Somatic Variant Classification

The foundational principle of somatic calling relies on a matched tumor-normal pair from the same individual. By comparing the Variant Allele Fraction (VAF) and genotype likelihoods at every locus, the algorithm subtracts the germline baseline. A variant present in the tumor but absent in the normal sample is classified as a somatic mutation. This direct subtraction eliminates the need for population-level frequency databases to filter private germline variants, though it requires precise read-backed phasing to avoid misclassifying low-purity tumor signals as heterozygous germline events.

Somatic classification hinges on the Variant Allele Fraction—the ratio of reads supporting the alternate allele to total depth. Unlike the fixed 50% or 100% ratios expected in germline heterozygosity or homozygosity, somatic VAFs are continuous and influenced by tumor purity and clonal heterogeneity. Advanced callers build statistical models of expected VAF distributions to distinguish true subclonal mutations from strand bias artifacts or homopolymer indel errors. A low VAF in the normal sample often indicates a sequencing error or circulating tumor DNA contamination rather than a true germline event.

The primary challenge is rejecting false positives that mimic somatic mutations. Key discriminative features include:

• Strand Bias: Variant alleles concentrated on a single DNA strand indicate oxidative damage during library preparation.
• Mapping Quality: Low mapping quality scores at the locus suggest mismapped reads or paralogous sequences.
• Base Quality Dropoff: Systematic reduction in base quality scores near the variant position flags sequencing cycle errors.
• Homopolymer Length: Indels in repetitive contexts are weighted against known polymerase slippage error rates. Classifiers integrate these features to assign a somatic probability score.

While single-sample somatic callers operate on one tumor-normal pair, joint genotyping across a cohort of tumors can rescue low-VAF mutations missed in individual samples. By sharing statistical evidence across multiple samples, the model learns locus-specific error profiles and identifies recurrent hotspot mutations. However, this approach risks cross-sample contamination and requires strict batch effect correction. Modern deep learning callers like DeepVariant treat each sample independently but leverage population-informed priors during training to improve rare variant sensitivity.

Simple alignment-based subtraction fails at loci with complex somatic events like multi-nucleotide variants (MNVs) or small insertions adjacent to deletions. Local reassembly performs targeted de novo assembly of reads mapping to a candidate region, constructing a De Bruijn graph to resolve the true haplotypes. This approach distinguishes a single complex somatic event from two adjacent simple germline variants, preventing incorrect classification. The reassembled contigs are realigned to the reference to generate a precise CIGAR string describing the somatic edit.

Modern classifiers output a three-way probability distribution: somatic, germline, or loss of heterozygosity (LOH). LOH events occur when the tumor loses one parental allele, converting a heterozygous germline site to an apparently homozygous somatic state. Distinguishing LOH from a true somatic mutation requires analyzing the B-allele frequency across a haplotype block and integrating copy number variation calls. This multi-class framework prevents the misclassification of copy-neutral LOH as a somatic point mutation.

The foundational experimental design for somatic analysis. By sequencing both cancerous tissue and healthy tissue (often blood or adjacent normal) from the same individual, algorithms can computationally subtract inherited germline variants to isolate acquired somatic mutations. Without a matched normal, population frequency databases must be used, increasing false positive rates for rare population-specific polymorphisms.

A critical quantitative metric defined as the proportion of sequencing reads supporting a variant allele relative to the total read depth at that locus. In somatic calling, VAF is used to:

• Distinguish heterozygous germline variants (~50% VAF) from subclonal somatic mutations (low VAF).
• Infer tumor purity and clonal architecture.
• Identify potential sequencing artifacts, which often exhibit very low VAF and strand bias.

A systematic sequencing error where a variant allele is observed predominantly on reads from one DNA strand (forward or reverse). True biological mutations are expected to be randomly distributed across both strands. Fisher's Exact Test or Wilcoxon rank-sum tests are commonly used to quantify strand bias. A significant skew indicates a technical artifact from oxidative damage during library preparation or sequencing chemistry errors.

The parallel process of identifying inherited genetic variations present in all cells. While somatic callers look for differences between tumor and normal, germline callers analyze a single sample against a reference genome using diploid genotype likelihoods. Accurate germline filtering is the first step in somatic classification; any variant with a high population frequency or a clear heterozygous genotype in the matched normal must be subtracted.

A machine learning technique using a Gaussian Mixture Model to assign a well-calibrated error probability to each variant call. VQSR trains on high-confidence truth sets (like Genome in a Bottle) and multiple annotation features to separate true variants from artifacts. The output is a VQSLOD score—the log odds of being a true variant—which is far more reliable than raw quality scores for filtering somatic candidates.

Statistical procedures applied to variant calling results to limit the expected proportion of false positives. The Benjamini-Hochberg procedure is commonly used to adjust p-values for multiple testing across millions of genomic loci. In somatic analysis, controlling the FDR is critical because the prior probability of a true somatic mutation is extremely low, and the number of tested positions is vast.