Convergent Random Forest predictor: methodology for predicting drug response from genome-scale data applied to...

Area of Science:

• Genomics
• Bioinformatics
• Translational Medicine

Background:

• Molecular profiling aims to identify biomarkers for predicting patient response to therapies.
• Analyzing high-dimensional gene expression data presents challenges due to sample size limitations and data variability.
• Identifying a minimal set of genes sufficient for accurate prediction is crucial for developing diagnostic tools.

Purpose of the Study:

• To present the Convergent Random Forest (CRF) method for identifying highly predictive biomarkers from genome-wide expression data.
• To select a small, non-redundant set of biomarkers suitable for a simple and robust diagnostic tool.
• To evaluate the CRF approach's performance in identifying predictive genes for therapy response.

Main Methods:

• The Convergent Random Forest (CRF) method combines Random Forest classification with gene expression clustering.
• CRF ranks and selects a small number of predictive genes from genome-wide expression data.
• The approach was evaluated on four different datasets, including rheumatoid arthritis patient response to anti-TNF therapy.

Main Results:

• CRF identified 8 transcripts predicting anti-TNF therapy response in rheumatoid arthritis patients with 89% accuracy.
• Compared to recursive support vector machines (RSVM), CRF consistently selected a smaller number of features (5-8 genes).
• CRF achieved similar or superior performance to RSVM on both training and independent testing datasets.

Conclusions:

• The CRF method is effective in identifying a minimal set of highly predictive biomarkers from complex genomic data.
• CRF offers a robust approach for developing simple and reliable diagnostic tools for predicting therapy response.
• The method demonstrates strong performance and efficiency compared to existing feature selection techniques.