Diploid Genotype Likelihood

Formally expressed as P(Data | Genotype), the likelihood function calculates the probability of observing the aligned reads at a locus given a specific combination of two alleles. It does not calculate the probability of the genotype being true; it calculates how well the genotype explains the data. The calculation integrates base quality scores, mapping quality scores, and the reference allele to model the error profile of the sequencing experiment.

For a diploid organism with reference allele R and alternate allele A, the likelihood calculation evaluates all 10 possible unordered genotypes:

• Homozygous Reference: R/R
• Heterozygous: R/A
• Homozygous Alternate: A/A
• Plus all combinations when multiple alternate alleles exist at a locus. The genotype with the maximum likelihood is the initial call, but the full likelihood distribution is retained for downstream Bayesian genotyping.

Raw likelihoods are transformed into Phred-scaled values for compact storage in VCF/BCF files:

• PL field: Phred-scaled genotype likelihoods, normalized so the most likely genotype has a PL of 0.
• Formula: PL = -10 * log10(P(Data | Genotype))
• A PL difference of 10 between two genotypes means the data are 10 times more likely under the higher-ranked genotype. This logarithmic scale prevents numerical underflow and enables integer encoding.

The likelihood calculation explicitly models multiple sources of uncertainty:

• Sequencing Error: The probability that a base call is incorrect, derived from the per-base quality score.
• Mapping Error: The probability that the read is aligned to the wrong genomic location, derived from the mapping quality.
• Allele Bias: Systematic over- or under-representation of alleles due to PCR amplification or capture efficiency. These error probabilities are marginalized over all possible true underlying bases to produce a robust likelihood.

Genotype likelihoods alone are frequentist; they do not incorporate prior knowledge. Modern callers combine likelihoods with a population prior (e.g., allele frequency from gnomAD) using Bayes' theorem:

• Posterior ∝ Likelihood × Prior
• This transforms the raw P(Data | Genotype) into the clinically actionable P(Genotype | Data).
• A common prior is the Hardy-Weinberg equilibrium expectation, which provides the expected frequency of each genotype given the population allele frequency.

For each candidate genotype, the likelihood is computed by summing over all possible haplotype configurations of the aligned reads:

• Each read is treated as an independent observation drawn from one of the two haplotypes.
• The algorithm marginalizes over the unknown read-haplotype assignment.
• For a heterozygous genotype R/A, the likelihood is the product over reads of 0.5 * P(read | R) + 0.5 * P(read | A), reflecting the equal probability of sampling from either haplotype.

The computational process of identifying inherited genetic variations present in all cells of an individual. This analysis operates under a diploid assumption, meaning the likelihood calculation must evaluate the probability of observing the sequencing data given all possible combinations of two alleles—homozygous reference (0/0), heterozygous (0/1), and homozygous alternate (1/1)—at each genomic locus.

The proportion of sequencing reads supporting a variant allele relative to the total read depth at a given locus. In a clean diploid sample, VAF serves as a critical prior for genotype likelihood calculations:

• Heterozygous (0/1): Expect VAF ≈ 50%
• Homozygous alternate (1/1): Expect VAF ≈ 100%
• Artifacts or somatic mosaicism: Often present with low VAF, requiring careful statistical modeling to distinguish from true heterozygous calls.

A preprocessing step that applies machine learning to correct systematic errors in the per-base quality scores reported by the sequencing instrument. Accurate Phred-scaled error probabilities are essential inputs to the diploid genotype likelihood model, as inflated quality scores can cause the model to overconfidently call false heterozygous variants, while deflated scores reduce sensitivity.

The computational process of determining which alleles on a single chromosome are inherited together from the same parent. While genotype likelihoods estimate the unphased diploid state (e.g., A/G vs G/A), phasing resolves the arrangement of variants along homologous chromosomes. Read-backed phasing uses paired-end reads spanning multiple heterozygous sites to physically link alleles, refining the likelihood calculation for adjacent variants.

The standardized text file format for storing gene sequence variations. The GT (Genotype) field encodes the diploid call, while the PL (Phred-scaled Likelihoods) field stores the normalized genotype likelihoods for all possible allele combinations. A typical PL entry like PL=0,30,100 indicates strong support for a homozygous reference call (0/0) over heterozygous (0/1) or homozygous alternate (1/1) states.

A systematic sequencing error where a variant allele is observed predominantly on reads from one DNA strand. Diploid genotype likelihood models must account for this artifact, as a true heterozygous variant should show roughly equal allele representation on both the forward and reverse strands. Failure to model strand bias inflates false positive heterozygous calls, particularly in regions with oxidative damage during library preparation.