A pathway-based classification of human breast cancer

Abstract

The hallmark of human cancer is heterogeneity, reflecting the complexity and variability of the vast array of somatic mutations acquired during oncogenesis. An ability to dissect this heterogeneity, to identify subgroups that represent common mechanisms of disease, will be critical to understanding the complexities of genetic alterations and to provide a framework to develop rational therapeutic strategies. Here, we describe a classification scheme for human breast cancer making use of patterns of pathway activity to build on previous subtype characterizations using intrinsic gene expression signatures, to provide a functional interpretation of the gene expression data that can be linked to therapeutic options. We show that the identified subgroups provide a robust mechanism for classifying independent samples, identifying tumors that share patterns of pathway activity and exhibit similar clinical and biological properties, including distinct patterns of chromosomal alterations that were not evident in the heterogeneous total population of tumors. We propose that this classification scheme provides a basis for understanding the complex mechanisms of oncogenesis that give rise to these tumors and to identify rational opportunities for combination therapies.

Fig. 1.

Meta-analysis of breast cancer gene expression patterns. (A) A breast cancer dataset comprise of 1,143 samples derived from 10 independently generated datasets was clustered by complete linkage hierarchical clustering based on the gene expression patterns of Affymetrix U133A probes. The intrinsic subtype of each sample is reported. (B) The number of clusters identified in random subsets of the 1,143 samples demonstrates saturation of the complexity of expression patterns as a function of dataset size. (C) Analysis of the number of clusters identified in random subsets of tumors with known ER status (n = 828) compared to the number of clusters identified in subsets of ER+ tumors (n = 596).

Fig. 2.

Patterns of pathway activity that characterize breast cancer. (A) Heat map depicting the two-way hierarchical clustering of the predicted probability of 1,143 breast tumor samples and 18 pathways. Low (blue) and high (red) pathway activity and predicted probabilities are shown. (B) Heat map depicting the correlation coefficient of pathway coregulation (red indicates a positive correlation; blue, a negative correlation).

Fig. 3.

Identification of breast tumor subtypes using patterns of pathway activity. (A) Scheme for the development of pathway-derived breast tumor subgroups. (B) Predicted probability of subgroup membership for 1,143 breast tumor samples where each row represents a sample; each column, a subgroup (samples are organized by subgroup). (C) Heat map depicting patterns of pathway activity in the 17 identified breast tumor subgroups organized by the relationship with intrinsic subtypes. Red indicates a high predicted probability, blue a low probability. Overall survival differences between pathway-derived subgroups classified as (D) basal-like (P = 0.0039) and (E) luminal A-dominant (P = 0.0046) were analyzed by a Kaplan-Meier survival curve and demonstrate a statistically significant difference in survival (log-rank test).

Fig. 4.

Prediction of subgroup membership. (A) Breast tumors in the validation dataset (n = 547) were classified into 17 pathway-derived subgroups and the probability of subgroup assignment was plotted for each sample (red indicates a high probability of subtype membership; blue, low probability). (B). 50 breast cancer cell lines were classified into 13 of 17 pathway-derived subgroups on the basis of patterns of pathway activity, and the predicted probability of subgroup membership is shown; 36 of 50 (72%) samples had a predicted probability of subgroup membership greater than 0.80.

Fig. 5.

Pathway-defined breast tumor subgroups exhibit unique patterns of DNA copy number changes. Patterns of DNA copy number changes were calculated for each subgroup. (A) The percent of samples in each of the 16 subgroups with identified copy number gains and losses are shown. Green indicates a region of amplification and red indicates a region of chromosomal loss; dark green and red indicate the percentage of samples with high copy number gains or homozygous deletion, respectively. Chromosomal borders are delineated by alternating gray and white regions. (B) Increasingly homogeneous patterns of copy number losses are evident in pathway-derived subgroups as compared to all breast tumors for subgroup 5 at 3p14.3 (P = 0.0009, unpaired t test), subgroup 7 at 4p15.1 (P = 0.0106, unpaired t test), and subgroup 6 at 11q21-24 (P = 0.0093, unpaired t test). (C) Increasingly homogeneous patterns of copy number gains are evident in breast tumor subgroups compared to all other samples. Subgroup 5 shows a amplification at 3q25.1 (P = 0.0211, unpaired t test) and subgroup 11 shows an amplification at 20p12-13 (P < 0.0001, unpaired t test). (D) Basal-like subgroups 2, 5, and 8 show copy number gains at 8q24 (P = 0.4575, ANOVA); only subgroup 5 shows copy number losses at 3p14 (P < 0.0001, ANOVA).