VCFcontam: A Machine Learning Approach to Estimate Cross-Sample Contamination from Variant Call Data

The quality of genotyping calls resulting from DNA sequencing is reliant on high quality starting genetic material. One factor that can reduce sample quality and lead to misleading genotyping results is genetic contamination of a sample by another source, such as cells or DNA from another sample of the same or different species. Cross-sample contamination by individuals of the same species is particularly difficult to detect in DNA sequencing data, because the contaminating sequence reads look very similar to those of the intended base sample. We introduce a new method that uses a support vector regression model trained on in silico contaminated datasets to predict empirical contamination using a collection of variables drawn from VCF files, including the fraction of sites that are heterozygous, the fraction of heterozygous sites with imbalanced allele counts, and parameters describing distributions fit to heterozygous allele fractions in a sample. We use the method described here to train a model that can accurately predict the extent of cross-sample contamination within 1% of the actual fraction, for simulated contaminated samples in the 0-5% contamination range, directly from the VCF file. DefinitionsO_ST_ABSLesser alleleC_ST_ABSThe allele in a heterozygous position that received less sequencing read support (which may be either the REF or ALT allele). Lesser allele fraction (LAF)The number of sequencing reads supporting the less frequently observed allele divided by the sum of reads supporting both alleles in the genotype at a given genomic position.