DNA layer

Genomics

Reading DNA and finding where it differs. Germline genomics asks which inherited variants explain disease or traits; somatic genomics asks which acquired variants drive cancer and clonal expansion; assembly asks what the genome actually looks like when no reference fits.

Research questions

Which variant explains this patient's rare disease? (diagnostic yield of WES is ~25–40%, WGS adds several percent more)
Which loci are associated with a complex trait, and which variant and gene in each locus is causal?
How much heritability do rare and structural variants explain beyond common SNPs?
Which somatic mutations drive a tumour, in what order, and which therapies do they predict?
What is the full structural diversity of human genomes across ancestries (pangenome)?
Can polygenic scores predict risk equally well across populations?

Technologies

Platform	Read length	Accuracy	Strengths	Weaknesses
Illumina SBS	2×150 bp	>99.9% (Q30+)	Cheapest per base, mature tooling, the reference standard for SNVs	Repeats, SVs, phasing, GC bias
Element AVITI, MGI/Complete Genomics, Ultima	150–300 bp	Q30–Q40+	Competitive cost and accuracy	Similar short-read limits; Ultima has homopolymer-specific errors
PacBio HiFi	15–25 kb	~Q30 (99.9%)	Accurate long reads: SVs, phasing, repeats, assembly, native 5mC	Higher cost per sample
Oxford Nanopore	10 kb to >1 Mb	~Q20+ with recent chemistry and basecallers	Ultra-long reads, real-time, portable, native base modifications, direct RNA	Systematic errors in homopolymers; basecaller version matters
Linked/optical (Bionano)	>150 kb molecules	n/a	Large SVs, scaffolding	No base-level sequence
SNP arrays	n/a	High for common SNPs	Very cheap; basis of biobank GWAS with imputation	Only known variants; rare variants are poorly imputed

How genomic data are generated

SampleBlood (EDTA), saliva, buccal swab, fresh-frozen or FFPE tissue; tumour plus matched normal for somatic work
DNA extractionSilica columns or magnetic beads; high-molecular-weight kits (>50 kb) for long reads
DNA QCConcentration (Qubit), purity (A260/280 ≈ 1.8), integrity (DIN, pulsed-field gel)
LibraryPCR-free WGS; exome capture with probes (~35–60 Mb); amplicon panels; size selection for HiFi/ONT
Sequencing2 × 150 bp to ~30× (germline) or 60–100× tumour; HiFi 15–25 kb; ONT ultra-long
Raw outputFASTQ (short reads), unaligned BAM (PacBio), POD5 signal (ONT)

Arrays follow a different route: DNA is amplified, fragmented, hybridised to probes on a BeadChip and scanned into IDAT intensity files, then genotypes are called from the two allele intensities (GenomeStudio, gtc2vcf) and missing variants are imputed.

Reading aligned sequences

After alignment, every read sits at its position on the reference. A genome browser such as IGV shows this as a pileup: bases that match the reference are drawn as plain bars, and only differences are shown. Real variants appear in a consistent fraction of reads from both strands; sequencing errors appear in single reads at low quality.

A caller summarises this evidence as one line of a VCF file:

#CHROM POS ID REF ALT QUAL FILTER INFO FORMAT SAMPLE chr2 21012345 . C T 812 PASS DP=41;AF=0.5 GT:AD:DP:GQ 0/1:21,20:41:99 │ │ │ │ │ │ │ │ │ └ genotype 0/1 = heterozygous; 21 ref and 20 alt reads; depth 41; genotype quality 99 │ │ │ │ │ │ │ └ site-level annotations └ keys for the sample column │ │ │ │ │ │ └ passed all filters │ │ │ │ │ └ Phred-scaled confidence that a variant exists │ │ │ │ └ alternative allele └ reference allele │ │ │ └ dbSNP identifier (rs…), or "." if none └ chromosome and 1-based position (GRCh38)

The germline short-read pipeline

QC and trimming (FastQC, fastp): adapters, quality, duplication.
Alignment to GRCh38 (BWA-MEM2; minimap2 for long reads). Uses the Burrows–Wheeler transform/FM-index to find exact seed matches quickly, then Smith–Waterman local alignment. Output: SAM/BAM/CRAM with mapping quality (MAPQ = −10 log₁₀ P(mapped position wrong)).
Duplicate marking (Picard, samblaster, GATK MarkDuplicatesSpark): PCR and optical duplicates inflate confidence.
Base quality score recalibration (BQSR): models systematic quality errors by machine cycle and context. Less important with modern instruments and with DeepVariant.
Variant calling: GATK HaplotypeCaller re-assembles local haplotypes with a de Bruijn graph and computes genotype likelihoods with a pair-HMM; DeepVariant encodes read pileups as images and classifies genotypes with a CNN; Strelka2, FreeBayes, Octopus and DRAGEN are alternatives. Cohort calling uses gVCFs and joint genotyping (GATK GenotypeGVCFs, GLnexus).
Filtering: VQSR (Gaussian mixture on annotations) for large cohorts, hard filters for small ones, or learned filters.
Annotation: VEP, ANNOVAR, SnpEff add gene, transcript and consequence, population frequency (gnomAD), predictions (CADD, REVEL, SpliceAI, AlphaMissense) and clinical databases (ClinVar).
Prioritisation: Exomiser and LIRICAL combine variant scores with patient phenotypes (HPO terms); inheritance-mode filters for trios.

Key idea

Variant calling is a statistical inference, not a lookup. A genotype call is the most likely genotype given read evidence, base and mapping qualities and a prior. Report and filter on genotype quality (GQ), depth (DP) and allele balance, and use benchmark regions to know where calls can be trusted.

Structural variants and copy number

Short-read SV callers combine three signals: discordant read pairs (unexpected insert size or orientation), split reads (one read spanning a breakpoint), and read depth. Assembly-based callers reconstruct breakpoint sequence. Manta, DELLY, LUMPY/smoove and GRIDSS2 are widely used, often merged with SURVIVOR or Jasmine. Long-read callers (Sniffles2, cuteSV, pbsv, SVIM) and assembly-based comparison (PAV, SVIM-asm, dipcall) recover far more SVs, especially insertions in repeats. For copy number: CNVkit and GATK gCNV (exomes and panels), CNVnator and Canvas (WGS), ASCAT, FACETS and PURPLE (tumour allele-specific copy number with purity and ploidy), and GISTIC2 (recurrent CNAs across tumours).

Somatic variant calling

Tumour samples are mixtures of tumour and normal cells, often with subclones, so somatic variants may appear at 5% VAF or lower. Use a matched normal, a panel of normals, and germline resource filtering (Mutect2 with FilterMutectCalls, Strelka2, VarDict, LoFreq). FFPE samples cause C>T deamination artefacts, and oxidation causes G>T artefacts; orientation-bias filters address both. Ultra-sensitive applications (minimal residual disease, liquid biopsy) need unique molecular identifiers (UMIs) and duplex sequencing to suppress errors below 10⁻⁶.

De novo assembly and pangenomes

When the reference is incomplete or absent, reads are assembled. Overlap-layout-consensus and string graphs (hifiasm, HiCanu, Flye, Shasta) handle long reads; de Bruijn graphs (SPAdes, MEGAHIT) handle short reads. Hybrid approaches (Verkko) combine HiFi with ultra-long ONT for telomere-to-telomere assemblies; Hi-C or parental data (trio binning) phase haplotypes. Assess with N50, BUSCO completeness, Merqury k-mer QV and Inspector. The Human Pangenome Reference Consortium builds graph references (minigraph-Cactus, PGGB) that reduce reference bias; vg giraffe aligns reads to graphs.

Population and statistical genetics

GWAS: test each variant for association with a trait while controlling population structure. PLINK 2 for QC and simple models; mixed models handle relatedness and structure at biobank scale (REGENIE, SAIGE, BOLT-LMM, fastGWA). Genome-wide significance is conventionally p < 5×10⁻⁸.
Imputation: infer untyped variants from haplotype reference panels (TOPMed, 1000 Genomes, HRC) with Beagle 5, Minimac4 or IMPUTE5; pre-phasing with SHAPEIT5 or Eagle2.
Heritability and genetic correlation: LD score regression (LDSC), GCTA-GREML, S-LDSC for enrichment in annotations (cell-type-specific enhancers).
Fine-mapping: SuSiE, FINEMAP, PolyFun (functional priors) output credible sets of likely causal variants.
Polygenic scores: PRS-CS, LDpred2, SBayesRC, PRS-CSx (multi-ancestry). Evaluate with incremental R² or AUC in held-out cohorts of each ancestry.
Rare-variant burden tests: SKAT-O, ACAT, gene-based burden in exome biobanks.

Tools

Benchmarks

Pitfall

Benchmarks such as Genome in a Bottle cover "high-confidence" regions, which historically excluded the hardest parts of the genome. A pipeline scoring 99.9% F1 there may do much worse in segmental duplications, the MHC or medically relevant genes with paralogs (e.g. SMN1/2, PMS2, CYP2D6). Check stratified performance (GIAB stratification BEDs with hap.py or vcfdist), not only the headline number.

Research gap

Variant calling against pangenome graphs, and benchmarks that represent diverse ancestries.
Joint calling of SNVs, SVs, repeat expansions and methylation from one long-read run.
Interpretation of non-coding and structural variants in rare disease, where most unsolved cases may sit.
Polygenic score portability, calibration and clinical utility across ancestries.
Somatic mosaicism and clonal haematopoiesis detection in normal tissues.