Somatic Variant Caller

The foundational comparative engine that performs a locus-by-locus subtraction of the matched normal sample genotype from the tumor sample. This algorithm explicitly models the joint probability of genotypes under both somatic and germline hypotheses, using a Bayesian framework to calculate the posterior probability that a variant is truly somatic. It rejects artifacts by requiring the variant to be absent in the normal and supported by high-confidence alternate allele reads in the tumor.

Leverages Unique Molecular Identifiers (UMIs) and duplex sequencing biochemistry to collapse PCR duplicates into a single consensus read. The algorithm groups reads by their UMI, aligns them, and requires concordant evidence from both the forward and reverse strands. This single-molecule consensus process mathematically eliminates random polymerase errors and base oxidation damage, pushing the reliable limit of detection below 0.1% Variant Allele Frequency.

A statistical noise model built from a curated cohort of healthy individuals sequenced on the same platform. The algorithm identifies recurrent technical artifacts—such as systematic sequencing errors in homopolymer regions or oxidized-guanine hotspots—that appear across normal samples. By flagging and removing these panel-wide noise signatures, the caller dramatically reduces false positives without sacrificing sensitivity to rare true variants.

A specialized filter that prevents misattribution of age-related clonal expansions in blood cells as tumor-derived mutations. The algorithm cross-references detected variants against known CHIP-associated genes (e.g., DNMT3A, TET2, ASXL1) and can integrate a matched buffy coat or peripheral blood mononuclear cell control. This ensures that a liquid biopsy finding reflects the solid tumor biology, not benign hematopoietic noise.

A machine learning step that empirically adjusts the per-base quality scores emitted by the sequencer. The algorithm models covariates of error—including read group, sequencing cycle, and preceding dinucleotide context—to generate accurate, platform-aware base quality estimates. This recalibration is critical for preventing high-confidence false positives caused by systematic under-estimation of error rates in specific sequence contexts.

Integrates the physical properties of cell-free DNA to validate the somatic origin of a variant. The algorithm analyzes fragment length distribution, end motif frequencies, and nucleosome footprinting around the variant locus. Tumor-derived ctDNA fragments are typically shorter than those from healthy cells. A variant supported by fragments with a tumor-like fragmentation profile receives a higher confidence score, adding an orthogonal layer of evidence beyond sequence alone.

The tumor-derived fraction of cell-free DNA (cfDNA) carrying cancer-specific somatic mutations. ctDNA typically constitutes <0.1% to >10% of total cfDNA, making it the critical analyte for liquid biopsy. The somatic variant caller must distinguish these rare tumor molecules from the overwhelming wild-type background. Key characteristics:

• Fragment length peaks at ~166 bp (nucleosomal protection)
• Half-life of 16 minutes to 2.5 hours in circulation
• Concentration correlates with tumor burden and disease stage

A random nucleotide barcode (typically 8-12 bp) ligated to individual DNA molecules before PCR amplification. UMIs enable computational deduplication by grouping reads sharing the same barcode, allowing the somatic variant caller to:

• Collapse PCR duplicates into a single consensus read
• Correct polymerase errors through family-level majority voting
• Quantify absolute input molecules rather than amplified copies
• Achieve error rates below 10⁻⁶ when combined with duplex strategies

The proportion of sequencing reads supporting a variant allele at a given locus: VAF = alt_reads / (alt_reads + ref_reads). In somatic variant calling, VAF is the primary signal for distinguishing:

• Somatic mutations (typically low VAF, often <5% in liquid biopsy)
• Germline variants (clustered near 50% or 100% VAF)
• Sequencing artifacts (ultra-low VAF with strand bias)
• Clonal hematopoiesis (intermediate VAF from blood-derived mutations) Accurate VAF estimation requires correction for local copy number and tumor purity.

The computational subtraction of inherited polymorphisms from a somatic variant callset. Strategies include:

• Matched normal subtraction: Direct comparison against the patient's own non-tumor tissue (gold standard)
• Population database filtering: Removal of variants present in gnomAD, 1000 Genomes, or dbSNP at population frequency >1%
• Panel of Normals (PoN): A curated set of healthy controls sequenced on the same platform to model and suppress recurrent technical artifacts Without rigorous germline filtering, a somatic variant caller will report thousands of false positives from benign polymorphisms.

A critical filter that removes somatic variants originating from age-related clonal expansions in hematopoietic stem cells rather than from a solid tumor. These variants appear in cfDNA because blood cells contribute significantly to the total cfDNA pool. Detection methods:

• Matched buffy coat sequencing: Directly sequencing the patient's white blood cell DNA
• Mutational signature analysis: CH variants often carry DNMT3A, TET2, or ASXL1 mutations
• Fragment length analysis: Hematopoietic-derived cfDNA shows distinct fragmentation patterns Failure to filter CH variants leads to false-positive tumor calls in early cancer detection.

A machine learning process that adjusts the per-base quality scores emitted by the sequencer using empirically observed error covariates. BQR models the relationship between reported quality and actual error rates across:

• Sequencing cycle (errors increase with read length)
• Dinucleotide context (specific motifs have elevated error rates)
• Machine tile position (spatial artifacts on the flow cell) The recalibrated quality scores provide the somatic variant caller with calibrated error probabilities, dramatically reducing false-positive variant calls from systematic sequencing noise.