Strand Bias Artifact

The defining feature is a significant imbalance in the number of forward-strand versus reverse-strand reads supporting the alternative allele. A true heterozygous variant is expected to show a roughly 50/50 split of the variant allele across both strands. An artifact occurs when the variant allele is seen almost exclusively on reads from one orientation, while the reference allele is seen on both. This is quantified using statistical tests like the Fisher's Exact Test or a Strand Odds Ratio (SOR).

The standard statistical method for detecting strand bias is the two-tailed Fisher's Exact Test applied to a 2x2 contingency table:

• Rows: Forward Strand, Reverse Strand
• Columns: Reference Allele Count, Alternative Allele Count A low p-value (typically < 0.001) indicates that the observed distribution is unlikely to occur by chance, flagging the site as a probable artifact. This test is a core component of the GATK StrandBiasBySample annotation.

The Strand Odds Ratio is a more robust metric than the Fisher's Exact Test p-value for high-coverage data. It estimates the ratio of the odds of seeing the alternative allele on the forward strand to the odds on the reverse strand. An SOR value significantly greater than 1.0 indicates bias. Unlike the p-value, the SOR does not become artificially inflated with increasing read depth, making it a preferred filter in modern pipelines like GATK's best practices.

Strand bias is not a single error but a symptom of upstream chemistry or processing failures:

• Oxidative DNA Damage (8-oxoguanine): During library preparation, guanine bases can be oxidized, causing DNA polymerases to misincorporate adenine opposite the lesion, producing G>T transversions predominantly on one strand.
• PCR Amplification Bias: Preferential amplification of one strand during the polymerase chain reaction can create a duplicate-heavy, biased read pile.
• End-Repair Artifacts: Enzymatic steps in library prep can introduce errors at fragment ends, which are then read only from the strand that starts at that end.

It is critical to distinguish strand bias artifacts from genuine biological phenomena that can mimic asymmetry:

• Allele-Specific Expression: True imbalance in transcription, not a sequencing error.
• Imprinting: Parent-of-origin-specific expression.
• Somatic Mosaicism: A true variant present in only a subset of cells. The key differentiator is that biological causes affect the genotype distribution, while strand bias is a read-level technical artifact tied to the sequencing library's physical preparation.

Hard-filtering on strand bias annotations is a standard quality control step:

• GATK: Apply filters using FS > 60.0 (Fisher Strand) and SOR > 3.0 for single nucleotide polymorphisms.
• DeepVariant: The model implicitly learns to down-weight pileup images with visible strand asymmetry, but explicit post-hoc filtering on the StrandBias info field is still recommended.
• Mutect2: Uses a specific strand_bias filter tuned for somatic variant detection, where artifacts are more common due to low tumor purity.

A common sequencing error where the true number of consecutive identical bases is miscounted, leading to false insertion or deletion calls in repetitive genomic regions.

• Homopolymer errors frequently exhibit strand bias because the polymerase slippage mechanism is strand-specific.
• These artifacts are particularly prevalent in pyrosequencing and semiconductor sequencing platforms.
• Local reassembly and haplotype-aware callers are more robust to homopolymer strand bias than alignment-based methods.

Non-uniform sequencing coverage across targeted exonic regions introduced during the hybridization and capture step of whole exome sequencing.

• Capture efficiency varies by probe binding thermodynamics, creating regional strand imbalances.
• GC-rich regions often show strand-specific dropout due to differential denaturation.
• Coverage depth normalization and strand-aware variant callers are required to correct for capture-induced strand bias artifacts.