Germline Variant Calling

The statistical engine of germline calling. For every candidate locus, the model calculates the probability of the observed sequencing data given each possible diploid genotype (e.g., 0/0, 0/1, 1/1). This calculation integrates base quality scores, mapping quality, and allele depth to produce a calibrated likelihood. The genotype with the highest posterior probability, after applying a prior like the Hardy-Weinberg equilibrium, is assigned to the sample.

The process of determining which alleles reside on the same parental chromosome. This is critical for resolving compound heterozygosity. The primary physical evidence comes from paired-end reads or long reads that span two or more heterozygous sites. By observing which alleles co-occur on the same DNA fragment, the algorithm can link them into a haplotype block, transforming an unphased VCF into a biologically accurate representation of chromosomal inheritance.

A machine learning approach to filtering artifacts. Instead of hard-filtering on single metrics, VQSR uses a Gaussian mixture model to evaluate the joint distribution of multiple variant annotations (e.g., strand bias, mapping quality rank sum, read position rank sum). The model is trained on a high-confidence truth set like Genome in a Bottle (GIAB) and assigns a well-calibrated VQSLOD score to every variant, allowing precise control over the false discovery rate.

A population-scale paradigm that simultaneously analyzes all samples in a cohort to resolve genotypes at every locus. This approach leverages population-wide linkage disequilibrium patterns and allele frequencies to refine individual genotype probabilities. It is particularly powerful for calling variants in low-coverage genomes, where a single sample lacks sufficient data, but information can be borrowed from the broader cohort to distinguish a true rare variant from sequencing noise.

The gold standard for validating variant calling accuracy. Hosted by the National Institute of Standards and Technology (NIST), GIAB provides high-confidence variant call sets for well-characterized human genomes (e.g., NA12878). Performance is measured using precision-recall curves across difficult genomic regions. A caller's ability to achieve high recall in Major Histocompatibility Complex (MHC) regions and high precision in low-complexity regions is a definitive test of its robustness.

The core statistical engine of germline calling. For each locus, the model calculates the probability of observing the aligned read data given each of the 10 possible diploid genotypes (AA, AC, AG, AT, CC, CG, CT, GG, GT, TT).

• Bayesian Framework: Combines prior probability of genotypes with sequencing data likelihood
• Error Modeling: Explicitly accounts for base calling errors and mapping uncertainty
• Key Output: Phred-scaled likelihoods (PL field in VCF) that downstream tools use for joint genotyping

The universal standard for storing germline variant calls. A VCF file contains a header with metadata lines and a data section where each row represents a single variant locus.

• GT Field: The called genotype (e.g., 0/1 for heterozygous, 1/1 for homozygous alt)
• GQ Field: Genotype Quality, the Phred-scaled confidence that the called genotype is correct
• AD Field: Allelic depths showing read counts supporting each allele
• INFO Column: Semi-colon delimited annotations like QD (Quality by Depth) and MQ (Mapping Quality)

Determining which alleles reside together on the same parental chromosome. Phasing transforms a list of unlinked heterozygous calls into complete maternal and paternal haplotypes.

• Read-Backed Phasing: Uses paired-end reads spanning multiple heterozygous sites to physically link alleles
• Statistical Phasing: Leverages population haplotype frequencies (e.g., using the Li-Stephens model) when reads don't span multiple variants
• Clinical Relevance: Essential for identifying compound heterozygous mutations where two deleterious variants affect the same gene on opposite chromosomes

A machine learning approach that separates true variants from sequencing artifacts by training a Gaussian mixture model on known truth resources.

• Training Data: Uses high-confidence variant sets like HapMap, 1000 Genomes, and Omni genotyping arrays
• Feature Vector: Each variant is annotated with metrics like strand bias (SB), mapping quality rank sum (MQRankSum), and read position rank sum (ReadPosRankSum)
• Output: A VQSLOD score representing the log-odds of being a true variant, enabling precise false discovery rate control at user-specified sensitivity tranches (e.g., 99.7% for truth sensitivity)

The gold-standard benchmarking resource from NIST that provides highly curated, integrated variant call sets for well-characterized human genomes.

• Reference Samples: Includes the Ashkenazi trio (HG002/HG003/HG004) and the Han Chinese trio (HG005/HG006/HG007)
• Integration Process: Combines calls from multiple technologies (short-read, long-read, linked-read) and orthogonal validation methods to define high-confidence regions
• Benchmarking: Enables precision-recall analysis stratified by variant type (SNP vs. indel) and genomic context (e.g., low-complexity regions vs. segmental duplications)

Simultaneously analyzing a cohort of samples to improve sensitivity for rare variants and resolve genotypes at sites where individual samples have low coverage.

• Information Sharing: A variant observed with weak evidence in one sample can be confidently called if the same allele is strongly supported in related individuals
• gVCF Format: An intermediate file storing genotype likelihoods for every position (including homozygous reference blocks) to enable scalable cohort-wide joint calling
• Population Priors: Incorporates allele frequencies from large reference panels to refine genotype probabilities