RF_Purify

RF_Purify estimates tumor purity from Illumina DNA-methylation array data using Random Forest regression models to predict ABSOLUTE- and ESTIMATE-derived purity without requiring reference or matching control samples.

Key Features:

• Random Forest Regression Models: Two Random Forest models trained on TCGA Illumina 450k DNA-methylation array data predict purity values optimized to ABSOLUTE (WGS-derived) and ESTIMATE (gene expression–derived) scores.
• Reference-Free Inference: Predicts tumor purity without prior knowledge of the tumor entity or matching control tissue.
• Array Platform Compatibility: Applicable to Illumina 450k and 850k (EPIC) array data when arrays are preprocessed with the Bioconductor package minfi.
• Validation and Benchmarking: Validated on independent test datasets and compared with ESTIMATE and LUMP on TCGA, including application across brain tumor subgroups.
• Implementation: Provided as an R package leveraging methylation array preprocessing via minfi.

Scientific Applications:

• Methylome studies of primary tumors: Enables systematic investigation of DNA methylation in bulk tumor samples by providing sample-level purity estimates.
• Adjustment for cellular composition: Supports analyses that require accounting for varying proportions of tumor cells, infiltrating immune cells, and stromal components.
• Comparative benchmarking: Facilitates comparison of methylation-derived purity estimates with orthogonal measures such as ABSOLUTE, ESTIMATE, and LUMP.
• Subtype and subgroup analyses: Allows assessment of purity heterogeneity across tumor subgroups, as demonstrated in brain tumor datasets.

Methodology:

Random Forest classifiers were trained on TCGA 450k methylation data using ABSOLUTE and ESTIMATE purity values and are applied to new methylation array datasets after preprocessing with minfi.