Integrated Analysis of RNA and DNA from the Phase III Trial CALGB 40601 Identifies Predictors of Response to Trastuzumab-Based Neoadjuvant Chemotherapy in HER2-Positive Breast Cancer

Abstract

Purpose: Response to a complex trastuzumab-based regimen is affected by multiple features of the tumor and its microenvironment. Developing a predictive algorithm is key to optimizing HER2-targeting therapy.Experimental Design: We analyzed 137 pretreatment tumors with mRNA-seq and DNA exome sequencing from CALGB 40601, a neoadjuvant phase III trial of paclitaxel plus trastuzumab with or without lapatinib in stage II to III HER2-positive breast cancer. We adopted an Elastic Net regularized regression approach that controls for covarying features within high-dimensional data. First, we applied 517 known gene expression signatures to develop an Elastic Net model to predict pCR, which we validated on 143 samples from four independent trials. Next, we performed integrative analyses incorporating clinicopathologic information with somatic mutation status, DNA copy number alterations (CNA), and gene signatures.Results: The Elastic Net model using only gene signatures predicted pCR in the validation sets (AUC = 0.76). Integrative analyses showed that models containing gene signatures, clinical features, and DNA information were better pCR predictors than models containing a single data type. Frequently selected variables from the multiplatform models included amplifications of chromosome 6p, TP53 mutation, HER2-enriched subtype, and immune signatures. Variables predicting resistance included Luminal/ER⁺ features.Conclusions: Models using RNA only, as well as integrated RNA and DNA models, can predict pCR with improved accuracy over clinical variables. Somatic DNA alterations (mutation, CNAs), tumor molecular subtype (HER2E, Luminal), and the microenvironment (immune cells) were independent predictors of response to trastuzumab and paclitaxel-based regimens. This highlights the complexity of predicting response in HER2-positive breast cancer. Clin Cancer Res; 24(21); 5292-304. ©2018 AACR.

Figure1.. CONSORT diagram of patient selection and characteristics.

(A) Sample flow chart to show how samples were selected. Starting with 305 patients, specimens were removed for multiple reasons including incomplete clinical data, low RNA yields, a normal-like non-tumor expression profile, being part of the TL= lapatinib and paclitaxel arm, thus leaving 203 patients. Of these, 137 had DNA exomes results, with this final 137 sample set also being split into a training and test set. (B) Clinical and intrinsic expression subtype characteristics with pCR rates using the 137 patient data set. P-values were calculated by Chi-aquare test. TH, trastuzumab and paclitaxel arm; THL, trastuzumab, lapatinib and paclitaxel arm; ER, estrogen-receptor; PgR, progesterone receptor.

Figure 2.. Performance of the Elastic Net model for pCR prediction using gene signatures on CALGB 40601.

(A) Area under the curve (AUC) from the Receiver operating characteristic curve analysis were estimated for Elastic Net models using gene signatures alone in CALGB 40601. Left, CALGB 40601 as the training set (N = 137), AUC = 0.80; Right, All test sets combined (CHERLOB + XENA + I-SPY + CALGB 40601 validation set, N= 143, AUC = 0.76). Sensitivity and specificity values were selected using Youden’s cutpoint where the sum of sensitivity and specificity is maximal. Mann-Whitney-Wilcoxon test was conduct to calculate p-values. (B) Barplots showing results of the Elastic Net model score split into three rank order groups and then comparing pCR rates for patients in CALGB 40601, or all test sets combined. ANOVA T-test was conducted to calculate p-values by comparing signature scores across all three groups.

Figure 3.. Elastic Net analysis using multi-dimensional data.

(A) Average AUC scores for various individual data type, or combined data type predictors, using test sets through 10 repeated Elastic Net analyses. Each bar shows the average AUC scores with 95% confidence intervals. (B) Frequently selected Elastic Net features coming from a multi-dimensional predictor. Features contributing to at least 6 out of 10 Elastic Net models using gene signatures, CNAs, mutations, and clinical ER/PgR status. GS, gene signature; CN, copy number; Mut, mutation; Gray and black bars indicate predictors which positively (37) and negatively (53) predict pCR; thus gray predictors are high in pCR samples and black predictors are high in non-pCR samples. Yellow arrows indicate CNAs features at Chromosome 6p; Green arrows indicate TP53 mutation status or signatures; Pink arrows indicate HER2-enriched signatures; A gray arrow indicates 21-gene Recurrence Score; Black arrows indicate immune signatures; Blue arrows indicate Clinical ER status, Luminal signatures and PgR gene signature.

Figure 4.. Hierarchical clustering of multi-dimensional features associated with pCR.

Supervised clustering of the 33 selected features among 137 samples. The features were grouped into two clusters with positive or negative predictors. The samples from left to right were ordered by their average scores derived from the 10 Elastic Net models grouped into high, middle, and low scores.

Figure 5.. Identification of DNA copy number alterations as biomarkers of trastuzumab-paclitaxel resistance and sensitivity.

(A) DNA copy number frequency landscape plots for pCR vs non-pCR tumors. The frequency of alterations in each group is indicated on the y-axis from 0 to 100 %. Segments of group-specific copy number gains or loss are plotted above or below the x-axis, respectively. Significantly different regions between pCR vs. non-pCR (t-test p<0.05 after Benjamini and Hochberg correction) are highlighted in red (gain) or in green (loss). (B) A Venn diagram comparing three types of gene-level copy number results. Genes in copy number segments contributing to ≥6 models out of the 10 Elastic Net testing, top 1% copy number genes from the Dawnrank analysis, and copy number genes with false discovery rate ≤ 1% from SAM analysis were plotted and identify MAPK14 and CDKN1A as possible driver genes for trastuzumab-paclitaxel sensitivity.