A Biobank of Breast Cancer Explants with Preserved Intra-tumor Heterogeneity to Screen Anticancer Compounds

Abstract

The inter- and intra-tumor heterogeneity of breast cancer needs to be adequately captured in pre-clinical models. We have created a large collection of breast cancer patient-derived tumor xenografts (PDTXs), in which the morphological and molecular characteristics of the originating tumor are preserved through passaging in the mouse. An integrated platform combining in vivo maintenance of these PDTXs along with short-term cultures of PDTX-derived tumor cells (PDTCs) was optimized. Remarkably, the intra-tumor genomic clonal architecture present in the originating breast cancers was mostly preserved upon serial passaging in xenografts and in short-term cultured PDTCs. We assessed drug responses in PDTCs on a high-throughput platform and validated several ex vivo responses in vivo. The biobank represents a powerful resource for pre-clinical breast cancer pharmacogenomic studies (http://caldaslab.cruk.cam.ac.uk/bcape), including identification of biomarkers of response or resistance.

Graphical abstract

Figure 1

Derivation of an Extensively Annotated Breast Cancer PDTX-PDTC Biobank Representing Breast Cancer Subtypes (A) Timeline of engraftment for established PDTX models (n = 31; ER+ in red; ER− in blue). Each square represents a time point of engraftment. Average ER+ and ER− re-implantation time is shown on furthermost right panel. Model IDs are color coded according to integrative cluster (IntClust). (B) Bar plots showing the IntClust distribution of PDTX models (n = 40; shadowed) and for comparison primary breast cancers from METABRIC (n = 1,980; dense). (C) Distribution of somatic mutations in tumors from the TCGA cohort (n = 495) and PDTX models (n = 30), stratified by ER status. See also Figure S1.

Figure 2

PDTXs Closely Match Originating Patient Cancer Samples (A) Heatmap of Pearson correlation scores across molecular data types (different sample sizes described in the main text). (B) Panels with individual examples for five types of molecular data. (Left panel: top) CNA plots for AB551 (originating sample [T], PDTX, and PDTC) are shown; (bottom) scatterplot of methylated CpGs (from RRBS data) in AB521M is shown. (Right panel: top) Scatterplots of pathway activity scores in AB521M are shown. (Middle) Scatterplots of variant allelic fractions in STG139 are shown. (Bottom) Mutational profiles in AB551 are shown.

Figure 3

Clonal Architecture and Clonal Dynamics of Breast Cancer PDTXs (A) Example plots of AB551 (left panel), HCI002 (middle panel), and STG282 (right panel). (Left graph) The mean cellular prevalence estimates of mutation clusters in originating patient samples (T) and subsequent xenograft passages (Xn; n for passage number) or PDTCs (XnCy; y for days in culture) are shown. PyClone was used to infer clusters and cellular prevalence using WES data. Line widths indicate the number of SNVs comprising each mutation cluster (numbers in brackets adjacent to each plot). Asterisks indicate clonal clusters with significant changes in cellular prevalence. (Right graph) Plots of distribution of variant allele frequency for selected genes within clusters. (B) PyClone plots (as in A) and cellular prevalence heatmap plots for STG139 and AB521 samples. (C) PyClone plot and plots of distribution of variant allele frequency for selected genes within clusters (as in A) of five spatially separated biopsies and their matched xenografts in CAMBMT1. See also Figure S3.

Figure 4

High-Throughput Drug Screening Using PDTCs (A) AUCs scatterplots showing reproducibility of PDTC drug testing. (Left panel) AUCs of technical replicates (n = 6,325; same sample, same compound) are shown. (Right plot) AUCs of biological replicates (n = 1,341; same model, different passages, same compound) are shown. r, Pearson correlation. (B) AUC scatterplots of all drugs targeting PI3K/AKT/mTOR pathway (n = 34 passages from 20 models). Red indicates Pearson correlation > 0.5. (C) AUC scatterplot for cisplatin and BMN-673 treatment across models tested (n = 15). (D) Illustration of the PI3K pathway with panels depicting difference in the AUC in models (n = 15) with versus without molecular alteration in pathway member. (Left panels) Inhibitors of PI3K alpha and PI3Kbeta are shown. (Right panels) Inhibitors of AKT1 and mTOR are shown. See also Figures S4, S5, and S6.

Figure 5

Drug-Drug Combination Studies in PDTCs (A) Synergism of paclitaxel in combination with 17-AAG. (Top panel) Bliss independence model residuals for paclitaxel combinations are shown. The 95% percentile of these differences (in percentage) is plotted. For each drug combination, the expected response is compared to the observed response in all the dose ranges in the combination. (Middle panel) Boxplots of distribution of residuals (Bliss independence model) for paclitaxel and 17-AGG combination in each PDTC model tested are shown. (Bottom panel) Detailed analysis for HCI008 (from top to bottom: single drug curves, bivariate isotonic fit for the combination, and residuals of the Bliss model for each dose combination) is shown. Red shades, synergistic effects; blue shades, antagonistic effects. (B) Synergism of IGF-1R/IR inhibitor (BMS-754807) with PI3K/mTOR inhibitor (NVP-BEZ235). Panels are the same as in A (bottom panel: detailed analysis for STG201).

Figure 6

Validation of Ex Vivo PDTC Drug Responses with In Vivo PDTX Testing Representative sensitive (gray panel) and resistant (pink panel) drug responses in several models. (Left plots) PDTC ex vivo dose response is shown. (Right plots) PDTX in vivo tumor growth curves are shown (sample sizes are indicated in the plot; average values and error bars representing SDs are shown). See also Figure S7.

Figure S1

Related to Figure 1 and STAR Methods (A) Copy number profiles based on the cis-features that define the 10 Integrative Clusters (Curtis et al., 2012) in the METABRIC discovery dataset (n = 997 breast tumors) and the PDTX biobank (n = 121 samples from 36 models). (B) Left panel: Goodness of fit scores for copy number-based classification into IntClust subtypes (Metabric cohort n = 905 tumors, PDTX cohort n = 87 samples from 36 models). Right panel: Kaplan-Meier (disease-specific survival) of IntClust subtypes for the METABRIC cohort. IntClust4 was further stratified into ER-+ and ER-. (C) Hierarchical clustering (Ward’s method) of cancer pathway activation scores across samples. Included are 15 randomly picked samples belonging to each of the 10 IntClusts from the METABRIC cohort and 19 PDTX models (plus 5 technical replicates). (D) Scatter plot of reproducibility and prediction accuracy of cancer pathway activation scores in PDTXs (n = 78 tumor/PDTX pairs from 16 models). Reproducibility was assessed using Spearman correlation of cancer pathway scores in matched tumors and PDTXs. Prediction accuracy was determined by fitting a generalized additive model and computing the deviance explained.

Figure S2

Related to STAR Methods and Figures 2A and 2B (A) Representative images from the histopathological analysis of PDTXs. Images from both an ER+ (STG335) and a triple negative (STG139) model at different passages are shown. (B) Calibration curve for estimation of mouse content based on the proportion of mapped reads from WES to the human genome. (C) Estimated percentage of mouse contamination from WES data. Top panel: data for each model at different passages is represented. Bottom panel: Box plots showing the distribution of percentage of human cells per model in all samples tested, including different mice from the same passage and the same model, and different passages of the same model (n = 94 samples from 29 models; estimates from the same model and passage have been averaged). (D) Representative FACS plot of single cells (left) and a FISH image (right) on an FFPE tissue section from an example PDTX sample. Table shows average percentage of mouse cells in different PDTX models tested by FACS.

Figure S3

Related to Figure 3 and STAR Methods (A) Box plots of MATH scores for each model analyzed (n = 238 samples from 39 models). Each box plot represents the distribution of scores for matched tumor (red T), PDTXs and PDTCs. (B) Plots of average change in clonal cluster prevalence. PyClone was used to infer clonal architecture for the set of samples from each PDTX model. For all PDTX models with more than two samples, the absolute change in clonal cluster prevalence was averaged over all clusters. Left plot: average change in clonal cluster prevalence with short-term culture of PDTX cells (n = 5 comparisons from 3 models). Middle plot: average change in clonal cluster prevalence with serial passaging (n = 38 passages from 12 models). Right plot: average change in clonal cluster prevalence with implantation (originating sample versus earliest PDTX passage; n = 24 pairs from 16 models). Dot size is proportional to number of samples analyzed. (C) PyClone individual cluster plots showing clonal mean cellular prevalences for 22 models. Width lines are proportional to the number of variants in each clonal cluster. The legend indicates the name of the cluster and the number of variants in it. Asterisks remark clusters whose cellular frequencies are significantly different between samples. Only clusters with at least one variant are shown in the plot.

Figure S4

Related to Figure 4 and STAR Methods (A) Representative microscopy images from 8 PDTX models of short-term cultures (PDTCs) at day 7 after plating. (B) Box plots showing the percentage of human DNA on PDTX (n = 94 samples) and PDTC (n = 15 samples) models (n = 29). Data estimated from our sequencing-based approach. (C) Changes in cell number and viability of PDTCs at each time-point. (D) Representative FACS image from PKH26 assay. Table showing quantification of PKH26 low cells in 2 PDTCs (STG195 and STG201) and a highly proliferative breast cancer cell line (MDAMB231), shown for comparison purposes, at different time-points. (E) Scatterplots comparing AUC values measuring drug response for 19 drugs in 8 models (AB521, AB555, AB582, STG195, STG316, STG321, STG335 and VHIO093) using 3 different viability assays. 10 different doses were tested. Curve fitting and computation of the AUC was done as described in the STAR Methods. Dots highlighted in color correspond to PI3K pathway inhibitors: GDC032 (PI3Kα) in red, GDC0941 (pan PI3K) in blue, AZD6482 (PI3Kβ) in purple, and AZD8055 (mTOR) in green.

Figure S5

Related to Figure 4 and STAR Methods (A) Drug responses classified into 8 different groups according to response curves. Left panel- the drug responses clustered into 8 groups. Right panel- plots with percentage of samples (n = 37 samples from 20 models) displaying each drug response pattern for each compound. Drug name (bold) and putative target indicated. (B) Unsupervised clustering of drug responses in tested models (n = 37 samples from 20 models) according to subtype of drug response pattern (color coded as in A). Vertical axis- drugs; horizontal axis- models.

Figure S6

Related to Figure 4 and STAR Methods (A) AUC and iC50 values (as percentage, see STAR Methods) for the EGFR/ERBB2 inhibitor BIBW2992 (Afatinib). Dots represent estimates using all technical replicates and error bars are standard errors of the estimates obtained using each technical replicate individually. (B) BRCA1 promoter methylation percentage measured by RRBS (n = 33 models). (C) BRCA1 expression measured by expression microarrays (n = 35 samples from 19 models).

Figure S7

Related to Figure 6 and STAR Methods (A) Table showing drug combinations tested. (B) Scatterplot showing AUC values for all drugs tested in both individual compound screen and as single agent in the drug-drug combination screen. Pearson correlation score is indicated (n = 125 comparisons on 8 models).