Abstract 2905: Validation techniques for PCA-based gene expression signatures.

Abstract

Abstract Background: Many gene expression signatures exist for measuring the biological state of a profiled tumor. One way to derive a gene-signature of a biological event is to perturb cell lines to mimic the event. The signature is applied to tumor data to predict the magnitude of the biological effect in tumors. Examples include pathway activation, metastasis scores, and chemo-resistance. Principal Component Analysis (PCA) summarizes a gene signature into a score, however many published gene signatures capture proliferation, rather than the intended process, as the first PCA component (PC) when applied to tumor datasets Venet et al. (Venet, 2011, PLoS Comput. Biol.). Determining why this happens and how this effect can be detected is an important for utilizing these signatures. Aim: To develop a set of tools to determine if a derived gene signature works as intended and is robustly represented when applied to a tumor dataset. Results: Differences observed between a cell line experiment and tumor data set is a problem of calibration transfer. The signature is derived on data with controlled variation and then applied to data often exhibiting much larger variation. PCA is a powerful tool for summarizing a gene signature into a score, but there are several pitfalls, particularly when the variation is larger than expected. We have developed a visual tool for evaluating the application of a signature, PCA characteristics are measured against thousands of random gene signatures to determine the significance of the findings. Four key concepts are measured using this tool. Coherence: Elements of a signature should be correlated beyond chance; if a common mechanism is measured, there should be coherence to the gene signature. This is the amount of variance explained in the first PC. Uniqueness: The general direction of the data can drive most of the observed signal. A gene signature possessing a unique direction (relative to the entire dataset) provides confidence that it is measuring a true effect. Robustness: If a signature is designed to measure a single biological effect, then this signal should be sufficiently strong and distinct compared to other signals within the signature. This is measured by calculating the ratio between the explained variance of the first and second PCA components. Transferability. The derived PCA gene signature score should describe the same biology in the tumor data set as it does in the cell line data. This can be verified by comparing the cell-line based PCA model with the tumor-based PCA model. If the sign and relative importance (loadings) of individual genes are similar between the two PCA models, this is an indication that the calibration transfer was successful. Conclusions: We have developed a technique for validating PCA-based gene signatures work as intended when applied to tumor data. Application of this technique can identify instances in which signature scores do not represent the desired biological effect. Citation Format: Anders E. Berglund, Eric A. Welsh, David A. Fenstermacher, Steven A. Eschrich. Validation techniques for PCA-based gene expression signatures. [abstract]. In: Proceedings of the 104th Annual Meeting of the American Association for Cancer Research; 2013 Apr 6-10; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2013;73(8 Suppl):Abstract nr 2905. doi:10.1158/1538-7445.AM2013-2905