Empirical evaluation of data normalization methods for molecular classification.

Area of Science:

• Bioinformatics
• Genomics
• Biostatistics

Background:

• Microarray studies are prone to data artifacts from experimental handling, potentially causing biased findings.
• Data normalization is a common method for artifact correction, primarily evaluated for single biomarker discovery.
• Its performance in developing multi-marker molecular classifiers, crucial for personalized medicine, is less understood.

Purpose of the Study:

• To evaluate the effectiveness of three common data normalization methods for molecular classification.
• To assess their performance in the context of confounding handling effects in microarray data.
• To investigate the impact of normalization on classification accuracy and cross-validation bias.

Main Methods:

• Extensive simulations using re-sampled microRNA microarray datasets.
• Evaluation of three widely-used data normalization techniques.
• Analysis of classifier performance on independent test data and through cross-validation.

Main Results:

• Normalization generally improved classifier accuracy on independent test data when handling effects were present.
• The degree of improvement varied based on signal strength, handling effect distribution, and classifier algorithm.
• Cross-validation consistently resulted in over-optimistic accuracy estimates across all normalization methods.

Conclusions:

• Data normalization can be beneficial for molecular classification accuracy in the presence of handling artifacts.
• Normalization does not resolve the inherent over-optimism bias associated with cross-validation for accuracy assessment.