- Variant calling is a fundamental process in bioinformatics used to identify genetic differences between a biological sample and a reference genome or between multiple samples. These differences include single-nucleotide variants (SNVs), insertions and deletions (indels), and larger structural variants (SVs), such as deletions, duplications, inversions, and translocations. Variant calling converts sequencing data into a set of genomic variants that can be used in clinical genetics, cancer research, population genetics, evolutionary studies, agricultural breeding, and microbial genomics. The two major sequencing approaches used for this purpose are short-read sequencing and long-read sequencing. Although both approaches aim to detect genetic variation, they differ in read length, error characteristics, alignment strategies, variant-calling algorithms, and the types of variants they can identify reliably.
- Short-read sequencing generally produces DNA fragments ranging from approximately 50 to 300 bases, depending on the sequencing platform and protocol. Illumina sequencing is a widely used short-read technology because it provides high-throughput data with relatively low per-base error rates. Short reads are particularly effective for detecting SNVs and small indels in genomic regions that can be mapped unambiguously to a reference genome. However, repetitive sequences, duplicated genes, segmental duplications, and complex genomic regions can make short-read alignment difficult. When reads cannot be assigned confidently to a unique genomic location, variant detection may become less reliable.
- Long-read sequencing produces substantially longer DNA sequences, often thousands or tens of thousands of bases. Major technologies include Pacific Biosciences (PacBio) Single-Molecule Real-Time (SMRT) sequencing, particularly HiFi sequencing, and Oxford Nanopore Technologies (ONT) sequencing. Long reads can span repetitive regions, structural rearrangements, and complex genomic regions that are difficult to resolve using short reads. They are particularly valuable for detecting structural variants, resolving complex indels, identifying repeat expansions, and analysing haplotypes. PacBio HiFi reads generally combine long read lengths with high consensus accuracy, while Oxford Nanopore reads can provide very long sequences and real-time data, with accuracy depending on the chemistry, basecalling model, and sequencing workflow.
- The variant-calling process usually begins with sequencing data in FASTQ format. Each FASTQ record contains a read identifier, nucleotide sequence, and corresponding quality scores. Before alignment, raw reads should be examined for sequencing quality, adapter contamination, low-quality bases, unusual sequence composition, and other technical problems. Short-read quality assessment is commonly performed using FastQC and MultiQC, while fastp and Cutadapt can be used for quality filtering and adapter removal. Long-read data can be assessed using tools such as NanoPlot, pycoQC, and LongQC, depending on the sequencing platform and workflow. Long-read preprocessing may include filtering by read quality and length, removing adapters, and identifying or removing contaminating sequences. Filtering should be appropriate for the technology and biological objective because excessively strict thresholds can remove useful reads.
- After quality control, reads are generally aligned to a reference genome. Short-read alignment tools include BWA-MEM2 and Bowtie2, although their suitability depends on the sequencing data and analysis objective. BWA-MEM2 is commonly used for genomic short-read alignment, while Bowtie2 supports a range of short-read mapping applications. Long-read alignment is commonly performed using minimap2, which supports different presets for PacBio and Oxford Nanopore data. The choice of preset and parameters is important because read characteristics differ between sequencing technologies. Alignment produces files in SAM format, which are commonly converted into compressed BAM or CRAM formats. SAMtools is widely used for sorting, indexing, and examining alignment files.
- An illustrative short-read alignment workflow using BWA-MEM2 and SAMtools is shown below:
# Align paired-end short reads to a reference genome
bwa-mem2 mem reference.fa reads_R1.fastq.gz reads_R2.fastq.gz > aligned.sam
# Convert SAM to BAM
samtools view -bS aligned.sam > aligned.bam
# Sort the alignment file
samtools sort aligned.bam -o aligned.sorted.bam
# Index the sorted BAM file
samtools index aligned.sorted.bam
# Generate alignment statistics
- samtools flagstat aligned.sorted.bam
- For long-read data, minimap2 can be used with a preset appropriate to the sequencing technology. The following example illustrates alignment of Oxford Nanopore reads:
# Align Oxford Nanopore reads to a reference genome
minimap2 -ax map-ont reference.fa nanopore_reads.fastq.gz > aligned.sam
# Convert SAM to BAM
samtools view -bS aligned.sam > aligned.bam
# Sort and index the BAM file
samtools sort aligned.bam -o aligned.sorted.bam
samtools index aligned.sorted.bam
For PacBio HiFi reads, minimap2 provides a different preset:
# Align PacBio HiFi reads
minimap2 -ax map-hifi reference.fa hifi_reads.fastq.gz > hifi_aligned.sam
# Convert, sort, and index the alignment
samtools view -bS hifi_aligned.sam | samtools sort -o hifi_aligned.sorted.bam
samtools index hifi_aligned.sorted.bam
- These examples illustrate common alignment steps rather than a complete production pipeline. Reference indexes, read-group information, sample metadata, software versions, and alignment parameters should be managed consistently. In particular, the reference genome assembly and its sequence identifiers must match those used in downstream analysis.
- Variant calling from short reads is commonly performed using local assembly, probabilistic models, or combinations of read alignment and haplotype reconstruction. GATK HaplotypeCaller is a widely used tool for germline SNV and small-indel discovery. It identifies active genomic regions, locally assembles candidate haplotypes, and evaluates read evidence to estimate genotypes. FreeBayes uses a Bayesian framework to identify variants and estimate genotypes from aligned reads. BCFtools call provides another widely used approach for variant calling from pileup or likelihood information. DeepVariant applies a deep-learning approach to classify candidate variants using read-alignment evidence represented in an image-like format. These tools differ in their algorithms, assumptions, supported data types, and recommended workflows; they should not be treated as interchangeable without considering the study design.
- An illustrative GATK HaplotypeCaller workflow is:
# Identify germline variants from aligned short reads
gatk HaplotypeCaller \
-R reference.fa \
-I sample.sorted.bam \
-O sample.g.vcf.gz \
-ERC GVCF
- The gVCF output contains variant and reference-confidence information and is commonly used in workflows that perform joint genotyping across multiple samples. For a single-sample workflow, HaplotypeCaller can also produce a regular VCF when configured accordingly. Joint genotyping can be performed with GATK GenomicsDBImport and GenotypeGVCFs, depending on the study design.
- An example using BCFtools is:
# Generate genotype likelihoods and call variants
bcftools mpileup -f reference.fa -Ou sample.sorted.bam | \
bcftools call -mv -Oz -o sample.variants.vcf.gz
# Index the VCF file
bcftools index sample.variants.vcf.gz
- The -m option requests multiallelic caller output, while -v requests variant sites. This is a simplified example; real analyses should consider ploidy, sample type, mapping quality, base quality, read groups, and the recommended settings for the particular dataset.
- Long-read variant calling uses algorithms designed to interpret longer alignments and the error patterns of long-read technologies. Long reads can provide direct evidence across variant breakpoints, making them particularly useful for detecting structural variants and complex genomic changes. For SNVs and small indels, callers can use read-level evidence, local realignment, haplotype information, or neural-network models. PacBio HiFi data can be analysed using tools such as DeepVariant and Clair3, depending on the supported workflow and software version. Clair3 supports long-read variant calling and has been used with both Oxford Nanopore and PacBio data. Long-read variant-calling tools should be selected according to the platform, chemistry, basecalling or consensus accuracy, and intended variant classes.
- Structural variant calling is a major strength of long-read sequencing. Tools such as Sniffles2 and cuteSV identify structural variants using long-read alignment patterns, including split reads, large insertions and deletions, and other signals around genomic breakpoints. SVIM is another tool developed for structural variant discovery from long-read alignments. These tools can identify deletions, insertions, duplications, inversions, and other structural changes, although the types of events and their accuracy depend on read length, coverage, genome complexity, and the calling algorithm. For short-read data, tools such as Manta, DELLY, GRIDSS, and LUMPY use paired-end, split-read, and related alignment evidence to identify structural variants. Short-read methods can perform well for many events but often struggle to resolve breakpoints within repetitive or highly complex regions.
- A simplified long-read structural variant calling example using Sniffles2 is:
# Call structural variants from long-read alignments
sniffles \
–input nanopore_aligned.sorted.bam \
–vcf nanopore_structural_variants.vcf \
–reference reference.fa
- The exact command-line options depend on the Sniffles2 version and the analysis configuration. For multi-sample analysis, Sniffles2 also provides population-oriented workflows that can combine evidence across samples. Researchers should consult the relevant software documentation and record all parameters used.
- Variant callers generally produce results in Variant Call Format (VCF) or its compressed binary representation, BCF. A VCF file contains metadata describing the reference and calling process, followed by a header and variant records. Each record typically includes the chromosome, genomic position, reference allele, alternate allele, quality score, filter status, information fields, and sample genotype information. A simplified VCF record may look like this:
#CHROM POS ID REF ALT QUAL FILTER INFO FORMAT SAMPLE
chr1 10500 . A G 60 PASS . GT:DP 0/1:35
- In this example, the reference allele is A and the alternate allele is G. The genotype 0/1 indicates a heterozygous genotype in a typical diploid setting, and DP=35 indicates a read depth of 35 under the caller’s definition. The quality score and filter status provide additional information about the confidence of the call. Actual VCF records may contain many additional annotations, including allele depth, genotype quality, mapping quality, strand-bias statistics, and variant-specific information.
- Variant quality assessment is essential because sequencing errors, alignment problems, low coverage, contamination, and systematic technical biases can produce false-positive and false-negative calls. Important metrics include read depth, genotype quality, variant quality, mapping quality, allele balance, strand bias, read-position bias, and the number of supporting reads. For a heterozygous variant in a diploid sample, the alternate allele fraction may be expected to be near 0.5 under ideal conditions, but biological factors, sequencing bias, copy-number changes, and sampling variation can cause substantial deviations. In cancer samples, tumour purity, subclonality, and local copy number can further affect allele fractions. Consequently, quality thresholds should be selected according to the organism, sequencing technology, ploidy, sample type, and study objective.
- Variant filtering may be performed using caller-specific filters, hard thresholds, statistical models, or machine-learning approaches. GATK workflows may use variant quality score recalibration when sufficient training data and suitable resources are available. Other workflows may use carefully selected thresholds for depth, genotype quality, allele balance, and mapping quality. BCFtools filter can be used to apply explicit filtering expressions. For example:
# Retain variants with a quality score of at least 30
bcftools filter -i ‘QUAL>=30’ \
sample.variants.vcf.gz \
-Oz -o sample.filtered.vcf.gz
# Index the filtered VCF file
bcftools index sample.filtered.vcf.gz
- This example applies only one filtering criterion and should not be interpreted as a complete filtering strategy. Filtering too aggressively can remove genuine variants, while permissive filtering can retain false positives. The balance depends on the purpose of the analysis and the quality characteristics of the data.
- Variant normalization is another important step in producing consistent variant representations. Insertions and deletions can sometimes be represented in more than one way within repetitive or low-complexity regions. Normalization commonly includes left alignment and representation of alleles in a consistent form. BCFtools norm can be used for this purpose:
# Normalize variants against the reference genome
bcftools norm \
-f reference.fa \
-m -any \
sample.filtered.vcf.gz \
-Oz -o sample.normalized.vcf.gz
# Index the normalized VCF file
bcftools index sample.normalized.vcf.gz
- Normalization improves consistency when comparing variant records across samples or databases. It does not guarantee that two biologically equivalent variants will be represented identically in every possible reference or transcript context. Reference assembly, chromosome naming, allele representation, and normalization conventions must therefore be considered when comparing datasets.
- Short-read and long-read sequencing provide complementary strengths. Short reads generally offer high per-base accuracy, high throughput, and established workflows for SNV and small-indel calling. They are often cost-effective for large cohorts and population-scale studies. However, their limited length can make it difficult to map reads uniquely in repetitive regions and to resolve complex structural variants. Long reads can span large genomic regions, improve haplotype resolution, and provide more direct evidence for structural variation and repeat expansions. Their performance depends on the sequencing platform and data quality, and some long-read datasets may have different error profiles or coverage characteristics from short-read datasets. PacBio HiFi reads generally provide high consensus accuracy, whereas Oxford Nanopore sequencing offers flexible read lengths and can generate ultra-long reads, with accuracy dependent on chemistry and basecalling.
- Coverage is an important consideration in both approaches. Higher sequencing depth can improve the detection of low-frequency variants and support more reliable genotype estimates, but it does not automatically eliminate systematic errors or mapping problems. Short-read germline variant discovery commonly uses relatively high coverage, while long-read workflows may require different coverage levels depending on the platform, variant class, and calling method. Somatic variant detection may require deeper sequencing because mutations can be present in only a fraction of cells. For structural variants, read length and the number of reads spanning breakpoints can be as important as overall depth.
- Combining short-read and long-read data can improve variant discovery. Short reads can provide strong evidence for SNVs and small indels, while long reads can help resolve structural variants, repetitive regions, and haplotypes. A combined strategy may use independently generated variant calls followed by comparison, evidence integration, or validation. However, variant records from different callers should not simply be merged without checking representation, genotype compatibility, reference assembly, and confidence. Differences in alignment, normalization, ploidy assumptions, and caller models can produce apparently different calls for the same underlying event.
- Variant phasing is another important application of long-read sequencing. Phasing determines which variants occur together on the same chromosome or haplotype. Because a long read may span multiple heterozygous sites, long-read data can directly connect variants across larger genomic distances than typical short reads. Short-read phasing is also possible using read pairs, statistical methods, family information, or specialised algorithms, but the distance over which variants can be linked is often more limited. Phased variants are useful for identifying compound heterozygosity, analysing disease-associated haplotypes, studying allele-specific expression, and understanding the structure of complex genomic regions.
- Variant annotation follows calling and filtering. Annotation tools add biological information about the possible consequences of variants, including effects on coding sequences, splice sites, regulatory regions, protein function, and known disease associations. Common tools include Ensembl Variant Effect Predictor (VEP), SnpEff, and ANNOVAR. Variant records may also be compared with resources such as dbSNP, ClinVar, gnomAD, and the GWAS Catalog, where relevant. Database matches should not be treated as automatic evidence of pathogenicity or functional effect. Interpretation requires consideration of population frequency, clinical evidence, gene–disease validity, inheritance pattern, phenotype, and the limitations of the underlying data.
- The performance of a variant-calling pipeline should be evaluated using appropriate reference datasets and validation methods. Precision measures the proportion of called variants that are true positives, while recall measures the proportion of true variants that are detected. The F1 score combines precision and recall into a single measure. Benchmarking tools such as hap.py can compare small-variant call sets with truth sets, while structural variant benchmarking can use tools such as Truvari. Reference materials from the Genome in a Bottle (GIAB) project are widely used for benchmarking human germline variant-calling workflows, although their regions and variant classes do not represent every genomic context or organism. Orthogonal validation may involve Sanger sequencing, targeted sequencing, PCR, digital PCR, or other suitable experimental approaches.
- Variant calling has broad applications across biomedical and biological research. In clinical genetics, it supports the identification of candidate variants associated with inherited disorders, although clinical interpretation requires additional evidence and professional assessment. In cancer genomics, somatic variant calling helps identify mutations, copy-number changes, and structural rearrangements that may contribute to tumour development or treatment response. In population genetics, variant datasets are used to investigate genetic diversity, population structure, demographic history, and natural selection. In agriculture and animal breeding, variant discovery supports marker development, genomic selection, trait mapping, and the identification of genetic changes associated with production, health, and adaptation. In microbial genomics, variant calling can help study pathogen evolution, strain relationships, antimicrobial resistance, and outbreak dynamics.
- Several limitations must be considered when interpreting variant calls. Reference bias can reduce detection of alleles that differ substantially from the reference genome. Repetitive sequences, segmental duplications, low-complexity regions, and poorly assembled genomic regions can lead to ambiguous alignments. Low coverage can cause missed variants, while sequencing errors and systematic biases can produce false positives. Differences in sequencing chemistry, library preparation, basecalling, alignment algorithms, and variant-calling models can affect results. Structural variants are especially challenging when breakpoints lie in repetitive regions or when multiple rearrangements occur together. Some variant classes, including repeat expansions, complex rearrangements, and mosaic variants, may require specialised methods rather than standard small-variant callers.
- Reproducibility is essential in variant-calling research. Researchers should record the reference genome assembly and version, sequencing platform, library preparation, basecalling software where relevant, read preprocessing steps, aligner and parameters, variant caller and version, filtering criteria, normalization procedures, and annotation resources. Sample identifiers and metadata should be managed carefully, particularly in clinical and human-genomics studies where genetic data are sensitive. Pipeline automation using workflow systems such as Nextflow, Snakemake, or Cromwell can improve consistency, traceability, and scalability. Software environments can be managed with Conda or containers, and intermediate files and logs should be retained according to project requirements.
- Variant calling from short and long reads is a central component of modern genome analysis. Short-read sequencing provides an efficient and accurate approach for many SNVs and small indels, while long-read sequencing improves the detection and interpretation of structural variants, repeat-associated changes, and haplotypes. Selecting an appropriate workflow requires understanding the sequencing technology, read characteristics, reference genome, variant classes, sample type, and analytical objective. Combining complementary technologies, validating important calls, and documenting computational procedures can improve the reliability and biological value of variant-discovery studies.