What a SNP Microarray Test Measures, and What It Misses

A SNP microarray is a fixed panel: several hundred thousand to a few million oligonucleotide probes, each interrogating one pre-chosen genomic position. It reports a genotype (AA/AB/BB) at those positions and nothing at the positions it does not carry. Two derived signals come out of the raw probe intensities: Log R Ratio (total intensity relative to a reference cluster, a copy-number proxy) and B Allele Frequency (the allelic ratio, 0 / 0.5 / 1 in a normal diploid region). Clinically, that combination is used as chromosomal microarray analysis: finding deletions and duplications, and finding runs of homozygosity that suggest consanguinity or uniparental disomy. It does not sequence anything. Any variant not on the manifest is invisible.

The assay, briefly

The dominant chemistry is Illumina Infinium. Genomic DNA is whole-genome amplified, fragmented, and hybridized overnight to bead-bound 50mer probes, then extended by one labeled base at the SNP position. Each bead type is replicated roughly 10-20 times on the array, so every call is an average over replicates, which is why array genotypes are accurate (>99.5% concordance on well-clustered markers) despite the crudeness of hybridization. The scalable single-base-extension design that made whole-genome arrays practical is described in the original Infinium paper ¹.

Two practical consequences follow from “fixed panel.” First, marker density determines resolution, and density is wildly uneven: consumer and research arrays like the Illumina GSA (~650k markers) are thin in gene deserts and thin in GC-rich first exons. Clinical arrays solve this by adding non-polymorphic copy-number probes to the backbone and tiling known disease genes at exon-level density, which is how single-exon deletions become detectable at all ². Second, ascertainment bias: markers were selected from HapMap-era European-heavy discovery panels, so allele frequency estimates and imputation quality degrade outside those populations.

What comes off the machine and how to process it

Raw output is a pair of .idat files per sample (Grn and Red). You need the product manifest (.bpm) and a cluster file (.egt) to make calls.

# genotype calling
iaap-cli gencall Manifest.bpm Clusters.egt ./out \
  --idat-folder ./idats --output-gtc --num-threads 8

# GTC -> VCF with LRR/BAF retained
bcftools +gtc2vcf --bpm Manifest.bpm --egt Clusters.egt \
  --gtcs ./out --fasta-ref GRCh38.fa --extra out.tsv -Ob -o array.bcf

From there, plink2 --bcf array.bcf --make-bed gets you .bed/.bim/.fam for anything population-genetic.

The failure modes worth knowing before you trust a single call:

• Strand. Illumina A/B allele coding is not reference/alternate. Flipping A/T and C/G sites is guesswork without allele frequency context. Use gtc2vcf, which resolves strand against the reference, rather than hand-mapping from a TOP/BOT manifest.
• Build. Older manifests are GRCh37. Mixing builds silently shifts coordinates. Check ##contig lines before any join.
• Call rate. Drop samples below 0.98 and markers below 0.98. A sample at 0.93 usually means degraded DNA, and its LRR will be too noisy for copy-number work regardless.
• Cluster file mismatch. Using a cluster file generated on a different chemistry version produces plausible-looking genotypes with elevated heterozygosity. Check per-sample het rate against the batch distribution.

Copy number: the part people buy the array for

CNV calling from LRR and BAF is a hidden Markov model problem. PennCNV is still the reference implementation:

detect_cnv.pl -test -hmm hhall.hmm -pfb hg38.pfb \
  -gcmodel hg38.gcmodel -log run.log -out sample.rawcnv sample.txt

Quality gates that matter: LRR standard deviation below 0.30, BAF drift below 0.01, and absolute waviness factor below 0.05 after GC correction. Above those thresholds the call count explodes with false positives. Require at least 10 consecutive markers and roughly 20 kb before you look at a call at all, and note that “20 kb” on a sparse array may mean a single-probe span in a marker-poor region. Effective resolution is set by local probe density. It is not set by the nominal array size.

BAF is what distinguishes a SNP array from an older array CGH design. A deletion shows LRR down plus loss of the 0.5 BAF band. A copy-neutral loss of heterozygosity shows flat LRR with the 0.5 band gone entirely, which array CGH cannot see. The signature is how uniparental disomy is detected, including segmental isodisomy that turns out to be more common in the general population than once assumed ³. Combining CGH-style backbone probes with SNP probes in one assay was adopted clinically for exactly this reason ⁴. Mosaicism shows as BAF band splitting: instead of one band at 0.5 you get two, separated in proportion to the mosaic fraction, usually detectable down to about 10-20% depending on marker density and LRR noise.

Runs of homozygosity and ancestry

plink --bfile array --homozyg \
  --homozyg-snp 50 --homozyg-kb 1000 \
  --homozyg-window-het 1 --homozyg-density 50

Total ROH length is informative about parental relatedness and flags regions where a recessive variant could be homozygous by descent. It is a pointer for follow-up sequencing. It is not a diagnosis.

For ancestry, array genotypes are more than sufficient. Tens of thousands of common markers resolve population structure at a fine scale, including substructure within Europe ⁵. Running plink2 --pca, projected onto 1000 Genomes reference samples, takes minutes.

Imputation extends the panel, within limits

Send your VCF to the TOPMed Imputation Server (Eagle v2.4 for phasing, Minimac4 for imputation). A 650k-marker array typically yields tens of millions of imputed sites, but keep only those with r² above 0.8, and understand that r² collapses for variants below roughly 1% frequency. Imputation recovers common variation you did not type. Imputation does not recover rare or de novo variants, which is precisely the class that drives most single-gene findings.

Where SNP arrays are used clinically, and what they miss

Chromosomal microarray is the established first-tier test for developmental delay, intellectual disability, and autism spectrum disorders with dysmorphic features, replacing G-banded karyotype for that indication ⁶. In prenatal work, SNP arrays give a rapid answer on genomic imbalance with a defined counseling workflow around incidental findings ⁷. Across consecutive ASD and learning-disability referrals, reported yields of clinically significant CNVs land in the high single digits to low teens, varying with array design and how strictly pathogenicity is called ^{8 9}. Higher-density arrays designed around neurodevelopmental gene content raise that yield ¹⁰.

What the assay cannot detect: single-nucleotide and small indel variants outside the manifest, balanced translocations and inversions (copy number is unchanged), repeat expansions, low-level mosaicism below the BAF-splitting threshold, triploidy on non-SNP arrays, and methylation. If your question is “what rare coding variants do I carry,” an array answers it badly and whole-genome sequencing answers it well. If your question is “do I have a 1.5 Mb deletion,” an array answers it cheaply and reliably.

One boundary worth stating plainly: a CNV call from research-grade array data is a hypothesis. Clinical interpretation, confirmation in a CLIA-certified laboratory, parental testing to establish inheritance, and any decision that follows belong with a medical geneticist or genetic counselor.

Questions people also ask

Can a microarray detect autism? No. It detects copy-number changes, some of which are recurrent CNVs associated with neurodevelopmental phenotypes. A positive finding identifies a genomic cause in a minority of tested individuals, and a negative result rules nothing out ^{9 8}. Autism is a clinical diagnosis made by behavioral assessment.

What is a chromosomal SNP microarray used for? Detecting deletions and duplications genome-wide, detecting copy-neutral loss of heterozygosity and uniparental disomy, and estimating mosaic fraction. In practice it is ordered for developmental delay, intellectual disability, multiple congenital anomalies, and for prenatal and products-of-conception analysis ^{6 7}.

What is the difference between a microarray and a SNP array? “Microarray” is the general format. Array CGH microarrays compare test and reference DNA intensities and report copy number only. SNP arrays genotype each probe position, so they additionally give B Allele Frequency, which reveals loss of heterozygosity, uniparental disomy, and mosaicism. Many clinical platforms now include both probe types on one chip ⁴.

What does a SNP microarray test for at the variant level? Genotypes at the specific positions on its manifest, plus copy number inferred from intensity. It does not test for anything absent from the manifest, and it does not read sequence.

Footnotes

1. Kevin L Gunderson, Frank J Steemers, Grace Lee, et al. A genome-wide scalable SNP genotyping assay using microarray technology. Nature Genetics, 2005. https://doi.org/10.1038/ng1547 ↩

2. Patricia B.S. Celestino-Soper, Chad A. Shaw, Stephan J. Sanders, et al. Use of array CGH to detect exonic copy number variants throughout the genome in autism families detects a novel deletion in TMLHE. Human Molecular Genetics, 2011. https://doi.org/10.1093/hmg/ddr363 ↩

3. Kensaku Sasaki, Hiroyuki Mishima, Kiyonori Miura, et al. Uniparental disomy analysis in trios using genome-wide SNP array and whole-genome sequencing data imply segmental uniparental isodisomy in general populations. Gene, 2013. https://doi.org/10.1016/j.gene.2012.10.035 ↩

4. Joanna Wiszniewska, Weimin Bi, Chad Shaw, et al. Combined array CGH plus SNP genome analyses in a single assay for optimized clinical testing. European Journal of Human Genetics, 2013. https://doi.org/10.1038/ejhg.2013.77 ↩ ↩²

5. Jinchuan Xing, W. Scott Watkins, David J. Witherspoon, et al. Fine-scaled human genetic structure revealed by SNP microarrays. Genome Research, 2009. https://doi.org/10.1101/gr.085589.108 ↩

6. Agatino Battaglia, Viola Doccini, Laura Bernardini, et al. Confirmation of chromosomal microarray as a first-tier clinical diagnostic test for individuals with developmental delay, intellectual disability, autism spectrum disorders and dysmorphic features. European Journal of Paediatric Neurology, 2013. https://doi.org/10.1016/j.ejpn.2013.04.010 ↩ ↩²

7. Malgorzata Srebniak, Marjan Boter, Grétel Oudesluijs, et al. Application of SNP array for rapid prenatal diagnosis: implementation, genetic counselling and diagnostic flow. European Journal of Human Genetics, 2011. https://doi.org/10.1038/ejhg.2011.119 ↩ ↩²

8. Jennifer L. Roberts, Karine Hovanes, Majed Dasouki, et al. Chromosomal microarray analysis of consecutive individuals with autism spectrum disorders or learning disability presenting for genetic services. Gene, 2014. https://doi.org/10.1016/j.gene.2013.10.020 ↩ ↩²

9. Caroline Nava, Boris Keren, Cyril Mignot, et al. Prospective diagnostic analysis of copy number variants using SNP microarrays in individuals with autism spectrum disorders. European Journal of Human Genetics, 2013. https://doi.org/10.1038/ejhg.2013.88 ↩ ↩²

10. Karen Ho, E. Wassman, Adrianne Baxter, et al. Chromosomal Microarray Analysis of Consecutive Individuals with Autism Spectrum Disorders Using an Ultra-High Resolution Chromosomal Microarray Optimized for Neurodevelopmental Disorders. International Journal of Molecular Sciences, 2016. https://doi.org/10.3390/ijms17122070 ↩