Data for this analysis were obtained from the Biorepository of Clinomics Biosciences Inc, which is a large collection of highly characterized human tissue samples. These samples span a wide range of common diseases, including many forms and stages of cancer, neurological disorders and heart disease. Clinomics has pioneered the development of an emerging new technology known as Tissue Microarray to enable researchers to simultaneously study hundreds of individual tissue samples in parallel, establishing the relative levels of protein expression in those samples and allowing conclusions as to the relevance of these proteins to disease to be made [10,11].

Immunohistochemistry methods were described previously [12,13]. Briefly, tissue specimens were treated with 3% H₂O₂ to quench endogenous peroxidase activity, followed by washing with PBS. After washing, the tissue specimens were first incubated with a specific primary antibody and then with a biotinylated secondary antibody. Substrate-Chromogen was applied to the specimens according to manufacturer’s instructions, followed by stained with hematoxylin. The stain was semiquantitatively examined by pathologists (Clinomics BioSciences, Inc.) using the Allred 8-unit system [14]. Staining was scored on a 0 –5 scale, with 0=no staining. Grades of 1 to 5 represent increased intensity of staining with 5 being strong, dark brown staining. For eachtumor, represented by one slide, the tumor epithelialcells proportion score and intensity score were determined. Peritumoral inflammatoryand stromal cells were not included in the evaluation. The proportionscore included the fraction of positively stained tumor cellsand was as follows: 0 = none, 1 = < 1/100th; 2 = 1/100th to1/10th; 3 = 1/10th to 1/3; 4 = 1/3 to 2/3; 5 = > 2/3. Theestimated average staining intensity of the positive tumor cellswas expressed as follows: 0 = none; 1 = weak; 2 = intermediate ;3 = strong [14]. Each protein was measured with six parameters: Cytoplasmic % Intensity defines percent intensity of stain within cytoplasm; Cytoplasmic % Positive defines percent of all cells positive within cytoplasm; Cytoplasmic Total Score defines the product of Cytoplasmic % Intensity and Cytoplasmic % Positive; Nuclear % Intensity defines percent intensity of stain within nucleus; Nuclear % positive defines percent of all cells positive within nucleus; and Nuclear Total Score defines the product of Nuclear % Intensity and nuclear % Positive.

Anti-EGFR antibody, anti-Her2/neu antibody, anti-phos-pho-EGFR antibody (Tyr845), and anti-phosphoHer2/neu (Tyr877) were from Cell Signaling Technology, Inc. (Beverly, MA).

Information of a breast cancer cohort was extracted from the Cell Signaling Database of Clinomics Biosciences Inc. The study cohort contained 269 breast tumor samples obtained from surgery. The patients had an average age of 63.7 years (ranging from 21 to 95 years). There were 24.9% with Stage 1, 45.0% with Stage 2, 17.8% with Stage 3, and 12.3% with Stage 4 breast cancer. Ninety-one patients (33.8%) were ER negative and 178 (66.2%) were ER positive. Forty-three (16.0%) patients were Her2/neu positive and 226 (84.0%) were Her2/neu negative (assayed by immunohistochemistry). There were 56.9% of patients received lumpectomy and node dissection, and 43.1% of patients received mastectomy and node dissection. Two hundred and two (75.1%) patients accepted CMF chemotherapy. One hundred and fifty-three (56.9%) patients went through localized (breast) radiation therapy. Among 33 patients who developed metastases, 10 patients were responders to the treatments, while 23 were none responders. Treatment response was defined according to RECIS [15]. Responders include complete response (CR) and partial response (PR), and none responders include stable disease (SD) and progressive disease (PR). Among the patients with Stage 0 to 3, 191 patients survived a 5-year disease-free interval, while 44 had recurrence within five years. The remaining patients’ survival information was censored (details in Table 1).

Table 1

Description of the clinical samples (N = 269)

There were two outcomes included in the current analyses: disease free survival after therapy and treatment response. The experiments were designed to construct prognostic models to predict disease-free survival and treatment response in individual patients with breast cancers. The patient cohort was separated into two groups: one for patients with Stage 0 to 3 (Group I, N = 235), the other for patients with metastatic diseases (Group II, N = 33). All patients in Group II received CMF chemotherapy. For Group I, biomarkers were identified to build prognostic models for 5-year disease-free survival. For Group II, treatment response prediction models were constructed using a separately identified biomarker set.

Based on the clinical information in the studied cohort, we performed three analyses for both survival prediction and treatment response prediction, respectively (The details are listed in Results section). The clinical-pathological parameters for outcome predictions were chosen using gain Ratio Attribution Evaluation method with software WEKA. The parameters with merit greater than zero in the evaluation were included in the analysis.

Different machine learning algorithms perform differently according to datasets and the given tasks. The evaluation of different pathologic factors based on a single classifier may not lead to an objective conclusion in general. Therefore, we developed a prediction system combining three selected classifiers, a nearest neighbor method (KStar)[16], Random Forests [17], and Neural Networks (MultilayerPerceptron)from the WEKA3.4 software package¹ [18]. These algorithms are accurate and stable in clinical outcome predictions on the studied cohort. The prediction accuracy of each classifier is evaluated by 10-fold cross validation, which has the least variance and bias of accuracy estimation among all validation methods including the leave-one-out method [19]. The average prediction accuracy of these classifiers is used as the evaluation criterion for breast cancer outcome predictions. If the prediction accuracy is unbalanced for a certain class label, the F-measure [20] was used as the evaluation criterion instead, which is defined as:

F = 2 × (recall × precision) / (recall + precision)

The average of the F-measure of all class labels was used as the final F-measure for a classifier. The average final F-measure of the three classifiers was used as the evaluation criterion of clinical outcome prediction (see Supplementary Information for details).

To assess the significance of the performance of an individual classifier, we computed the probability of the observed prediction accuracy occurring by chance (random prediction using a fair coin flip). The probability of doing at least as well as our prediction models by chance was calculated using Binomial Distribution functions [21] in software package R (http://www.r-project.org/). Statistical significance test was used to assess the significance of different prediction results on the studied cohort.

In order to identify specific high risk breast cancer patients, we used the clinical information in the studied cohort to predict whether or not a patient would survive a 5-year disease free interval. As we discussed previously, the traditional pathologic factors are inclined to be subjective and ER, PR, Her2/neu, as well as the demographic factors tend to be objective. The clinical information was separated into two categories: one includes the traditional pathologic factors and the other includes ER, PR, Her2/neu, plus the demographic factors. To compare the predictive power of these two different kinds of clinical information in a survival prediction, we conducted following analyses on the studied cohort. (1) Using pathologic factors (histology, stage, pT, pN, and nodes positive) to predict breast cancer survival. (2) Using ER, PR, Her2/neu, age, and smoking to predict breast cancer survival. (3) Using a comprehensive profile (age, histology, stage, pT, pN, nodes positive, ER, PR, Her2/neu, and smoking) to predict breast cancer survival. Our analyses demonstrated that the traditional pathologic factors entailed overall accuracy 94.6% in the survival prediction (Table 2, Analysis 1) and the parameters of ER, PR, Her2/neu, plus patient’s age and smoking status resulted in survival prediction accuracy 82% (Table 2, Analysis 2), indicating that the traditional pathologic parameters are more accurate in the individualized risk assessment (p <0.000007). However, when two kinds of predictive factors were incorporated, the survival prediction accuracy was further increased from 82% to 95.8% (p <0.000001) for individual breast cancer patients (Table 2, Analysis 3). It should also be pointed out that the addition of ER, PR, Her2/neu, plus patient's age and smoking status to the traditional pathologic factors did not significantly (p >0.05) improve the prediction of the traditional factors alone. These results are consistent with the observations in Paik et al. [22]. Our results demonstrated that the identified comprehensive classifier can accurately define a patient’s risk and be used to develop an optimal individualized therapy of breast cancer (see Supplement Information for detailed results).

Table 2

Breast cancer disease free survival prediction. Various breast cancer factors were analyzed for the predictive power in survival based on this study cohort (N=235). In each analysis, the clinical-pathological parameters used as predictors in the survival prediction were marked in the corresponding column. The average prediction accuracy of three methods, (KStar), Random Forests, and Neural Networks, was reported as the final result. P value represents the probability that the random prediction can perform at least as well as our prediction models

In order to assess a breast cancer patient’s treatment response, we used the clinical information in the studied cohort to predict whether or not a patient is a responder to chemotherapy. It enables clinicians to determine the most suitable treatment options for each individual patient. To find out the best predictive parameters for the breast cancer treatment response, we conducted the following analyses on the studied cohort of the 33 patients with metastatic disease: (1) Using traditional pathologic factors (histology, stage, pT, pN, nodes positive, metastasis site, surgery procedure, and chemotherapy) to predict a patient’s response to treatments. (2) Using a patient’s ER, PR, Her2/neu, and demographic information (ER, PR, Her2/neu, age, smoking, surgery procedure, and chemotherapy) to predict a patient’s response to treatments. (3) Using a comprehensive profile (age, histology, stage, pT, pN, nodes positive, metastasis site, ER, PR, Her2/neu, smoking, surgery procedure, and chemotherapy) to predict a patient’s response to treatments. Using the traditional pathologic parameters, the accuracy of treatment response prediction is 94.7% (Table 3, Analysis 1). Using ER, PR, Her2/neu, and the patient’s age and smoking status, the accuracy of treatment response prediction is 87.8% (Table 3, Analysis 2). Again, the traditional pathologic factors entailed higher accuracy in the treatment response prediction (p <0.16). However, when the parameters in Analysis 1 and Analysis 2 were combined, the treatment response prediction accuracy was increased from 87.8% to 95.3% (p <0.14) (Table 3, Analysis 3). Here, the addition of the biomarker panel to the traditional histological features did not significantly improve the prediction accuracy (p >0.05). Our data demonstrate that the comprehensive profile encompassing both the traditional pathologic factors and ER, PR, Heu2/neu plus patient’s age and smoking status was more accurate in predicting the treatment response of breast cancer patients. Our constructed classifier can accurately predict a patient’s response to treatments, which enables the selection of the most suitable combination of therapeutic regimens for each individual breast cancer patient. Our results confirmed previous findings. Kendal previously reported that positive lymph node status is an independent risk factor for breast cancer prognosis [23]. Terry et al. identified the association between prolonged cigarette smoking and increased breast cancer risk [24]

Table 3

Treatment response prediction. Various breast cancer factors were analyzed for the predictive power in treatment response based on this study cohort (N=33). In each analysis, the clinical-pathological parameters used as predictors in the treatment response prediction were marked in the corresponding column. The average prediction accuracy of three methods, (KStar), Random Forests, and Neural Networks, was reported as the final result. P value represents the probability that the random prediction can perform at least as well as our prediction models

To more accurately reflect the functionalities of these two proteins in vivo. We performed the following analyses on the studied cohort to investigate the relative strength of the predictive power of total Her2/neu, total EGFR, phospho-EGFR and phospho-Her2/neu in breast cancer survival prediction. The six antibody measurements for each phosphorylated protein were used in disease-free survival prediction. As a control, we first used a patient’s age, ER and PR status to predict the patient survival after therapy (Table 4, Analysis 1). In this analysis, all patients were classified as low risk (5-year survival). It indicates that the predictive power of a patient’s age, ER, and PR is weak, because they cannot identify any high risk patients. However, since majority of the patients (82.1%) survived 5 years in the studied cohort, the overall prediction accuracy is 82.1%. In this situation, we cannot use overall prediction accuracy as the evaluation criterion. As a resolution, we used the average F-measure [20] of the 5-year survival and non-5-year survival prediction results as the evaluation criterion. The F-measure is a combined measurement of recall and precision. It thus overcomes the problems caused by unbalanced prediction error rates. We then evaluated the survival prediction model by adding the expression level of total Her2/neu and total EGFR (Table 4, Analysis 2), as well as phospho-Her2/neu and phospho-EGFR (Table 4, Analysis 3), respectively, to the prediction model. Our results showed that the survival prediction accuracy was increased with the addition of total Her2/neu and total EGFR measurements (p <0.15). Furthermore, our analyses showed that the incorporation of phospho-Her2/neu and phospho-EGFR entailed the highest survival prediction accuracy (p <0.05), while the incorporation of total Her2/neu, total EGFR, phospho-Her2/neu and phospho-EGFR resulted in the second highest survival prediction accuracy compared to the control (p <0.09) (Table 4, Analysis 4).

Table 4

The predictive power of EGFR, Her2/neu, phospho-EGFR, phospho-Her2/neu in breast cancer survival assessment on this study cohort (N=235). Six parameters for antibody measurements were included in the analysis of phospho-EGFR, phospho-Her2/neu, and EGFR. P value represents the statistical significance of the improvement over the control analysis

Both Her2/neu and EGFR belong to the EGFR family. There is growing evidence showing that the interaction among EGFR family members contributes to oncogenic transformation in human tumors. Studies found that human tumors are prone to co-express multi-EGFR family members, which are correlated to the more aggressive phenotypes, enhanced transforming activities, and a poor clinical outcome [25]. Therefore, we compared the predictive power of the combination of phospho-Her2/neu and phospho-EGFR vs. either phospho-Her2/neu or phospho-EGFR. The six antibody measurements for each phosphorylated protein were used in disease-free survival prediction. The protein expression patterns of phospho-EGFR and phospho-Her2/neu in clinical samples are shown in Fig. (1). Among the measured antibody scores, phospho-EGFR Cytoplasmic % Positive, phospho-EGFR Nuclear % Positive, phospho-EGFR Nuclear % Intensity, phospho-Her2/neu Cytoplasmic % Positive, phospho-Her2/neu Cytoplasmic % Intensity, phospho-Her2/neu Nuclear % Positive, and phospho-Her2/neu Nuclear % Intensity were differentially expressed among non-5-year survival and 5-year survival (p <0.05). Furthermore, we found that age-ER-PR-phospho-Her2/neu-phospho-EGFR was more accurate in predicting the survival of breast cancer patients than either age-ER-PR-phospho-Her2/neu or age-ER-PR-phospho-EGFR (Table 5). Our results imply that some interaction may exist between phospholated Her2/neu and phospholated EGFR in breast cancer. These observations are consistent with the clinical results in Her2/neu-targented trastuzumab treatment for breast cancer patients. It was found that some breast cancer patients with Her2/neu overexpression had no response to the treatment. In vitro experiments showed that the response to trastuzumab treatment was dependent not only on Her2/neu expression, but on the expression of other EGFR family member as well [26]. Our results are consistent with the recent publication showing that there is a simultaneous phosphorylation of Her2/neu and EGFR in Her2/neu overexpressed metastatic breast cancer [27]. Taken together, our results demonstrated that the combination of phospho-Her2/neu and phospho-EGFR is more accurate in survival prediction of individual breast cancer patients than either phospho-Her2/neu (p <0.10) or phospho-EGFR (p <0.21).

Fig. (1) Protein expression patterns of pEGFR and pHer2/neu in clinical samples. Antibody measurements: 1. pEGFR Cytoplasmic % Positive, 2. pEGFR Cytoplasmic % Intensity, 3. pEGFR Cytoplasmic Total Score, 4. pEGFR Nuclear % Positive, 5. pEGFR Nuclear % Intensity, 6. pEGFR Nuclear Total Score, 7. pHer2/neu Cytoplasmic % Positive, 8. pHer2/neu Cytoplasmic % Intensity, 9. pHer2/neu Cytoplasmic Total Score, 10. pHer2/neu Nuclear % Positive, 11. pHer2/neu Nuclear % Intensity, 12. pHer2/neu Nuclear Total Score.

Table 5

Comparison of the predictive power among phospho-Her2/neu, phospho-EGFR, and phospho-Her2/neu/phospho-EGFR in breast cancer survival assessment on this study cohort (N=235). P value represents the statistical significance of the improvement over the control analysis