Whole Genome Array: What It Measures and What It Misses

“Whole genome array” is a genome-wide assay built on a fixed set of probes: hundreds of thousands to a few million positions spread across all 24 chromosomes, interrogated by hybridization. It covers the whole genome in the sense that no chromosome is excluded. That coverage does not mean every base is read. Two distinct platforms share the name. A clinical chromosomal microarray (CMA) is designed to find copy number changes (deletions, duplications, and stretches of homozygosity) and is ordered and interpreted by a clinician. A SNP genotyping array (the Illumina Global Screening Array, Infinium Omni, Affymetrix/Thermo Axiom family) calls genotypes at a curated marker set and is the engine behind consumer genotyping and most GWAS cohorts. Neither reads novel variation. If a variant is not on the chip and not in strong linkage disequilibrium with something that is, the array is silent about it.

Resolution: what each platform can and cannot see

A modern clinical CMA such as CytoScan HD carries roughly 2.6 million probes, of which about 750,000 are SNP probes and the rest are non-polymorphic copy number probes. Probe spacing is not uniform. Coverage is dense over known disease genes and recurrent microdeletion regions and sparse elsewhere. Effective resolution is usually quoted around 25-50 kb in targeted regions and 100-200 kb in backbone regions, and labs typically apply reporting thresholds like 200 kb for gains and 50 kb for losses to keep the noise out of the report.

What CMA finds: aneuploidy, unbalanced translocations, microdeletion and microduplication syndromes, mosaicism above roughly 20-30% cellularity, and long runs of homozygosity that suggest consanguinity or uniparental disomy. What it does not find: balanced translocations and inversions (no net copy change, so the log R ratio is flat), single-nucleotide variants, small indels, repeat expansions, and most of the structural variation under 10 kb.

A SNP genotyping array is a different tradeoff. 650,000 markers on the GSA, up to 2.4 million on Omni2.5, all preselected. You get high per-call accuracy at genotyped sites (concordance with sequencing above 99% for common variants on well-clustered probes), near-zero ability to detect anything rare, and systematically poor performance on indels and in segmental duplications. Rare pathogenic variants get thin representation by design: array content is optimized for tagging common haplotypes, and variants at 0.01% frequency cannot be tagged efficiently. The ENCODE pilot made the structural point early: functional elements are distributed across intergenic and non-coding sequence far beyond the exons and tag SNPs that array content historically prioritized.¹

Methylation arrays are the third member of the family. The Illumina MethylationEPIC BeadChip covers roughly 850,000 CpG sites out of about 28 million in the genome, which is under 3% of CpGs, concentrated in promoters, CpG islands, and enhancer regions. A careful evaluation against whole-genome bisulfite sequencing found EPIC probes reproducible and well correlated with sequencing at covered sites, while also documenting probes affected by underlying SNPs and cross-reactive probes that need filtering before analysis.² If you run EPIC data, filter cross-reactive and SNP-overlapping probes before you compute anything downstream.

The file formats and the pipeline we would use

Illumina arrays produce paired .idat files per sample, one per channel (Grn and Red), a few megabytes each. From there:

• illuminaio::readIDAT() in R, or Illumina’s Array Analysis CLI, converts IDAT plus a manifest (.bpm) and cluster file (.egt) into GTC.
• gtc_to_vcf (Illumina’s Python tool, with --genome-fasta-file pointed at the same build as your manifest) converts GTC to VCF. Check the build. Most older array manifests are GRCh37, and silently mixing builds is the single most common failure mode in array work.
• For association or relatedness work, plink2 --vcf sample.vcf --make-bed gives you .bed/.bim/.fam. Run --geno 0.02 --mind 0.05 --hwe 1e-6 as baseline QC and check --check-sex against reported sex before trusting anything.

For copy number from array intensity, the two quantities that matter are log R ratio (LRR) and B allele frequency (BAF). LRR is the normalized total intensity, a proxy for copy number. BAF is the fraction of signal from the B allele, which distinguishes AB from AA/BB and reveals loss of heterozygosity. PennCNV consumes LRR/BAF directly and segments with a hidden Markov model. ASCAT and its derivatives do the same for tumor samples with purity and ploidy estimation. Two QC numbers to check before you believe any CNV call: LRR standard deviation (below 0.2-0.3 for usable data) and BAF drift. A sample with LRR SD of 0.35 will generate dozens of spurious calls.

For CNV from sequencing instead, CNVkit on a 30x CRAM (cnvkit.py batch sample.cram -r reference.cnn) gives .cnr/.cns files and reaches breakpoint resolution arrays cannot, because read depth and split reads are continuous signals rather than fixed probe positions.

When an array is the right instrument and when it is not

We would reach for an array in three situations: you need a cheap genotype panel across thousands of samples, you are doing GWAS or polygenic score work where imputation against a reference panel fills the gaps, or a clinician has ordered CMA as a first-tier test for a specific indication. Imputation is the reason arrays remain useful. With a reference panel like TOPMed or HRC and Beagle 5 or Minimac4, 650k genotyped markers expand to tens of millions of imputed sites at high accuracy for common variants. Imputation quality (INFO/R2) degrades sharply below 1% minor allele frequency, so filter at R2 > 0.8 and do not treat imputed rare variants as observed.

For a single individual who wants a durable molecular baseline, 30x whole-genome sequencing is the better buy. It captures SNVs, indels, structural variants, mitochondrial variants, and copy number in one assay, and it does not become obsolete when the next array manifest ships. The cost argument that kept arrays dominant has collapsed: the $1000 genome was a stated engineering target two decades ago, and short-read sequencing passed it years back.³ The remaining cost is interpretation. Generation is not the expensive part any more.

Interpretation is where sequencing gets harder. That step does not get easier. Clinical genetics guidance on whole-genome sequencing is explicit that unfiltered genome-wide analysis produces unsolicited findings and variants of uncertain significance at a rate that requires a defined filtering strategy and pre-test agreement about what will be returned.⁴ That guidance exists because the problem is real. A typical genome carries millions of variants, thousands of rare protein-altering ones, and a handful that some database labels pathogenic on thin evidence. Variant classification for anything you would act on belongs with a clinical geneticist and a CLIA-certified lab. It does not belong with a script you wrote on a Saturday.

Questions people also ask

How long do microarray results take? Clinical CMA is typically 2-4 weeks from sample receipt, with prenatal or urgent cases run in 5-10 days. Consumer genotyping arrays usually quote 4-8 weeks, most of which is batching and logistics rather than assay time. The hybridization and scan itself takes about a day and a half.

What is a normal result from a microarray test? A normal CMA report states no clinically significant copy number gains or losses were detected at the resolution of the assay, often with a note about the platform’s limits. Benign copy number variants are present in everyone and are not reported. Normal does not exclude balanced rearrangements, single-nucleotide variants, or small structural changes below the reporting threshold.

How do I interpret the results of a microarray test? Reports use ISCN nomenclature, for example arr[GRCh37] 15q11.2(22,805,313-23,094,530)x1, meaning one copy where two are expected across that interval. The key fields are the interval, the copy number, the genes contained, and the classification. The classification is pathogenic, likely pathogenic, uncertain significance, likely benign or benign. Classification is a clinical judgment made against ClinGen, DECIPHER, and population frequency data, and a variant of uncertain significance needs a genetic counselor or clinical geneticist to put in context.

What genetic disorders does a microarray test for? CMA detects dosage imbalance wherever it occurs. That covers aneuploidies, recurrent microdeletion and microduplication syndromes, and non-recurrent deletions and duplications. Examples of the recurrent syndromes are 22q11.2, 7q11.23, 15q11-q13, and 16p11.2. CMA is not a test for a named list. Reported diagnostic yields in developmental delay, intellectual disability, and congenital anomaly cohorts are commonly in the 15-20% range.

What are the three most common chromosomal abnormalities? Among liveborn autosomal aneuploidies, trisomy 21, trisomy 18, and trisomy 13, in that order of frequency. Sex chromosome aneuploidies (47,XXY and 45,X) are also common and frequently detected incidentally on arrays run for other reasons.

Footnotes

1. Ewan Birney, Paul Flicek, Damian Keefe, et al. Identification and analysis of functional elements in 1% of the human genome by the ENCODE pilot project. Nature, 2007. https://doi.org/10.1038/nature05874 ↩

2. Ruth Pidsley, Elena Zotenko, Timothy J. Peters, et al. Critical evaluation of the Illumina MethylationEPIC BeadChip microarray for whole-genome DNA methylation profiling. Genome biology, 2016. https://doi.org/10.1186/s13059-016-1066-1 ↩

3. Simon T. Bennett, C.L. Barnes, Anthony J. Cox, et al. Toward the $1000 Human Genome. Pharmacogenomics, 2005. https://doi.org/10.1517/14622416.6.4.373 ↩

4. Carla van El, Martina C. Cornel, Pascal Borry, et al. Whole-genome sequencing in health care. European Journal of Human Genetics, 2013. https://doi.org/10.1038/ejhg.2013.46 ↩