Robust inflammatory breast cancer gene signature using nonparametric random forest analysis

IBC is a rare form of breast cancer associated with poor prognosis compared to other subtypes, and this is attributed to its therapy resistance and a high metastatic potential [1‐3]. Moreover, the majority of IBC patients present with late-stage disease wherein the cancer has spread beyond the primary site [4]. To better diagnose and treat IBC patients, the IBC research community is working on defining an IBC-specific molecular signature. The largest study was published through the establishment of the World IBC Consortium which identified 79 genes, molecular subtype-independent, IBC signature [5]. Shortly after, another 132 genes, subtype-independent, IBC signature was reported [6]. However, both signatures were seen in ~ 16.4% and ~ 25% of breast cancer TCGA samples of primarily non-IBC patients, respectively, signifying low specificity in discriminating IBC from non-IBC samples [5, 7‐9]. Nevertheless, thus far a robust tumor cell-intrinsic signature that can define IBC from non-IBC or can stratify IBC patients has remained elusive [8, 9]. Indeed, a recent comparison of existing IBC signatures found minimal or no overlap among the proposed genes and none of the signatures could be validated in an independent dataset [9].

In this report, we reanalyzed publicly available gene expression datasets using the nonparametric machine learning random forest (RF) approach. RF is superior to classic statistical approaches used previously on these datasets because (1) It can handle many predictors at once while assigning each a predictor importance score. (2) It uses bootstrap-aggregated (bagged) decision trees to minimize overfitting, allowing for a robust model that can be validated in independent datasets. By restricting our analysis to microdissected IBC tumor epithelium and matching IBC samples with similar receptor-status to non-IBC samples, we have identified an IBC signature of 59 genes that only misclassified one patient out of a total 106 patients in pre-treatment datasets.

We have identified a robust gene signature that can characterize IBC from non-IBC with an aim to better understand and potentially develop a tailored treatment regimen for IBC patients. G59 is the first IBC signature to be successfully validated in an independent dataset and shows the highest accuracy (100% (45/45) in GSE45581 and (60/61) 98.4% in GSE111477) in its prediction [9]. This is a significant improvement in accuracy as previous signatures accuracy range between 68 and 88% [5, 8, 9], a range similar to our analysis [Fig. 2c(ii)]. Importantly, G59 shows higher specificity in primarily non-IBC TCGA samples compared to previous signatures [5, 7‐9].

The low prediction accuracy in primarily post-treatment tumor samples highlights the fact that chemotherapy induces changes in gene expression [12]. Interestingly, SUM149 and SUM190, the two cell lines used in most of the IBC research [13], were derived from patients who had already received chemotherapy treatment [14]. Our analysis suggests the need for establishing IBC cell lines from untreated patients to fully capture IBC-specific profile.

G59 is a more curated version of the 132 gene list selected by Dr. Woodward [6] for IBC assessment with 49% similarities. Most of the genes in G59 code for membrane proteins, suggesting that IBC cells are highly communicative with the tumor microenvironment, likely playing an essential role in directing their disease progression. The novel implication of IL-2 inflammatory as well as chemokine pathways in IBC (Fig. 2f right) adds to the proposed inflammatory pathways involvement [8, 15].

Our finding highlights the need to integrate contemporary statistical approaches to identify molecular signatures previously missed by traditional statistical methods. Most important, the IBC-specific molecular signature we have identified paves the way for IBC functional studies, validation, and potentially successful therapeutic interventions.