Study design and data analysis considerations for the discovery of prognostic molecular biomarkers: a case study of progression free survival in advanced serous ovarian cancer

Li-Xuan QinEmail author and
Douglas A. Levine

BMC Medical GenomicsBMC series – open, inclusive and trusted20169:27

DOI: 10.1186/s12920-016-0187-4

Received: 8 December 2015

Accepted: 1 June 2016

Published: 10 June 2016

Open Peer Review reports

Abstract

Background

Accurate discovery of molecular biomarkers that are prognostic of a clinical outcome is an important yet challenging task, partly due to the combination of the typically weak genomic signal for a clinical outcome and the frequently strong noise due to microarray handling effects. Effective strategies to resolve this challenge are in dire need.

Methods

We set out to assess the use of careful study design and data normalization for the discovery of prognostic molecular biomarkers. Taking progression free survival in advanced serous ovarian cancer as an example, we conducted empirical analysis on two sets of microRNA arrays for the same set of tumor samples: arrays in one set were collected using careful study design (that is, uniform handling and randomized array-to-sample assignment) and arrays in the other set were not.

Results

We found that (1) handling effects can confound the clinical outcome under study as a result of chance even with randomization, (2) the level of confounding handling effects can be reduced by data normalization, and (3) good study design cannot be replaced by post-hoc normalization. In addition, we provided a practical approach to define positive and negative control markers for detecting handling effects and assessing the performance of a normalization method.

Conclusions

Our work showcased the difficulty of finding prognostic biomarkers for a clinical outcome of weak genomic signals, illustrated the benefits of careful study design and data normalization, and provided a practical approach to identify handling effects and select a beneficial normalization method. Our work calls for careful study design and data analysis for the discovery of robust and translatable molecular biomarkers.