MicroRNA signatures from FFPE tissue distinguish patients with colorectal cancer of UICC stage I from those with metastatic disease of stage IV.

Abstract

e14055 Background: miRNAs have a length of 21-23 bp. They are more stable than mRNAs, especially in formalin-fixed paraffin embedded (FFPE) tumor tissue. The increased stability of miRNA makes them attractive as potential biomarkers. Methods: In the frame work of a prospective, multicenter, IRB approved study we recruited more than 4.000 patients (pts) with colorectal cancer in all four UICC stages from 33 primary care und university hospitals in Germany. Informed consent was obtained from all pts. For this study we used a set of 357 FFPE tissues from stage I (n=179) and stage IV (n=178). All FFPE tissues were quality controlled by an in-house pathologist. Total RNA including miRNA was isolated using miRNeasy FFPE kits and QiaCube robots (Qiagen). 250ng total RNA was labeled using the FlashTag kit (Genisphere) and then hybridized onto GenChip miRNA arrays (Affymetrix). Results: After all QC, 311 array data sets remained. A training set (N=114, stage I=58, stage IV=56) and an test set (N= 197, stage I=103, stage IV=94) were formed. miRNA signatures with a length of 2 to 1.000 probesets were identified using FARMS condensation followed by a nested bootstrap approach with random forest and SVM for feature selection. The best signature (1.000 probesets) showed a sensitivity (S+) of 73% and a specificity (S-) of 71%. This signature was validated in the test set and achieved a S+ of 71% (95% CI: 0.610-0.801) and a S- of 66% (95% CI: 0.560-0.751). Furthermore, the potential of randomly drawn training sets of different sizes was examined in a bootstrap approach by which training sets of size 20, 40, 80, 114, 180, and 280 were repeatedly sampled. A discovery pipeline was then applied which also used a double-nested bootstrap approach. Resulting classifiers were cross-validated on all samples outside of its training set, adding validation performance to the prospective estimates of the discovery pipeline. Training sets of 180-280 samples yielded better results than the original 114 sample set, and outperformed all small sets of 20-80 samples. Conclusions: We will apply the best miRNA signatures to identify pts with stage II and stage III disease who are at high risk of metastasis.