Protein Signature Predicts Response to Neoadjuvant Treatment With Chemotherapy and Bevacizumab in HER2-Negative Breast Cancers

Abstract

Purpose: Antiangiogenic therapy using bevacizumab has proven effective for a number of cancers; however, in breast cancer (BC), there is an unmet need to identify patients who benefit from such treatment.

Patients and methods: In the NeoAva phase II clinical trial, patients (N = 132) with large (≥ 25 mm) human epidermal growth factor receptor 2 (HER2)-negative primary tumors were randomly assigned 1:1 to treatment with neoadjuvant chemotherapy (CTx) alone or in combination with bevacizumab (Bev plus CTx). The ratio of the tumor size after relative to before treatment was calculated into a continuous response scale. Tumor biopsies taken prior to neoadjuvant treatment were analyzed by reverse-phase protein arrays (RPPA) for expression levels of 210 BC-relevant (phospho-) proteins. Lasso regression was used to derive a predictor of tumor shrinkage from the expression of selected proteins prior to treatment.

Results: We identified a nine-protein signature score named vascular endothelial growth factor inhibition response predictor (ViRP) for use in the Bev plus CTx treatment arm able to predict with accuracy pathologic complete response (pCR) (area under the curve [AUC] = 0.85; 95% CI, 0.74 to 0.97) and low residual cancer burden (RCB 0/I) (AUC = 0.80; 95% CI, 0.68 to 0.93). The ViRP score was significantly lower in patients with pCR (P < .001) and in patients with low RCB (P < .001). The ViRP score was internally validated on mRNA data and the resultant surrogate mRNA ViRP score significantly separated the pCR patients (P = .016). Similarly, the mRNA ViRP score was validated (P < .001) in an independent phase II clinical trial (PROMIX).

Conclusion: Our ViRP score, integrating the expression of nine proteins and validated on mRNA data both internally and in an independent clinical trial, may be used to increase the likelihood of benefit from treatment with bevacizumab combined with chemotherapy in patients with HER2-negative BC.

FIG 1.

ViRP score based on protein or mRNA in relation to response. (A) ViRP score for each patient with corresponding relative tumor size after end of treatment. (B) ViRP scores in pCR and non-pCR patients. (C) ViRP scores in RCB low (0 and I) and high (II and III) patients. (D) Correlation between protein and mRNA ViRP scores. (E) mRNA ViRP scores in pCR and non-pCR patients in the NeoAva cohort. (F) mRNA ViRP scores in pCR and non-pCR patients in the PROMIX cohort. pCR, pathologic complete response; RCB, residual cancer burden; ViRP, VEGF inhibition response predictor.

FIG 2.

Predictive performance of the ViRP score. (A) ROC curve for ViRP score prediction of pCR. (B) ViRP score prediction of probability for pCR (green line) with 95% CI (blue contour). True response by pCR (red dots) marked on bottom and non-pCR (blue dots) marked on top. (C) ROC curve for ViRP score prediction of low RCB (0 and I). (D) ViRP score prediction of probability for low RCB (0 and I) (green line) with 95% CI (blue contour). True response by low RCB (red dots) marked on bottom and high RCB (blue dots) marked on top. AUC, area under the curve; pCR, pathologic complete response; RCB, residual cancer burden; ROC, receiver operating characteristic; ViRP, VEGF inhibition response predictor.

FIG 3.

Response rates by selection of patients using the ViRP score. (A) Fraction of patients obtaining pCR when treated with standard CTx in all NeoAva patients (N = 55) or Bev plus CTx in all NeoAva patients (N = 54) and selected based on ViRP score (N = 22). (B) Fraction of patients obtaining low RCB (0 and I) when treated with standard CTx in all NeoAva patients (N = 55) or Bev plus CTx in all NeoAva patients (N = 54) and selected based on ViRP score (N = 22). pCR, pathologic complete response; RCB, residual cancer burden; ViRP, VEGF inhibition response predictor.

FIG A1.

CONSORT diagram.

FIG A2.

Responders by pCR in study population (N = 109). pCR, pathologic complete response.

FIG A3.

Mixed distribution of protein variance. Mixed distribution of protein expression variances (solid green and blue lines); only proteins with variance above intersection (red line) were selected for input in Lasso regression to determine the ViRP (Appendix Table A3). ViRP, VEGF inhibition response predictor.

FIG A4.

Cluster heatmap of protein expression prior to treatment. Hierarchical clustering (Euclidean distance and complete linkage) of protein expression (red = high to blue = low). Top bars recapitulate relative tumor size, addition of bevacizumab to CTx, and PAM50 classification of tumors. Hierarchical clustering was performed using the R-package Clustermap (https://github.com/cbsteen/clustermap) with median centered and log²-transformed RPPA data, and visualized using values normalized to a selected range [− x₀, x₀] by application of the transformation f(x) = x₀tanh(x/x₀). The number of clusters were determined using the Partitioning Algorithm by Recursive Thresholding (PART) with N = 10,000 permutations (Nilsen et al, Statistical Applications in Genetics and Molecular Biology, 2013).

FIG A5.

Correlation of protein expression and relative tumor size after treatment. Rho-values derived from the Spearman correlation of protein expression to relative tumor size for patients receiving the Bev plus CTx combination (x-axis) and CTx treatment only (y-axis) with first principal component of relation between the treatment arms (stitched red line). Venn diagram (upper left corner) of proteins in each treatment arm significantly associated with relative tumor size.

FIG A6.

ROC curves for prediction of pCR with mRNA ViRP scores in the NeoAva and PROMIX trials. (A) ROC curve for mRNA ViRP score prediction of pCR in NeoAva. (B) ROC curve mRNA ViRP score prediction of pCR in PROMIX. AUC, area under the curve; pCR, pathologic complete response; ROC, receiver operating characteristic; ViRP, VEGF inhibition response predictor.

FIG A7.

In silico predicted downstream biological effects of bevacizumab treatment mediated through the ViRP proteins. In silico simulation with Ingenuity Pathway Analysis (version 47547484, Qiagen) was used to predict biologic effects (lower row) affected by the ViRP member proteins (second bottom row) when influenced by bevacizumab treatment (top row) through a set of intermediate proteins (second top row). Red or orange color predicts activation, blue color indicates inhibition, and yellow color indicates findings inconsistent with the state of downstream molecule. ViRP, VEGF inhibition response predictor.

FIG A8.

Relative importance of proteins in ViRP signature. Relative contribution of each protein in the ViRP signature for prediction of pCR. pCR, pathologic complete response; ViRP, VEGF inhibition response predictor.

FIG A9.

Correlation between protein and mRNA expression. Rho-Spearman correlation of all individual protein-mRNA in the data set (N = 210) with ViRP proteins highlighted in red. ViRP, VEGF inhibition response predictor.

FIG A10.

ViRP scores in the CTx treatment arm. (A) ViRP score for each patient with corresponding relative tumor size. (B) ViRP scores in pCR and non-pCR patients. (C) ViRP scores in RCB low (0 and I) and high (II and III) patients. (D) Fraction of patients obtaining pCR when treated with standard CTx in all NeoAva patients (N = 55) or selected based on ViRP score (N = 17). (E) Fraction of patients obtaining low RCB (0 and I) when treated with standard CTx in all NeoAva patients (N = 55) or selected based on ViRP score (N = 17). pCR, pathologic complete response; RCB, residual cancer burden; ViRP, VEGF inhibition response predictor.

FIG A11.

ViRP score in relation to tumor cell content. (A) ViRP score for each patient with corresponding tumor cell content. (B) Tumor cell content in patients with low or high ViRP score. ViRP, VEGF inhibition response predictor.