Cross platform comparison of multigene predictors of response to neoadjuvant paclitaxel/FAC chemotherapy in breast cancer generated by cDNA arrays and Affymetrix GeneChips

Abstract

503 Background: Transcriptional profiling is under study as a clinical research tool to classify cancer into clinically relevant sub-groups. Comparison of results across multiple laboratories and platforms is critical for the future development of a useful diagnostic test. Methods: In this study, we compared muligene predictors developed on 2 distinct profiling platforms Affymetrix GeneChip and Millennium cDNA arrays using pre-treatment fine needle aspirates of 33 breast cancer patients. RNA was hybridized from the same samples to both arrays. We independently identified gene expression signatures associated with pathologic complete response (pCR) to neoadjuvant paclitaxel followed by 5-FU, doxorubicin, cyclophosphamide chemotherapy on each platfrom and examined their predictive value on the other platform. Results: Only 30 % of all measurements of corresponding genes on both platforms showed Person correlation coefficient of ≥ 0.7, 54 % of cDNA clones matched with ≥ 1 probe sets on the Affy chip and there was large variation in probe set to probe set correlation with the matching cDNA result. The top response discriminating gene lists consisted of 182 genes for Affy p≤0.006 and 45 for cDNA, p≤ 0.002, with only 17 gene overlap. The top genes performed well in intra-platform supervised hierarchical clustering each correctly separating 91% of cases. However cross-platform application resulted in substantially lower clustering accuracy of 45% (Affy to cDNA) and 79% (cDNA to Affy). When only the 17 overlapping genes were used for clustering 67% (Affy) and 64% (cDNA) of cases were correctly clustered. Similar results were obtained when linear discriminant analysis and Generic Algorithm were used for outcome prediction. Conclusion: Distinct predictive marker sets can be identified from the same samples using different gene expression profiling platforms. Predictors perform well on one platform but cross-platform comparisons are less informative. Misclassification rates may be slightly improved with careful gene selection of probes with high correlation coefficients. No significant financial relationships to disclose.