Improving variant calling using population data and deep learning

BMC Bioinformatics

Table 1 Variant calling accuracy when inferring using 35x WGS data from HG003 and 30x WGS data from HG00733

Dataset	Variant type	Population	Precision	Recall	F1
HG003	INDEL	Agnostic	0.997351	0.993922	0.995634
		1000genomes	0.997462	0.993977	0.995716
		AFR	0.997337	0.993623	0.995476
		AMR	0.997355	0.993787	0.995568
		EAS	0.997021	0.993062	0.995038
		EUR	0.997364	0.993801	0.995579
		SAS	0.997333	0.993692	0.995509
	SNP	Agnostic	0.998131	0.993769	0.995945
		1000genomes	0.998461	0.993868	0.996159
		AFR	0.998475	0.993671	0.996067
		AMR	0.998472	0.993816	0.996138
		EAS	0.998444	0.993489	0.995961
		EUR	0.998471	0.993808	0.996134
		SAS	0.998464	0.993782	0.996117
HG00733	SNP	Agnostic	0.997700	0.993789	0.995740
		1000genomes	0.997783	0.994116	0.995946
		AFR	0.997783	0.993956	0.995866
		AMR	0.997802	0.993950	0.995873
		EAS	0.997813	0.993409	0.995606
		EUR	0.997813	0.993932	0.995868
		SAS	0.997810	0.993862	0.995832

Bold numbers indicate best performance in each dataset-variant type group
Methods: default DeepVariant (Agnostic), population-aware DeepVariant using allele frequencies from the entire 1000Genomes (1000genomes) and five 1000Genomes superpopulations (AFR, AMR, EAS, EUR and SAS). Higher values correspond to higher accuracy

ISSN: 1471-2105