Table 2.

Accuracy of CNV-calling algorithms.

Algorithm(s)	Platform	Validation Method	Accuracy	Study Conclusion	Reference
Adapted method on SW-ARRAY and GIM	Affymetrix	qPCR or Mass Spec Validation	2.5% false positives, ~90% singleton validation	Developed a multistep algorithm to better call CNVs.	[41]
Birdsuite, CNAT, CNVPartition, GADA, Nexus, PennCNV and QuantiSNP	Affymetrix, Illumina	Comparison of HapMap samples to Kidd et al., Korbel et al. and Redon et al., data [5,40,56]	Assay sensitivity ranged 20%−49% with some algorithms predicting more events (i.e., GADA, 546 predicted CNVs).	PennCNV had the greatest sensitivity (49%). Little agreement between studies and within studies.	[37]
cnvHap, CNVPartition, PennCNV and QuantiSNP	Aglient, Illumnina	Compared samples either with previously characterized (by aCGH) CNVs or HapMap samples from Kidd et al. [40]	cnvHap had very good sensitivity (68%) for larger CNVs (>10kb) in Kidd et al. This reduced to 31% for smaller CNVs (<5kb).	cnvHap has increased sensitivity compared with other CNV algorithms.	[52]
PennCNV, Aroma.Affymetrix, APT and CRLMM	Affymetrix	Compared concordance between calling algorithms.	Greater concordance in deletion (51.5%) than duplications (47.9%). The probable false positive rates for CRLMM and PennCNV were 26% and 24%.	PennCNV appeared to detect all the CNV and more than CRLMM predicted	[57]
CNVPartition, PennCNV and QuantiSNP	Illumnina	Agreement between algorithms	Agreement varied from 59%−62% for deletions, to 43%−57% for duplications.	Use of multiple algorithms increased the positive predictive value, as did the number of probes and the minimum size (kb).	[35]
CNVPartition, PennCNV and QuantiSNP	Illumnina	MLPA validation, measures were taken to reduce false positive calls.	All algorithms show better specificity than sensitivity. QuantiSNP was the most sensitive, predicting 28% of CNVs. PennCNV was better at discriminating copy number state.	Applying methods to reduce false positives results in low sensitivity.	[42]
ADM-2, Birdsuite, CNVfinder, CNVPartition, dCHIP, GTC, iPattern, Nexus, Partek, PennCNV, QuantiSNP	CGH arrays and SNP arrays (Affymetrix and Illumina)	Experiments were repeated in triplicate and CNV calls were compared. CNV calls were also compared to 5 references (‘gold standards’).	Algorithm replication has <70% reproducibility. CNV calls between any two algorithms is typically low (25%–50%) within a platform. Overlap with DGV was high, whereas overlap with references [39,40] was low.	Newer high resolution arrays outperform older arrays in both CNVs’ call and reproducibility. Algorithms developed for specific array platforms outperformed adapted and independent algorithms.	[58]
Birdsuite, Partek, Genomics Suite, HelixTree and PennCNV	Affymetrix	Comparison with HapMap CNV in two studies [39,40].	Overlap ranged between 42% and 70% when including 20 probes for Kidd et al. [40] and 26%−48% in Conrad et al. [39]	Birdsuite outperformed the other 3 algorithms over multiple permutation.	[38]
		qPCR validation of rare CNVs (a single CNV event in >1000 bipolar samples)	For each algorithm between 10 or 11, CNVs were tested. Partek and Birdsuite both validated all (5/5) deletion events tested.	Birduite and Partek had high positive predictive values, particularly with deletions. HelixTree performed poorly.
CNVPartition, PennCNV and QuantiSNP	Illumnina	Comparison to a previous CGH study [59]. qPCR validation of 3 candidate loci in 717 horses.	50 CNVs were called by all 3 algorithms. QuantiSNP had the highest overlap with CNVs predicted from CGH arrays (25%). Validation rates were greater than 80% for the 3 loci.	CNVPartition predicted the least CNVs, suggesting a high false negative rate.	[60]
GenoCN, PennCNV and QuantiSNP	Illumnina	Comparison of HapMap sample to Conrad et al.[39] Compared both CNVs (i.e. Gain or Loss) and normal calls.	All algorithms show much better specificity than sensitivity. PennCNV had the worst sensitivity, predicting <15% of Conrad et al. [39] CNVs in 3 samples	The three HMM algorithms all performed with varied results. They were all highly specific (>98%), but sensitivity remains to be an issue for all three algorithms.	[36]
cnvHap, COKGEN, GenoCNV, HaplotypeCN, PennCNV and QuantiSNP	Affymetrix	Compared 270 HapMap samples which have been previously described. Compared simulated data to test haplotype phasing between cnvHap and HaplotypeCNV.	GenoCNV has the most sensitivity (28%) when using Kidd et al. [40]; however, the concordance rate in PennCNV was greater (36% and 9%, respectively).	Algorithm performance varied with reference study. GenoCNV was the most sensitive but had the lowest concordance rate. HaplotypeCNV, cnvHap and PennCNV (under a specific permutation) were compared separately, with HaplotypeCN outperforming the other two.	[61]
Birdsuite, dCHIP, GTC and PennCNV	Affymetrix	Comparison to a previous CGH study [62].	GTC had the highest portion of CNV matching (50% overlap) to CGH, 66%. Larger CNVs were called with greater accuracy.	Birdsuite called the most CNVs; however, PennCNV outperformed all algorithms with greater specificity and sensitivity.	[63]

Abbreviations: aCGH, array comparative genomic hybridisation; APT, Affymetrix Power Tools; CNV, copy number variant; CRLMM, corrected robust linear mixture model; DGV, Database of Genomic Variants (http://dgv.tcag.ca/dgv/app/home ); HMM, hidden Markov model; GTC, Genotyping Console; kb, kilobases; MLPA, Multiplex ligation-dependent probe amplification; qPCR, quantitative polymerase chain reaction.