In the context of contemporary advancements in genomics and precision medicine, Variant Calling (VC) has emerged as a pivotal element in elucidating individual genetic variations and facilitating disease diagnosis and treatment. The accurate identification of a given variant, whether it is a single nucleotide variant (SNV), an insertion-deletion (indel), or a structural variant (SV), relies on a series of complex and detailed processes. These processes include high-quality sequencing, comparison, variant identification, and filtering. Due to the continuous iteration of sequencing technologies and algorithmic tools, the accuracy and sensitivity of variant detection are constantly improving. The field has evolved from short-read-long sequencing platforms (e.g., Illumina) to long-read-long platforms (e.g., PacBio and Oxford Nanopore), from classic BWA-MEM comparison to variant identification tools such as GATK, Strelka2, DRAGEN, etc., and to assays optimized using deep learning. The optimization of each step is driving precision medicine and genomic research forward. Concurrently, variant detection confronts distinct technical challenges in germline and somatic applications, including low-frequency variant detection, tumor heterogeneity, and reference genome bias. These challenges have prompted researchers to perpetually explore novel solutions, such as single-cell sequencing, whole-genome mapping, and the integration of artificial intelligence models.
This article systematically examines the entire process, from raw sequencing data to the identification of high-confidence genomic variants. It compares the performance characteristics of mainstream variant detection tools and summarizes the key factors and optimization strategies affecting detection accuracy. This article provides practical references for scientific research and clinical applications. The article then provides a synopsis of the variant detection process, a comparison of mainstream tools, and an examination of accuracy challenges.
Foundations of Variant Calling Workflow
Variant detection is a critical component of genomics and precision medicine research. The objective of variant detection is to identify genetic differences in a sample relative to a reference genome. These genetic differences may include SNVs, indels, and SVs. The standard process involves a series of steps, including sequencing, matching, variant identification, and result filtering. Each of these steps contributes to the accuracy and sensitivity of the final test.
Process overview: sequencing, alignment, and variant detection
The first step in mutation detection is to acquire high-quality sequencing data. Short read-length platforms such as Illumina are widely used for their high accuracy and are suitable for point mutation and small indel detection, while long read-length platforms such as PacBio and Oxford Nanopore are suitable for detecting structural variants and large fragment insertions due to their ability to span repetitive regions and complex structures. After sequencing, tools such as FastQC and Trim Galore are used for quality assessment and clipping to remove junctions and low-quality sequences.
The cleaned reads need to be aligned to the reference genome. Commonly used alignment tools such as BWA-MEM (for short read lengths) and Minimap2 (for long read lengths) are used to output SAM/BAM files. This is followed by sorting, de-duplication (Picard MarkDuplicates), and quality calibration (GATK BQSR) to enhance the accuracy of subsequent analysis. For indel regions, local re-matching is also performed to correct matching errors due to insertions/missings.
Best Practices workflow for variant discovery in germline DNA sequence (Poplin et al., 2017)
On the basis of the comparison data, variant detection tools are used to identify variants. GATK HaplotypeCaller improves detection accuracy by local reassembly and is suitable for germline samples; FreeBayes handles multiple samples; and DeepVariant employs deep learning to improve the recognition of complex regions. After calling is complete, hard filtering or variant quality recalibration (VQSR) needs to be applied to remove false positives. Ultimately, VCF files are output for downstream annotation and functional analysis.
Germline and somatic cell variation detection
Germline variants are found in germ cells and whole body cells, and are common in individual or family studies, with a frequency of 50% or 100%. The test focuses on accuracy and consistency, and it is recommended to use tools such as HaplotypeCaller, FreeBayes, etc., and refer to dbSNP, gnomAD, and other databases to remove common polymorphisms. Usually, 30× depth of coverage is sufficient.
Inherited (germline) genomics variants vs acquired (somatic) variants (Alldredge et al, 2019)
Somatic variants, on the other hand, are acquired during individual development or disease and are especially common in tumors, where mutation frequency is low and accompanied by clonal heterogeneity. Detection usually requires paired tumor-normal samples, the use of specialized tools such as Mutect2, Strelka2, and VarScan2, and the reduction of false positives by a panel of normals or deep learning methods, and higher sequencing depths are recommended for detection.
In terms of technical challenges, germline variants mainly need to distinguish between true rare variants and sequencing noise, while somatic cell variants need to solve the problems of low-frequency detection, tumor heterogeneity, and sample contamination. In recent years, long read-length sequencing, single-cell sequencing, and deep learning models (e.g., NeuSomatic) have continuously improved the sensitivity and accuracy of somatic mutation detection.
Comparing Variant Calling Tools
In the era of high-throughput sequencing, accurate detection of genomic variants is the key. This chapter compares the current mainstream variant detection tools: GATK, FreeBayes, Strelka2, and Illumina DRAGEN, firstly introducing the core algorithms, application scenarios, and advantages of each tool, then evaluating their performances in terms of consistency, speed, and accuracy based on benchmarking, and finally summarizing the features and future trends of each tool.
Analysis of Mainstream Mutation Detection Tools
Currently, the mainstream variant detection tools include GATK, FreeBayes, Strelka2, and Illumina DRAGEN, each of which has its characteristics in terms of algorithm design, application scenarios, and performance.
GATK (Genome Analysis Toolkit), developed by Broad Institute, is one of the most widely used mutation detection tools. Its core algorithm is based on HaplotypeCaller, which combines local de novo assembly and Bayesian modeling to reconstruct and score candidate regions, thus improving the accuracy of mutation detection.GATK is particularly suitable for large-scale whole-genome or exome sequencing analyses and is able to efficiently detect SNVs and Indel, and provide a set of standardized analysis processes (e.g. BQSR). Its main advantages are its comprehensive functionality, active community, and stable results, making it one of the most commonly used standard tools in research and clinical studies.
Flowchart of the variant calling workfow (Bathke et al., 2021)
FreeBayes is a flexible open-source variant detector that supports multi-sample coanalysis and modeling of complex isoforms (including polyploids). Its algorithm uses hash indexing and Bayesian inference for local reconstruction and probability estimation of read segments and can better handle population samples and structurally complex mutation regions.FreeBayes has the advantage of being flexible and scalable and is suitable for personalized analysis or non-human species research. However, it is less sensitive to mutation identification in low-coverage regions, and the default parameters may miss some of the true variants, requiring the user to optimize the parameters according to the data characteristics.
Strelka2 is a highly efficient mutation detection tool from Illumina specifically designed for tumor-normal pair analysis in cancer research. Combining Bayesian statistics with local splicing methods, it can detect both SNV and Indel and is highly sensitive to low allele frequency variants.Strelka2 excels in handling low-frequency somatic mutations, making it suitable for high-precision scenarios such as cancer genomics and clinical sequencing. Its strengths also include well-optimized algorithms, fast operation speed, low false alarm rate, and high accuracy in multiple benchmark tests.
Illumina DRAGEN is a field-programmable gate array (FPGA) hardware-accelerated variant detection platform that provides an end-to-end solution from read segment comparison to variant calling. Its algorithms implement graph splicing and parallel computation on a dedicated chip, making it much faster than traditional software tools, typically completing genome-wide analyses in minutes to hours.DRAGEN's high accuracy, especially in indel detection, makes it ideal for use in clinical and large-scale genomic projects that require rapid response. Its drawback is its reliance on proprietary hardware, which limits deployment flexibility in general-purpose server or cloud environments.
Tool performance comparison: consistency, speed, and accuracy
Benchmarking results show that GATK and DRAGEN typically perform best in SNV and Indel detection, with F-scores up to high levels.Strelka2 has a clear advantage in somatic mutation detection and is particularly sensitive to mutations at low allele frequencies.FreeBayes is slightly less accurate overall, but sensitivity and specificity can be weighed by tuning the parameters. The results of GATK and DRAGEN were highly reproducible and stable concerning parameter and input data quality dependence, whereas FreeBayes results fluctuated greatly with different samples and parameter settings.
In terms of speed, DRAGEN is in the lead by FPGA acceleration, and can usually complete whole genome calling in minutes to hours; Strelka2 is also optimized to analyze paired samples relatively quickly, while GATK has a large amount of computation in the whole process, which requires more time and computational resources; FreeBayes is moderately fast, but has better support for multiple samples, which is suitable for parallel analysis. Researchers can choose the right tool according to their needs in different situations: DRAGEN for extreme speed, GATK for large-scale population variation studies, Strelka2 for low-frequency somatic variation, and FreeBayes for flexible customization.
Each tool has its own advantages: DRAGEN is extremely fast and accurate, but relies on proprietary hardware; GATK is comprehensive and mature; Strelka2 is sensitive to somatic cell variation detection and suitable for tumor analysis; FreeBayes is open-source and flexible, and is suitable for multi-sample and polyploidy research. In the future, the development of new technologies such as deep learning and graph genome will make variant detection more accurate and efficient, and meet the needs of precision medicine and large-scale sequencing.
Accuracy Challenges in Variant Calling
High-throughput Sequencing (HTS) technology is widely used in genomic research and clinical diagnosis, and the accuracy of Variant Calling is increasingly required. However, Variant Calling still faces many challenges due to a variety of technical and biological factors. In this paper, we will briefly discuss the Sources of Error and Benchmarking Approaches.
Sources of Error
Sequencing depth is one of the key factors affecting detection sensitivity. At low depth, the number of reads supported by variants is reduced, especially low-frequency variants (low variant allele frequency) are more likely to be missed; while too high depth may amplify the sequencing noise. Therefore, a reasonable depth setting, combined with quality control measures such as deduplication and base quality score recalibration (BQSR), is the basis for ensuring accuracy.
Reference genome bias is also an important source of error. Current mainstream reference sequences (e.g., GRCh38) cannot cover all population variants, which makes it difficult to compare regions with large differences from the reference, thus affecting the identification of variants. The rise of personalized references and pangenome strategies is providing new paths to reduce such biases.
In addition, alignment artifacts are prevalent in regions rich in complex repeats or structural variants. Read fragments are prone to mismatch in these regions, generating false variant signals, and random errors introduced during PCR amplification can exacerbate the problem. Optimizing comparison algorithms and enhancing data filtering are essential to reduce false positives caused by artifacts.
Benchmarking Approaches
In order to objectively assess the performance of variant detection, Truth sets are often used for benchmarking. Truth Sets are derived from the results of multi-platform and multi-tool cross-validation of standardized samples (e.g., NA12878) and can be used to calculate true positives, false positives, and false negatives, thus assessing the sensitivity and precision of the detection method. However, the existing truth sets are mostly limited to easy-to-compare and high-coverage regions, and the coverage of complex variants is insufficient, which may underestimate some of the detection errors.
Trio Analysis, on the other hand, identifies variants that violate Mendelian Inconsistency by comparing the genotypic identity of parents and offspring. Theoretically, the vast majority of offspring variants should be observed in one parent, and the presence of offspring-specific variants that are not present in either parent is most likely a detection error (except for a very small number of de novo mutations). Triptych analysis not only estimates the false-positive rate but also reveals systematic underdetection, providing an important addition to the comprehensive assessment of variant detection methods.
Integration with Downstream Analyses
Variant detection is a key part of the bioinformatics process, but its value relies on a tight interface with downstream annotation and biological interpretation. Standardized interface design and standardized data formats enable rapid and efficient transfer of assay results to functional annotation, pathogenicity assessment, and mechanistic studies.
Framework for variation discovery and genotyping from next-generation DNA sequencing (DePristo et al, 2011)
Mechanisms for Bridging Variation Detection and Annotation
Variation detection tools (e.g., GATK, DeepVariant) usually output standard VCF files, which are directly used as inputs to annotation tools (e.g., VEP, SnpEff.) The loci, variant types, and quality information recorded in the VCFs provide the basis for the subsequent annotation and functional prediction.
Basic annotation identifies gene locations, variant effects, and coding influences by comparing reference databases. Further, the integration of clinical databases such as ClinVar and OMIM allows the determination of the pathogenic potential of the variants. For SV, tools such as nanotatoR combine genomic rearrangement information for annotation.
Modern processes such as Agilent Alissa Reporter have integrated variant detection and annotation, dramatically improving efficiency and standardization.
Standardization in multi-tool processes
In a collaborative multi-tool process, standardized inputs and outputs are central to ensuring reproducibility and consistency of results.
Unified comparison file formats (e.g., BAM) and variation file formats (e.g., VCF) enable seamless integration across detection and annotation modules, and the unified way of recording information (e.g., INFO field) in the VCF standard also facilitates comparison and integration of results across tools.
The modular design further enhances process flexibility, e.g., the standardized interval file (BED) supports parallel computation, speeding up large-scale analysis. Unified quality filtering criteria (e.g., QUAL value setting) and reference datasets (e.g., GIAB) also facilitate the alignment of process evaluation with international standards.
Through standardization, variant detection results flow accurately and efficiently into downstream analysis, supporting functional interpretation and clinical applications.
References
- Bathke, Jochen, and Gesine Lühken. “OVarFlow: a resource optimized GATK 4 based Open source Variant calling workFlow.” BMC bioinformatics vol. 22,1 402. 13 Aug. 2021, doi:10.1186/s12859-021-04317-y
- DePristo, Mark A et al. “A framework for variation discovery and genotyping using next-generation DNA sequencing data.”Nature genetics vol. 43,5 (2011): 491-8. doi:10.1038/ng.806
- Poplin R, Ruano-Rubio V et al. “Scaling accurate genetic variant discovery to tens of thousands of samples.” bioRxiv, 201178 (2017). doi:10.1101/201178
- Alldredge, Jill, and Leslie Randall. “Germline and Somatic Tumor Testing in Gynecologic Cancer Care.” Obstetrics and gynecology clinics of North America vol. 46,1 (2019): 37-53. doi:10.1016/j.ogc.2018.09.003
