Merging microarray data, robust feature selection, and predicting prognosis in prostate cancer.

Area of Science:

• Bioinformatics
• Genomics
• Cancer Research

Background:

• Individual microarray studies often lack statistical power for prognostic biomarker discovery due to limited sample sizes.
• Publicly available datasets offer opportunities to increase power by combining data across studies.
• Integrating data across different institutions and platforms presents methodological challenges.

Purpose of the Study:

• To develop and validate a novel approach for combining microarray data from multiple sources.
• To introduce and assess a new algorithm, robust greedy feature selection (RGFS), for identifying predictive genes.
• To improve the accuracy of prognostic models for prostate cancer by leveraging combined datasets.

Main Methods:

• A novel algorithm, robust greedy feature selection (RGFS), was developed for predictive gene selection.
• Two prostate cancer microarray datasets were combined and analyzed.
• The Kolmogorov-Smirnov goodness-of-fit test was used to confirm the appropriateness of data combination.
• Various predictive models, including logistic regression and linear discriminant analysis (LDA) with different feature selection methods, were constructed and compared.

Main Results:

• The best logistic regression model with stepwise forward selection used 7 genes and had a 31% misclassification rate.
• LDA models combined with various feature selection algorithms showed misclassification rates between 19% and 33%.
• The novel RGFS algorithm, when combined with LDA, yielded the best model with only two genes and a 15% misclassification rate.
• Gene sets selected during cross-validation varied substantially when using standard feature selection methods.

Conclusions:

• Combining microarray data across studies is a viable strategy to enhance statistical power for biomarker discovery.
• The robust greedy feature selection (RGFS) algorithm is effective in identifying a minimal set of highly predictive genes.
• The proposed method significantly improves the accuracy of prognostic models for prostate cancer compared to traditional approaches.