Our study proves that important biological interdependencies can be detected at the level of tumour tissue immunophenotype based on the multivariate analysis of DA data. In a cohort of 109 patients with ductal carcinoma of the breast, we were able to detect biologically relevant interdependencies and heterogeneity largely reflecting the main intrinsic subtypes of the disease and providing new data and insights into the breast cancer biology.
The design of our study enabled us to avoid significant human impacts and assumptions while obtaining the results: we performed an automated image analysis with automated detection of tumour tissue of the ductal carcinoma of the breast TMAs stained for 10 IHC markers, followed by factor analysis of the data set. In some sense, our results represent an automated readout of the IHC data in the TMAs. Factor analysis revealed latent factors governing the interdependent variance of the immunophenotype: being orthogonally independent by definition, the factors can be seen as independent biological processes standing behind the IHC profile variability in the disease entities. We then produced integral characteristics (factor scores) for individual patients and tested their associations with main conventional categories of the breast ductal carcinoma.
The factors of the immunophenotype variance, established in our study, are in line with the current knowledge of breast cancer biology, however, new insights emerge.
We found that major factor of the IHC profile variation in the ductal breast carcinoma was characterized by a strong inverse relation between the expression of hormone receptors (ER, PR, AR) along with anti-apoptotic marker BCL2, on one side, and Ki67 (proliferation) and HIF-1α (hypoxic stress, angiogenesis, see below), on the other side. While the corresponding correlations were detected by the pairwise correlation analysis (Table 2), they were only moderate, true interdependencies being obscured by multiple interactions in the dataset and difficult to interpret. We named the factor 1 the "i-Grade" since its pattern reflected the interdependent variance of the IHC markers known to represent the axis from aggressive (Ki67, HIF-1α) to more indolent (HR, BCL2) behaviour of the disease . In particular, biological meaning of this factor includes the axis of anti-apoptotic ← → proliferative effects that could be indicative of the variance of the tumour growth behaviour. Importantly, the anti-apoptotic effects were closely related to the expression of HR, while proliferative effects were paralleled by the marker of increased hypoxic stress and angiogenesis (HIF-1α). Also, our interpretation of the biological nature of the factor 1 (i-Grade) was further confirmed by pronounced bimodal distribution of the factor scores and strong associations with higher histological grade and the more aggressive intrinsic subtypes (HER2 positive and Triple-negative).
The pattern of the factor 1 (i-Grade) reveals important interactions between the HR, BCL2, Ki67, and HIF-1α. Our data support the notion that BCL2 may be a useful addition to the current scoring schemes as reported recently by Dawson et al. . Furthermore, a combined mitotic/BCL2 or Ki67/BCL2 index reflects true biological variation in breast cancer and may provide more relevant prognostic information [14, 20]. These latter observations were based on a combination of semi-quantitative scores of the two biomarkers; Ki67/BCL2 index was represented by subtraction of Ki67-BCL2 scores. As a matter of fact, factor 1 (i-Grade) scores showed very strong correlation (r = 0.89, p < 0.0001, data not shown) with the difference between Ki67 and BCL2 expression (the percentage of Ki67 positive cells - the percentage of BCL2 positive cells) in our data set. We therefore provide quantitative and multivariate analysis-based evidence supporting the suggested semi-quantitative and empirical definition of the Ki67/BCL2 index . In addition, our data suggest that inclusion of HIF-1α into this integrated index of the disease aggressiveness might bring more accuracy to this potential prognostic indicator.
Factor analysis, performed in the HR-positive subgroup of cases (LA, LB, and LB HER2-positive), extracted factor 3 (Ki67-ER) resembling the i-Grade by its associations to the histological grade and the intrinsic subtypes. The peculiarities of this "aggressiveness index" in the HR-positive subpopulation of breast cancer should be noted. First, the factor 3 (Ki67-ER) was no longer the main source of variability and did not present with bimodal distribution of the score values. This change can be explained by the decreased variation of Ki67 expression in the data set after exclusion of the most proliferative subtypes (Triple-negative and HER2 positive). Second, the pattern of the factor 3 suggests that Ki67/ER index rather than Ki67/BCL2 index might be a more accurate measure of the aggressiveness of the HR-positive disease. Or, at least, ER is tightly co-expressed with BCL2 and therefore sufficient to be used in combination with Ki67 in the HR-positive tumours. Although clinical utility of BCL2 as an independent prognostic marker has been suggested [19, 21, 22], this notion has to undergo scrutiny of large prospective trials and multivariate analyses. Third, although the Ki67-ER was strongly associated with the histological grade and intrinsic subtypes, it did attribute to the Ki67-ER-High category 16 and 21% cases of the Grade 1 and Luminal A tumours, respectively. This suggests that the combined index may provide an added value to the conventional categories securing against potential misinterpretations in, at least, borderline cases. It has been reported that Ki67 index can classify G2 breast cancer into low and high risk subgroups . It has been shown that image analysis of Ki67 correlates strongly with human evaluation, however, evaluation bias is possible and the results are potentially dependent on different Ki67 antibodies used . Therefore, improved stratification based on combinatorial index could be useful for better discrimination of Luminal A and B subtypes, in particular, and decision on chemotherapy. Interestingly, suggestions to develop specific sets of prognostic IHC biomarkers in lymph node-negative and positive breast cancer subgroups  also can be viewed as combinatorial approach, leading ultimately to the concept of multimodal metamarkers of the disease .
HER2 expression was independent of other IHC parameters in the patients with ductal carcinoma, however, in the HR-positive subset, HER2 "competed" mostly with PR but not ER expression (Figure 6). This finding is intriguing since the independent significance of ER and PR (as well as BCL2) is not entirely clear because of significant correlations between these biomarkers. Loss of PR expression might indicate somewhat worse prognosis compared to ER+/PR + tumours, whereas various combinations of BCL2, p53, HER2 expression might provide additional prognostic information as recently reviewed by Rakha et al. . In our study, we have found relevant pairwise correlations between ER, PR, and BCL2, however, factor analysis sheds the light into their true interdependencies: in the context of ductal breast carcinoma, we confirm that the HR and BCL2 expression is indeed highly inter-dependent and governed by the same latent factor of variation which is also characterized by inverse relation to Ki67 and HIF-1α. In the context of HR-positive tumours, high ER expression is seen in less proliferative (Ki67) cases, whereas low PR expression may indicate higher HER2 expression. Therefore, differential ER and PR expression may reflect differences in proliferative activity and HER2 expression of the HR-positive tumours; consequently, carrying related prognostic information (if the continuous increase of HER2 expression could be viewed as a potential feature of more aggressive behaviour). As noted already\, our analyses suggest that BCL2 correlates closely with the expression of hormone receptors (both in the whole group of patients and HR-positive tumours) and does not carry an independent information in the data sets.
Our study highlights the potential significance of relatively new and less-explored biomarkers in breast cancer. SATB1, a genome organizer that recruits chromatin-remodelling enzymes to regulate chromatin structure and gene expression, has been recently implicated to promote growth and metastasis of breast cancer and indicate poor prognosis [26, 27]. It was not confirmed by other studies and remains controversial as a prognostic factor and a potential target for therapy [28–31]. In particular, the expression levels of SATB1 mRNA in 2058 breast cancer samples were not related to disease-free survival among ER negative cancers, however, high SATB1 expression among ER positive tumours showed beneficial prognosis; nevertheless, even in ER positive cancer no independent prognostic value in multivariate analysis with standard parameters was observed . In the study of Patani et al., high SATB1 expression levels were more often found in ER negative tumour samples . Our study presents first evidence on correlation of IHC expression of SATB1 with ER and other markers. The factor 1 (SATB1/HIF-1α-AR/ER/BCL2) pattern in the HR-positive tumours revealed an inverse relation between the two groups of markers with the opposite factor loadings. Indeed, it is likely that the prognostic effects of SATB1 can be caused by a possible confounding effect of its inverse relation to ER expression . Furthermore, we did not find any associations of SATB1 expression with the histological grade or other categories tested. Nevertheless, it is remarkable that SATB1 was closely associated with HIF-1α in HR-positive and the whole cohort of ductal carcinoma; the factors with involvement of SATB1 and HIF-1α caused a major portion of variation in both groups, especially, in HR-positive tumours. This implies that SATB1 and HIF-1α may be important markers of the disease, whereas their biological and clinical significance remains to be elucidated.
HIF-1α is broadly expressed in many human cancers and frequently correlates with poor prognosis; it affects many key aspects of tumour initiation, progression, invasion, inflammatory cell recruitment and metastasis, and represents an attractive target for anti-cancer therapies as reviewed recently . Yamamoto et al.  reported on 171 cases of invasive breast cancer examined, nuclear HIF-1α expression was detected in 37% cases (a cut-off of 5% was used as in previous study ). HIF-1α was closely associated with indicators of aggressive phenotype, such as high histological grade, lymph node metastasis, large tumour size, high proliferation rate, negativity of hormone receptors, HER2 positivity and increased VEGF expression; elevated levels of HIF-1α expression were associated independently with shorter disease-free and overall survival ; hypoxia and HIF-1α might be related to the worse prognosis found in CD44 + CD24-/low positive breast tumors . Association of HIF-1α expression with unfavourable prognosis in patients with breast cancer has been demonstrated by previous studies [34, 36, 37]. In our study, we confirm the association of HIF-1α and Ki67 expression, both by pairwise correlation and similar factor loadings on the factor 1 (i-Grade) in the group of ductal carcinoma. Also, the i-Grade scores were strongly associated with histological grade, therefore, confirming aggressive nature of HIF-1α and Ki67 (confirmed also by one-way ANOVA using HIF-1α and Ki67 as dependent variables, data not shown). However, interpretation of the factor patterns in our analyses presents some peculiarities. First, in the group of ductal carcinoma, we find that HIF-1α participates in two factors: factor 1 (i-Grade) along with Ki67 and factor 2 - along with SATB1. In the group of ductal carcinoma, we find that HIF-1α participates in two factors: factor 1 (i-Grade) along with Ki67 and factor 2 - along with SATB1. Second, in the group of HR-positive ductal carcinoma, HIF-1α participates in the most important factor 1 along with SATB1 providing factor loadings opposite to those of AR, ER and BCL2, independently of Ki67 expression (factor 3). This further supports the notion that HIF-1α and SATB1 may convey important biological messages other than the aggressiveness of the disease reflected by Ki67 expression and histological grade, at least in HR-positive disease.
Expression of p16 was governed by an independent factor in our analyses and was remarkable for significantly higher levels in the TN subtype compared to all other intrinsic subtypes. We therefore support the reports [38–41] on increased p16 expression in basal/triple-negative breast cancer also suggesting frequent inactivation of the retinoblastoma tumour suppressor (Rb) and up regulation of the cyclin-dependent kinase inhibitor p16 in these tumours. Furthermore, down regulation of p16 expression has been observed in some basal-like breast cancer cell lines, suggesting that such cells can be divided into two groups according to Rb and p16 status, predictive of reduced chemo sensitivity in p16 depleted cancers . Interestingly, factor analysis performed in the subgroup of 17 patients with Triple-negative tumours (data not shown), revealed strong association of p16 expression with Ki67 and their strong inverse relation to AR (but not ER or PR) expression. AR expression has been reported as a marker of better prognosis in Triple-negative breast cancer [43–45], however, our findings warn that this effect may be caused by the confounding effect due to the inverse relation between AR and Ki67 with p16. This data awaits confirmation on a larger set of patients.
In our study we extracted 5 factors from the dataset of 10 IHC markers, arriving to clinically and biologically meaningful interpretation of the results. However, since the results of factor analysis may be influenced the number of factors extracted, defined by the investigator, we have also tested the robustness of our results by extracting 3 or 4 factors (not shown). The pattern of 4 factors extracted was largely the same, except p16 loadings moderately contributing to the i-Grade (-0.41) or Ki67-ER (0.49), in all or HR-positive tumours respectively. Similarly, extraction of 3 factors resulted in redistribution of p53 loadings whereas the interactions of the other markers remained essentially stable.
In summary, our study demonstrates that factor analysis of multiple IHC biomarkers measured by automated DA is an efficient exploratory tool clarifying complex interdependencies in the breast ductal carcinoma IHC profiles. We find that a major factor of the aggressive disease behaviour (i-Grade) is characterized by opposite loadings of ER/PR/AR/BCL2 and Ki67/HIF-1α. The i-Grade factor scores represent integral quantitative characteristics that reveal bimodal distribution and are strongly associated with the histological grade and relevant intrinsic subtypes. In HR-positive tumours, the aggressiveness of the tumour is best reflected by the combination of Ki67 and ER, rather than Ki67 and BCL2. These indices may provide additional risk stratifications for the currently used grading systems whereas their clinical utility and cut-offs remain to be established by analysis of clinical outcomes. Interestingly, we found inverse relation between HER2 and PR expression in the HR-positive tumours which, along with the inverse relation between Ki67 and ER, may shed the light into the differential information conveyed by the ER and PR expression. Remarkably, SATB1 along with HIF-1α reflected the second major factor of variation in our patients; furthermore, in the HR-positive group they were inversely associated with the HR and BCL2 expression and represented the major factor contributing to the variation in the IHC data set. However, this factor was not associated with the clinicopathologic categories studied. Biological meaning of this variation remains unclear: HIF-1α and SATB1 may convey important biological messages other than the aggressiveness of the disease reflected by Ki67 expression and histological grade. Meanwhile, we support the notion that the suggested prognostic significance of SATB1 may be related to its inverse relation to the ER expression. Finally, our analysis confirms high expression levels of p16 in Triple-negative tumours.