Illumina Methylation Array Preprocessing

Illumina methylation array preprocessing

Summary

Preprocessing and quality control of Illumina methylation array data (450K and EPIC) via detection p-value and bead count filtering, followed by type-2 probe correction and normalization. This skill removes low-quality probes and corrects technical biases prior to downstream analysis.

When to use

Your input is raw .idat files or a beta-valued matrix from Illumina HumanMethylation450 or EPIC arrays, and you need to remove unreliable probes (those with detection p-value > 0.01 or insufficient bead counts) before performing differential methylation or other downstream analyses. Apply this skill early in the pipeline when data quality and probe reliability directly affect downstream statistical power.

When NOT to use

• Input is already a curated, pre-normalized feature matrix from a published study without raw .idat or detection p-value metadata — skip to downstream analysis.
• Your array platform is not Illumina 450K or EPIC (e.g., custom or non-Illumina bisulfite sequencing data) — choose a platform-specific preprocessing pipeline.
• You have already applied aggressive probe filtering (< 5,000 probes remaining) — additional filtering may eliminate too much data to support statistical inference.

Inputs

• .idat files (raw Illumina methylation intensity data)
• beta-valued matrix (pre-imported methylation beta values)
• HumanMethylation450 or EPIC array sample metadata (sample IDs, phenotypes)

Outputs

• Filtered and normalized beta-value matrix
• Quality control report (probes retained vs. removed, detection p-value and bead count distributions)
• Corrected type-2 probe intensity values
• Batch-effect diagnostics (optional SVD visualizations)

How to apply

Load the raw methylation array data using ChAMP's data import functions (from .idat files or a beta-valued matrix). Apply champ.filter() with default parameters to execute two successive filtering steps: (1) removal of probes with detection p-value > 0.01, indicating unreliable signal, and (2) removal of probes with fewer than 3 beads in at least 5% of samples, which reflects insufficient technical replication. After filtering, apply type-2 probe correction using SWAN, PBC, or BMIQ (default is BMIQ) to adjust for the distinct hybridization kinetics of Infinium type-2 probes. Finally, apply Functional Normalization or another normalization method to correct for technical variation while preserving biological signal. Document probe retention rates and quality metrics (detection p-value distributions, bead count distributions pre- and post-filtering) to validate the filtering and correction steps.

Related tools

• ChAMP (Main preprocessing and analysis pipeline for Illumina methylation arrays; provides data loading, filtering, type-2 probe correction, normalization, and quality control functions) — https://github.com/YuanTian1991/ChAMP
• ChAMPdata (Data and annotation package supporting ChAMP; provides CpG-probe manifests and test datasets for 450K, EPIC v1, EPIC v2, and mouse arrays) — https://github.com/YuanTian1991/ChAMPdata
• minfi (Alternative methylation array analysis package offering Functional Normalization and other preprocessing methods compatible with ChAMP)
• BMIQ (Type-2 probe correction method (default in ChAMP) to adjust for Infinium type-2 probe bias)
• SWAN (Alternative type-2 probe correction method available in ChAMP)

Evaluation signals

• Probe count comparison: pre-filter vs. post-filter probe counts should show removal of probes with detection p-value > 0.01; exact retention percentage should be documented.
• Bead count distribution: verify that probes with < 3 beads in ≥ 5% of samples are removed by examining bead count histograms before and after filtering.
• Detection p-value distribution: all remaining probes should have detection p-value ≤ 0.01; no probes in the filtered matrix should exceed this threshold.
• Type-2 probe correction validation: type-2 probe beta values should show reduced systematic bias relative to type-1 probes (visible as reduced clustering by probe type in quality control plots).
• Quality control report completeness: documentation must include number of probes retained/removed, reason codes (detection p-value vs. bead count), and summary statistics for both filtering stages.

Limitations

• Default filtering thresholds (detection p-value > 0.01, < 3 beads in ≥ 5% samples) may be overly stringent for small sample sets or underpowered studies, reducing probe coverage.
• No methods section was provided in the source article, so filter step rationale and alternative parameter choices are not explicitly justified.
• ChAMP development version (GitHub) is noted as 'under intensive modification and upgrade'; users should rely on the formally released Bioconductor version for stable, peer-reviewed implementations.
• Functional Normalization requires sufficient samples (typically ≥ 8) to estimate control probe variation; small pilot studies may require alternative normalization strategies.
• SVD batch effect diagnostics require metadata on known batch variables; unmeasured or confounded batch factors will not be detected.

Evidence

• [other] champ.filter() applies two successive filtering steps by default: (1) removal of probes with detection p-value > 0.01, and (2) removal of probes with fewer than 3 beads in at least 5% of samples per probe.: "champ.filter() applies two successive filtering steps by default: (1) removal of probes with detection p-value > 0.01, and (2) removal of probes with fewer than 3 beads in at least 5% of samples per"
• [readme] ChAMP package is designed for conduct DNA methylation array analysis, providing service from data loading, to final gene set enrichment analysis: "ChAMP package is designed for conduct DNA methylation array analysis, providing service from data loading, to final gene set enrichment analysis"
• [intro] Type-2 probe correction methods include SWAN, Peak Based Correction (PBC) and BMIQ (the default choice): "Type-2 probe correction methods include SWAN, Peak Based Correction (PBC) and BMIQ (the default choice)"
• [intro] The popular Functional Normalization function offered by the minfi package is also available: "The popular Functional Normalization function offered by the minfi package is also available"
• [intro] a variety of different data import methods (e.g. from .idat files or a beta-valued matrix): "a variety of different data import methods (e.g. from .idat files or a beta-valued matrix)"
• [readme] Current latest version: 2.29.1, which support EPICv2, but it must be used along with ChAMPdata >= 2.23.1: "Current latest version: 2.29.1, which support EPICv2, but it must be used along with ChAMPdata >= 2.23.1"