was the most useful cutoff.
Multivariate model. A multivariate Cox proportional hazards model was constructed based on the univariate and conditional analyses. pN, age, MYC, ER (including interactions with PGR and BCL2), BCL2 (including interaction with TP53), and ERBB2 (including interaction with ER) remained independent for both disease-specific survival and overall survival (Table 4). The overall Ps (log-rank statistic) of the multivariate model were highly significant at 3.22E–12 and 3.58E–09, respectively, for disease-specific survival and overall survival.
|
|
Using the NPI low-risk threshold (score = 3.4) as a cutoff to assign patients to good and bad prognosis groups, the NPI correctly identified 81% of the patients who died within 5 years by placing them in the intermediate-risk or high-risk category (sensitivity). However, 67% of the patients who were predicted to die ended up surviving (false positives), which could have led to their overtreatment with more aggressive therapy. In contrast, at this same 81% sensitivity, the multimarker model had a false-positive rate of only 13% (Fig. 2A). Using the 81% sensitivity cutoff, Kaplan-Meier survival analysis also revealed the superiority of the multimarker model over the NPI in accurately separating patients into good and bad prognosis categories (Fig. 2B and C).
|
Discussion |
|---|
|
TopAbstractMaterials and MethodsResultsDiscussionReferences |
|---|
Our analysis revealed several marker dependencies and interactions. For example, although BCL2, ERBB2, and PGR were significant univariate factors, they provided better predictive value when considering their dependencies on ER status and/or interactions with each other. The same was true for the interaction between BCL2 and TP53. In addition, preliminary evidence suggests that different cutoffs for TP53 staining may be relevant in different patient subsets based on the status of other molecular markers. Advanced, nonlinear machine learning methodologies were used to integrate clinicopathologic features with the molecular markers into a predictive model that performed significantly better than standard guidelines.
The predictive value of ER in tamoxifen response is well established. PGR is an estrogen-regulated gene product (37). Thus, the presence of PGR may be a surrogate indicator of a functional estrogen response pathway, particularly in cases where ER is present at functional levels that are too low to detect (false negative). Consistent with several previous studies (e.g., ref. 38), we found that PGR was a predictive factor for tamoxifen treatment in univariate analysis (Table 2).
PGR negativity may arise when ER is detected but is a nonfunctional mutant or variant (false positive; ref. 38) due to signaling through alternative growth pathways, such as EGFR/ERBB2 or insulin growth factor-I receptor (39, 40). In the subpopulation of ER-positive patients, PGR negativity increased the risk of tamoxifen treatment resistance (HR, 2.1; P = 0.02). Thus, although PGR shows univariate statistical significance, its role in predicting tamoxifen treatment response, along with BCL2, may be better elucidated in the context of ER status (Table 4).
EGFR and ERBB2 are growth factor receptor tyrosine kinases that initiate cell survival and proliferation signaling cascades. They are elevated in 15% and 25% of breast cancers, respectively. In the presence of the appropriate peptide growth factors, activation of these pathways may overcome the growth inhibitory effects of tamoxifen on the ER pathway. In addition, there is substantial cross-talk between the ER pathway and the ERBB2 and EGFR growth factor pathways (6, 7, 9). For example, there is evidence that various downstream members in these pathways (e.g., extracellular signal-regulated kinases 1 and 2 and AKT) can directly activate ER. Reciprocally, there is evidence that ER can directly activate members of the ERBB2 and EGFR pathways. Interestingly, binding of ER by either estrogen or tamoxifen may be sufficient for this activation. In fact, a preclinical study indicates that tamoxifen can actually stimulate cell proliferation in ERBB2-positive breast cancer cells, shifting tamoxifen from an antagonist to an agonist role (9). Consistent with this finding, ERBB2-positive patients given tamoxifen can experience even higher rates of recurrence than untreated patients (41).
In agreement with these previous studies, we found that EGFR or ERBB2 positivity were predictive of tamoxifen resistance. Unfortunately, there were an insufficient number of EGFR-positive patients to include it in the multivariate model. Interestingly, although ER and ERBB2 levels tend to be inversely related in breast cancers, ERBB2 positivity was a statistically significant risk factor in the ER-positive patient subpopulation. This suggests that ERBB2 worked mainly by reducing the effective inhibition of the ER pathway by tamoxifen perhaps through growth factor pathway cross-talk.
Several studies indicate that a low BCL2 level is associated with worse outcome in tamoxifen-treated breast cancers (e.g., refs. 42–44). This is counterintuitive, as BCL2 is an antiapoptotic factor that might be expected to inhibit drug-induced apoptosis in the tumor cells. However, there is evidence that, similar to PGR, the BCL2 gene itself is ER regulated. Thus, high BCL2 may be indicative of an intact ER pathway that is driving tumor growth and should be sensitive to endocrine therapy (45). Furthermore, in those tumors in which it is highly expressed, BCL2 may be the leading antiapoptotic factor, so down-regulation with tamoxifen may be an effective inducer of apoptosis. In addition, in the multivariate analysis, low BCL2 predicted worse outcome independent of ER status, indicating that BCL2 participates in both ER pathway–dependent and ER pathway–independent mechanisms.
Mutations in the tumor suppressor TP53, most of which lead to elevated basal levels of the protein, are observed in 30% of breast cancers. TP53 can be activated by stresses, such as DNA damage, leading to its regulation of genes that induce growth arrest or apoptosis through either transcription-dependent or transcription-independent mechanisms. Numerous studies show that mutant TP53 is associated with resistance to endocrine therapies, including tamoxifen (e.g., refs. 42, 46).
In this study, high TP53 predicted worse outcome independent of ER. Interestingly, this was related to the same observation made with low BCL2. Patients with either high TP53 or low BCL2 were at similar risk, but having both of the markers in this state did not further increase risk. Other studies have implemented subpopulation grouping based on BCL2 and TP53 with varying results. For example, one study reported that TP53 status was only significant in the BCL2-positive subset (44). Other reports have shown utility in separating ER-positive and ER-negative patients by TP53 and/or BCL2 status to determine subgroups with different prognoses (42, 44, 47). We saw similar, although not identical, results in our patient set. However, we determined that directly combining BCL2 and TP53 and separately assessing BCL2 in the context of ER status achieved better predictive power than these alternative methods. One confounding factor was that our study included only tamoxifen-treated patients, whereas the others also included patients who received various cytotoxic chemotherapy regimens.
Although mutations that lead to loss of TP53 function are well characterized, there is also evidence that some TP53 mutants exert gain-of-function effects. Such mutants have altered transcriptional activities and/or protein-binding targets, favoring growth and/or apoptosis resistance. Our observation that low to moderate TP53 does not significantly affect outcome compared with high TP53 raises the possibility that high levels of TP53 gain-of-function mutants can contribute to resistance independent of the ER pathway.
When amplified, MYC can inappropriately stimulate cell division through its functions in metabolism, replication, differentiation, and apoptosis (48). Approximately 11% of analyzable patients exhibited MYC amplification in this study, which is consistent with previous findings (48). Although MYC amplification is reportedly associated with ER negativity (23), it was a strong predictor of poor outcome independent of all other variables, including nodal and hormone receptor status. This may be related to the cellular functions described above, but it could also be a more general indicator of elevated genomic instability (23).
Current standard guidelines to classify patients into risk categories for recurrence include the NIH, St. Gallen, and NPI. Although these guidelines were not developed to specifically predict resistance to tamoxifen, they are used here in the same manner they would be used by an oncologist in hormone receptor–positive patients to help determine whether these candidates for tamoxifen monotherapy are at sufficiently high risk of recurrence to justify more aggressive therapy, such as cytotoxic chemotherapy.
The NIH and older St. Gallen guidelines categorize a very large number of patients in the "intermediate-risk" or "high-risk" categories. Although this results in a very low "false negative for bad outcome" rate, it leads to the overtreatment of a sizable proportion of patients. The NPI improves performance by using an algorithm based on multivariate analysis of clinicopathologic factors from retrospective studies of breast cancer patients, and the new 2005 St. Gallen guidelines improve performance by incorporating the molecular marker ERBB2. However, they still incorrectly categorize a relatively large number of patients in the intermediate and high categories and fail to identify an important subset of higher-risk patients.
In attempts to produce superior predictive/prognostic models, several gene expression profiles are under development (e.g., refs. 14–16). However, a variety of serious questions have been raised about the experimental and statistical methodologies used in many of these studies (e.g., refs. 17–22). For example, data overfitting is a common problem, in which thousands or even tens of thousands of genes are analyzed in a relatively small number of patients. In many cases, the validation sets are not entirely independent of the training sets or they are too small to establish reliable 95% confidence intervals (95% CI) for prediction accuracy. It is also interesting to note that application of more sophisticated algorithms to the standard clinicopathologic data, such as the NPI, or the use of an artificial neural network may essentially match the performance of gene expression signatures in the same patient set (21).
Beyond the statistical issues, gene expression assays can only measure transcript levels, which do not always correlate with functional protein levels, and they cannot detect protein mislocalization. In addition, the assays are relatively complicated and costly, often requiring sophisticated and/or proprietary technology and multiple steps, including methods to try to reduce the contribution of adjacent nontumor tissue and to account for RNA degradation. Thus, several significant challenges still face this promising technology.
In this study, we used a machine learning methodology, allowing us to account for multiple marker interactions in a limited patient set for the development of a prognostic model. We employed a false discovery rate method using q-values to limit the number of false-positive identifications to compensate for multiple comparisons testing. In addition, we prevented overfitting and added robustness to the modeling process by employing leave-one-out cross-validation. We have directly shown that the model is superior to not only the cruder NIH and old St. Gallen guidelines but also the more sophisticated NPI and 2005 St. Gallen guidelines. We speculate that it also performs as well as or better than gene expression profile signatures without many of the associated drawbacks. Furthermore, the model integrates clinicopathologic features with demonstrated prognostic utility and molecular markers with established roles in drug response and general tumor aggressiveness, which can be collected with well-characterized and cost-effective immunohistochemistry and fluorescence in situ hybridization assays.
To produce the survival curves (Fig. 2B and C), a specific threshold value was chosen to categorize patients with good or bad prognosis with tamoxifen treatment alone. However, the multimarker model can be used to produce a risk of recurrence percentage as a continuous function of the score, so patients and their oncologists could be provided with a more specific risk rating. Alternatively, two thresholds could be chosen, one at specificity x to recommend against more aggressive therapy and the associated side effects and another at sensitivity y to recommend more aggressive therapy. Although there would be an intermediate group of patients with ambiguous scores using this strategy, the set of patients would be significantly smaller than it is with the current prognostic standards in which a majority or plurality of patients end up in the intermediate category.
Although this study has been fully cross-validated, it will be useful to independently validate it on a completely new set of patients. In addition, it will be useful to extend the study to additional markers with putative roles in tamoxifen resistance, such as ER coregulators (8), and tumor aggressiveness, such as factors reflecting proliferation, invasion, anchorage-independent growth, and/or angiogenesis. These markers could further increase the sensitivity and specificity of the predictive model. Given the mechanism of action of tamoxifen and the specific molecular markers included in this study, our model likely will contribute significantly to predictive tests for other endocrine therapies, such as aromatase inhibitors, which are becoming the standard treatment in postmenopausal women. Considering the very different costs and side effects associated with tamoxifen and aromatase inhibitor treatments, it would be interesting to compare their efficacy based on molecular modeling. Our goal is to produce a test that will enable the timely administration of the most biologically and cost-effective treatments, contributing positively to both life expectancy and quality of life for breast cancer patients.
|
Footnotes |
|---|
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.
Received 7/20/05; revised 10/20/05; accepted 12/ 7/05.
|
References |
|---|
|
TopAbstractMaterials and MethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  I. Zlobec, R. Steele, L. Terracciano, J. R Jass, and A. Lugli Selecting immunohistochemical cut-off scores for novel biomarkers of progression and survival in colorectal cancer J. Clin. Pathol., October 1, 2007; 60(10): 1112 - 1116. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. M. Giltnane, L. Ryden, M. Cregger, P.-O. Bendahl, K. Jirstrom, and D. L. Rimm Quantitative Measurement of Epidermal Growth Factor Receptor Is a Negative Predictive Factor for Tamoxifen Response in Hormone Receptor Positive Premenopausal Breast Cancer J. Clin. Oncol., July 20, 2007; 25(21): 3007 - 3014. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 F. Bertucci, D. Birnbaum, and A. Goncalves Proteomics of Breast Cancer: Principles and Potential Clinical Applications Mol. Cell. Proteomics, October 1, 2006; 5(10): 1772 - 1786. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Cancer Research | Clinical Cancer Research |
| Cancer Epidemiology Biomarkers & Prevention | Molecular Cancer Therapeutics |
| Molecular Cancer Research | Cancer Prevention Research |
| Cancer Prevention Journals Portal | Cancer Reviews Online |
| Annual Meeting Education Book | Meeting Abstracts Online |
