Robust and interpretable PAM50 reclassification exhibits survival advantage for myoepithelial and immune phenotypes

Abstract

We introduce a classification of breast tumors into seven classes which are more clearly defined by interpretable mRNA signatures along the PAM50 gene set than the five traditional PAM50 intrinsic subtypes. Each intrinsic subtype is partially concordant with one of our classes, and the two additional classes correspond to division of the classes concordant with the Luminal B and the Normal intrinsic subtypes along expression of the Her2 gene group. Our Normal class shows similarity with the myoepithelial mammary cell phenotype, including TP63 expression (specificity: 80.8% and sensitivity: 82.8%), and exhibits the best overall survival (89.6% at 5 years). Though Luminal A tumors are traditionally considered the least aggressive, our analysis shows that only the Luminal A tumors which are now classified as myoepithelial have this phenotype, while tumors in our luminal class (concordant with Luminal A) may be more aggressive than previously thought. We also find that patients with basal tumors surviving to 48 months exhibit favorable continued survival rates when certain markers for B lymphocytes are present and poor survival rates when they are absent, which is consistent with recent findings.

Keywords: Breast cancer; Prognostic markers.

Fig. 1

History of the molecular classification of breast cancer. Names are shown at the chronological level at which they were introduced. The Her2+ breast tumors were already well-known in the 1990s for highly favorable response to the drug trastuzumab (herceptin), which was approved by the FDA for metastatic Her2+ breast cancer in 1998. The hierarchical clustering of Perou et al. used genes whose expression differentiates between samples from different tumors better than between samples from the same tumor, finding four main classes: ERBB2+ (or Her2+), Basal, Luminal, and Normal breast-like. Sorlie et al. explicitly incorporated clinically relevant outcome data such as overall survival, uncovering three Luminal subtypes, Luminal A, B, and C. Luminal A has higher overall survival than Luminal B, and Luminal B has higher overall survival than Luminal C. Later investigators found only two Luminal subtypes to be sufficiently robust. Parker et al. introduced the 50 gene set that became known as the PAM50 (Prediction Analysis of Microarray) and introduced a straightforward centroid-based classifier for breast tumor RNA expression patterns along the PAM50 with five classes: Basal, Her2, Luminal A, Luminal B, and Normal. The authors used this classification as a key component in the model that became the Prosigna predictor of Risk Of Relapse (ROR). Prat and Perou introduced the Claudin-low subtype carved largely out of the Basal group. The authors find that the Claudin-low subtype has poor prognosis compared to Luminal A, but no worse than the other subtypes. The Topological Data Analysis of Nicolau et al. confirmed the distinction between more luminal, more basal, and more normal-like subtypes along branches of a graph structure modeling the distribution of breast tumor samples. They found a subgroup of patients exhibiting a very high survival rate, largely characterized by expression of MYB. Our proposed classification uses the method of Nicolau et al. and incorporates gene sets and priors (e.g., the basal-to-luminal stratification) known to be relevant to breast cancer biology. (Below right) Our proposed system with seven classes defined by four elementary phenotypes (see also Figs 2, 7)

Fig. 2

RNA expression heatmap of the 1904 METABRIC breast tumor samples. (Above) Organized first by PAM50 subtype and then by the TDA signatures classes assigned by the Mapper-derived classifier along the PAM50 gene set (BAG1, MYBL2, GPR160, and TMEM45B omitted due to missing values). (Below) Organized first by TDA signature class then by PAM50 subtype

Fig. 3

Kaplan–Meier survival analysis of the subgroups of the TCGA and METABRIC cohorts defined by PAM50 subtypes and the major corresponding TDA signature classes. The Myo/Luminal class has the highest survival rate, statistically significantly greater than the primary corresponding PAM50 subtype, the Normal subtype. In the TCGA data set, the log-rank test for PAM50 Normal versus Myo/Luminal yields p = 0.023, while in the METABRIC data set (with approximately twice as many samples) the test yields p = 0.003

Fig. 4

Stratification of the Myo/Luminal class by three TDA signature classes. The Kaplan–Meier plot shows slightly different survival rates, with Myo/Luminal A having the best prognosis; better than PAM50 Luminal A. TP63 expression (a known myoepithelial marker; see Fig. 5) somewhat robustly defines the Myo/Luminal class. Kaplan–Meier survival analysis plots are shown comparing the survival probabilities between TP63+ and TP63− phenotypes across the whole METABRIC cohort. TP63+ confers a survival advantage comparable to that of PGR+

Fig. 5

Mapper analysis of the 290 GTEx normal mammary tissue samples using the basal–luminal score as filter function. (Above) Along the PAM50 gene set, using the basal–luminal score as filter function. (Below) The same sample set, in the same order, showing the expression of the marker genes of Santagata et al., which define the normal mammary cell-type classification proposed by those authors. A substantial group displays the Myo/Luminal/Her2 phenotype. According to the Santagata et al. classification, these samples are primarily a combination of the myoepithelial type M2 (TP63+/KRT5+) and the luminal–epithelial type L7 (VDR+/KRT5+)

Fig. 6

Survival analysis of FOXC1+/MIA+/PHGDH+ and CD79A+/CD38+/IGLL1+ phenotypes. (Above) The FOXC1+/MIA+/PHGDH+ phenotype, observed in the Myo/Luminal B class but not the Myo/Luminal A class, confers a survival disadvantage for approximately the first 48 months after diagnosis, and a survival advantage afterwards. (Below) Of the top 100 genes out of 18,543 exhibiting statistically significant mean differences between the FOXC1+/MIA+/PHGDH+ short-term and long-term survivors, several are B-lymphocyte-related, including: CD79A (immunoglobulin-alpha), CD38, and IGLL1 (immunoglobulin lambda-like polypeptide 1). FOXC1+/MIA+/PHGDH+ is also observed in the PAM50 Basal subtype. Within the Basal subtype, CD79A+, CD38+, and IGLL1+ confer a significant survival advantage after 48 months

Fig. 7

Mapper analysis of the 1904 METABRIC breast tumor samples, along the PAM50 gene set, using the basal–luminal score as filter function. The circular nodes represent clusters in the strata or bins defined by the filter function at the chosen level of resolution. For example, there are three clusters in the stratum shown in yellow; two clusters shown higher and labeled with unsupervised signature number [4], and one cluster shown lower labeled with unsupervised signature number [5]. All three have the same basal–luminal score range, indicated by color. In Table 1, the salient signatures are recorded. These signatures differ slightly, in two ways, from the seven classes we finally propose as in Fig. 1. First, for the sake of simplicity we merge the two myoepithelial–related gene groups (b, c) into a single-gene group, consequently merging Myo/Luminal A and Myo/Luminal B into Myo/Luminal. Second, on account of the salient signatures observed in the heatmaps in Fig. 2, we split Her2/Basal [5] into Her2/Basal and Luminal/Basal/Her2. Where blanks appear, the corresponding gene group (a–e) is neither uniformly positively nor uniformly negatively expressed