A new method for class prediction based on signed-rank algorithms applied to Affymetrix microarray experiments.

Area of Science:

• Bioinformatics
• Genomics
• Medical Informatics

Background:

• Microarray data analysis faces challenges due to evolving technology, hindering data integration for class prediction.
• Affymetrix technology offers quantitative signals and detection calls, but integrating data across different chip types is difficult.
• A new prediction method was developed using only detection calls from recent Affymetrix chip types for improved class prediction.

Purpose of the Study:

• To develop a novel prediction method for class belonging based on Affymetrix detection calls.
• To apply the method to biological data from normal B cells and multiple myeloma (MM) patient groups.
• To evaluate the method's performance in predicting clinical features and compare it with existing methods.

Main Methods:

• A call-based data reduction step filtered non-class-discriminative probe sets.
• A predictor was developed with correction for multiple testing via iterative probe set deletion.
• Error rates were determined using leave-one-out and 5-fold cross-validation.

Main Results:

• The method successfully predicted patient sex, classifying gender with high accuracy across different groups.
• Immunoglobulin light and heavy chains expressed by malignant myeloma clones were accurately predicted.
• The method demonstrated superior performance compared to the Prediction Analysis of Microarray (PAM) method.

Conclusions:

• The normalization-free method is valuable for quality control and correcting errors in patient reports.
• The method is extendable to multiple class predictions for clinical groups and international projects.
• It serves as a powerful tool for mining published Affymetrix data and classifying samples with binary features.