Gene Signature-based Prediction of Tumor Response to Cyclophosphamide

Abstract

Cyclophosphamide (CY) is a clinically used cytotoxic agent that is effective in a wide range of tumor types including breast and small cell lung cancers. However, by far not all patients benefit from CY therapy. We used patient tumor explants grown in nude mice as an experimental model system to identify a gene signature that, based on a tumor's gene expression profile, predicts its CY response. Forty-nine human tumor xenografts of different histologies were defined as the training set. Correlation of the gene expression profiles of untreated tumors to the sensitivity of the same tumors to CY led to the identification of 129 transcripts as predictive biomarkers for CY response. Interestingly, the products of 12 of these genes were known to interact at least indirectly with CY. A leave-one-out cross-validation approach led to a correct prediction of the CY response of the training set tumors in 15 out of 18 cases (83%) as compared to a response rate of 18 out of 49 (32%), following random testing. For an independent set of 25 previously untested tumors with known gene expression profiles (validation set) CY sensitivity was predicted correctly for 6 out of 8 tumors (75%), and CY resistance for 15 out of 17 tumors (88%). In comparison, random testing of the same tumors resulted in a response rate of 8 out of 25 (32%). For the same 25 tumors, the median minimum T/C value for predicted responders was 1% as compared to 49% for predicted non-responders. Finally, for tumor types considered as CY sensitive such as small cell lung and breast cancers as well as melanoma, the combined real and predicted response rates for 37 tested and 26 untested tumors was 49%. In contrast, for tumor types considered as CY resistant, including colon and renal cancer, the combined real and predicted response rate for 37 tested and 75 untested tumors was only 13%. Taken together, we identified a gene signature that can predict tumor response to CY and warrants clinical validation.