Background: Molecular signatures are sets of genes, proteins, genetic variants or other variables that can be used as markers for a particular phenotype. Reliable signature discovery methods could yield valuable insight into cell biology and mechanisms of human disease. However, it is currently not clear how to control error rates such as the false discovery rate (FDR) in signature discovery. Moreover, signatures for cancer gene expression have been shown to be unstable, that is, difficult to replicate in independent studies, casting doubts on their reliability. Results: We demonstrate that with modern prediction methods, signatures that yield accurate predictions may still have a high FDR. Further, we show that even signatures with low FDR may fail to replicate in independent studies due to limited statistical power. Thus, neither stability nor predictive accuracy are relevant when FDR control is the primary goal. We therefore develop a general statistical hypothesis testing framework that for the first time provides FDR control for signature discovery. Our method is demonstrated to be correct in simulation studies. When applied to five cancer data sets, the method was able to discover molecular signatures with 5% FDR in three cases, while two data sets yielded no significant findings. Conclusion: Our approach enables reliable discovery of molecular signatures from genome-wide data with current sample sizes. The statistical framework developed herein is potentially applicable to a wide range of prediction problems in bioinformatics.
Bibliographical noteFunding Information:
The authors would like to thank Drs. José M. Peña and Albert Compte for helpful discussions. This work was supported by grants from the Ph.D. Programme in Medical Bioinformatics at Karolinska Institutet (RN), Clinical Gene Networks AB, Vinnova (JT), Swedish Research Council (JT) and Linköping University.
ASJC Scopus subject areas
- Structural Biology
- Molecular Biology
- Computer Science Applications
- Applied Mathematics