Characterization of an oxaliplatin sensitivity predictor in a preclinical murine model of colorectal cancer

Abstract

Despite advances in contemporary chemotherapeutic strategies, long-term survival still remains elusive for patients with metastatic colorectal cancer. A better understanding of the molecular markers of drug sensitivity to match therapy with patient is needed to improve clinical outcomes. In this study, we used in vitro drug sensitivity data from the NCI-60 cell lines together with their Affymetrix microarray data to develop a gene expression signature to predict sensitivity to oxaliplatin. To validate our oxaliplatin sensitivity signature, patient-derived colorectal cancer explants (PDCCE) were developed in nonobese diabetic/severe combined immunodeficient (NOD/SCID) mice from resected human colorectal tumors. Analysis of gene expression profiles found similarities between the PDCCEs and their parental human tumors, suggesting their utility to study drug sensitivity in vivo. The oxaliplatin sensitivity signature was then validated in vivo with response data from 14 PDCCEs treated with oxaliplatin and was found to have an accuracy of 92.9% (sensitivity = 87.5%; specificity = 100%). Our findings suggest that PDCCEs can be a novel source to study drug sensitivity in colorectal cancer. Furthermore, genomic-based analysis has the potential to be incorporated into future strategies to optimize individual therapy for patients with metastatic colorectal cancer.

Figure 1. NCI-60 Oxaliplatin Signature Bayesian Regression Heatmap and Leave-One-Out Performance

A subset of cell lines that represent the extremes of sensitivity to oxaliplatin were identified from the NCI-60 panel. The left panel is the expression plot for genes selected for discriminating the oxaliplatin sensitive and resistant cell lines from the NCI-60 set based on the expression levels of the 120 genes selected during the Bayesian binary regression analysis. Blue represents lowest expression and red highest. Each column represents an individual cell line and each row one of the discriminating genes in the oxaliplatin predictor. From these 120 genes, phenotype discriminatory power from sample misclassification was optimized using leave-one-out performance (right panel). The acuracy of the predictor on each samples in the leave-one-out performance are displayed with 95% CI.

Figure 2. Global Comparison of Human Tumors and Matching PDCCEs

A. PDCCEs and matching human tumors were sectioned on histology slides and stained with hemotoxalin and eosin (H&E) to confirm presence of tumor tissue. PDCCE slides were also stained for carcinoembryonic antigen (CEA) by immunohistochemistry. B. Global gene expression from 14 human tumors and matching PDCCEs were estimated using RMA and normalized for batch effects using ComBat. After filtering, samples were subjected to unsupervised hierarchical clustering using pearson correlation and complete linkage. Patient-Derived Colorectal Cancer Explants are prefixed with “PDCCE”. C. The R library pvclust was used to determine statistical significance of clusters, with Approximate Unbiased (AU) values in red and Bootstrap Probability (BP) values in green. Cluster pairs with statistically significant expression correlations (AU, BP > 95%) are boxed in blue. From this it was found that 71% (10/14) of samples cluster together with statistical significance.

Figure 3. Metastasis-Derived Colorectal Cancer Explants and Tumor Growth Inhibition Versus Predicted Oxaliplatin Sensitivity

A. Displayed is a representative sample from each phenotype, with both saline-treated and oxaliplatin-treated PDCCEs, to illustrate differences in oxaliplatin response. B. In fresh frozen PDCCEs, a Pearson correlation between tumor growth inhibition and predicted oxaliplatin sensitivity was found to be statistically significant (p = 0.002). The signature was found to have an accuracy of 92.9%. C. In FFPE PDCCEs, a Pearson correlation between tumor growth inhibition and predicted oxaliplatin sensitivity was found to be statistically significant (p = 0.025). The signature was found to have an accuracy of 84.6%.