Hidden in the Noise: Low-Variant Allele Frequency Mutations and Their Impact on Precision Oncology

Abstract

Intratumoral heterogeneity is a defining feature of cancer, yet standard sequencing and reporting practices often overlook somatic variants present at low variant allele frequencies (VAFs), commonly below 5%. Increasing evidence indicates that these rare alleles can represent clinically meaningful subclones involved in tumor evolution, therapeutic resistance, minimal residual disease, and metastatic dissemination. However, detecting and interpreting low-VAF variants is technically and analytically challenging because background error rates, library artifacts, genomic context, and caller assumptions increasingly overlap with true signal as allele fraction decreases. In this review, we integrate biological and clinical evidence supporting the relevance of low-VAFs and evaluate constraints across sequencing strategies, including whole genome and whole exome approaches and deep targeted panels. We discuss why detectability depends strongly on variant class and genome architecture, with SNVs generally more tractable than indels and structural variants. We then summarize practical approaches that improve sensitivity and specificity beyond increasing depth, including proper tissue handling, molecular enrichment, unique molecular identifiers, duplex-consensus methods, advanced error modeling, and orthogonal validation. Finally, we highlight emerging single-cell, spatial, and multiomic technologies that resolve rare variants in a cellular context. Collectively, these advances support incorporating low-VAF detection into precision oncology frameworks.

1. Introduction

Cancer genomes are mosaics of evolving clonal and subclonal populations shaped by mutational processes, chromosomal instability, therapy selection, and microenvironmental pressures [1,2]. Although clinical sequencing workflows have traditionally emphasized variants present at higher variant allele frequencies (VAFs) due to artifacts in clinically available formalin-fixed, paraffin-embedded (FFPE) tissue, a growing body of evidence indicates that biologically and therapeutically consequential alterations often exist below conventional reporting thresholds, frequently at VAFs < 5% [3,4,5]. VAF represents the proportion of sequencing reads supporting a given variant, and low-VAF mutations (≤5%) typically reflect subclonal tumor populations, while very low-VAF variants (≤1%) approach the background error rate of many sequencing platforms and require specialized detection methods. These low-frequency variants can mark emergent resistant subclones, minimal residual disease, and metastatic precursors, yet they are often filtered as technical noise due to declining signal-to-noise ratios at low allele fractions. The challenge is further influenced by tissue preservation, variant class, and genomic context. Single-nucleotide variants (SNVs) are generally detectable at far lower VAFs than insertion and deletions (indels) and structural variants (SVs), while repetitive and low-complexity regions remain difficult even at substantial depth [6]. This review synthesizes current evidence for the biological and clinical significance of low-VAFs, outlines the technical and analytical constraints that limit their detection, and evaluates emerging strategies to improve their interpretation in precision oncology.

Low-frequency variants arise naturally from the evolutionary dynamics of cancer, driven by ongoing mutation, selective pressures from the tumor microenvironment, and therapeutic intervention. As tumors expand and diversify, new subclonal populations emerge while others are suppressed or eliminated, resulting in a constantly shifting distribution of variant allele frequencies. At any given time point, subclones may be present at low or ultralow-VAF either because they represent newly emerging populations or residual clones following treatment [7]. These dynamics create a biological expectation that clinically relevant mutations will often exist at low abundance. However, detecting these variants remains challenging due to sequencing noise, preservation-induced artifacts, and sampling limitations, which can obscure true subclonal signals or generate false positives (see Section 3). Accordingly, understanding how intratumoral heterogeneity gives rise to low-VAF variants provides the foundation for interpreting their clinical significance, as discussed in the following section.

2. Biological and Clinical Context

Different cancer types exhibit substantial intratumoral genetic heterogeneity, manifested as variable gene expression across individual cells and the emergence of genetically distinct subclonal populations [2]. A large-scale analysis of over 5000 tumors from The Cancer Genome Atlas (TCGA) used PhyloWGS to infer clonal structure and demonstrated wide variation in clonal diversity across 32 cancer types, with bladder cancer, lung adenocarcinoma, and ovarian cancer exhibiting the highest mean number of clones per tumor [1]. Importantly, clonal diversity correlated positively with mutation burden and copy number alterations, indicating that both mutational and chromosomal instability drive intratumoral heterogeneity. Because subclones often comprise only a fraction of the tumor mass, many biologically relevant mutations are expected to occur at low-VAFs.

Subclonal heterogeneity is evident across neoplasms and is particularly well characterized in acute lymphoblastic leukemia (ALL), where nearly half of pediatric diagnostic samples harbor exclusively subclonal alterations [8]. Similar patterns have been observed in solid tumors, including hepatocellular carcinoma, where subclone-specific gene expression changes were obscured by bulk RNA sequencing but resolved when subpopulations were examined individually [9]. These findings underscore that bulk analyses systematically underestimate tumor complexity and preferentially miss low-VAF events.

Intratumoral heterogeneity also has clear prognostic implications. Quantitative measures such as Mutant-allele Tumor Heterogeneity (MATH) have been associated with significantly worse overall survival in multiple cancer types, including bladder and pancreatic cancers [10,11]. Moreover, heterogeneity is dynamic: selective pressures from the tumor microenvironment and therapy drive temporal shifts in subclonal composition, allowing low-frequency variants to expand and mediate recurrence or therapeutic resistance [12]. This microcosm of natural selection provides a rationale for how low VAFs can emerge to cause recurrence and resistance in the face of traditional and targeted therapies (Figure 1). Thus, it is expedient to support technologies and tissue management that allow for the most accurate detection of low-VAFs.

Assay choice critically determines whether low-VAF variants are detected. Comparative analyses of whole-genome sequencing (WGS), whole-exome sequencing (WES), transcriptome sequencing, and targeted panels show that each approach yields distinct therapy recommendations [14]. While WGS provides the most comprehensive genomic context, its distributed sequencing depth limits sensitivity for low-VAFs, with reliable detection often falling off below 5% VAF even at high coverage [15]. In contrast, targeted panel sequencing concentrates depth on selected loci, enabling more accurate quantification of low-VAF mutations. Clinical studies consistently show superior detection of subclonal and resistance-associated variants using panels, including EGFR T790M mutations that are frequently present < 5% VAF in lung cancer [4,5,16]. Additional panel-based workflows showed 100% sensitivity for variants above 3% VAF [17], further illustrating the diagnostic strength of focused sequencing methods.

WES represents an intermediate strategy, offering greater coverage of coding regions than WGS and improved sensitivity for low-VAF variants within those regions, but at the expense of detecting structural variants and non-coding alterations. As a result, WES is commonly used in precision oncology research when coding mutations are the primary focus, whereas WGS is reserved for applications requiring broader genomic characterization. However, increased sensitivity comes at the cost of reduced discovery potential. Targeted panels are inherently limited to predefined regions and may miss novel structural variants or unexpected genomic events that are detectable by WGS or transcriptome sequencing [18,19,20]. Collectively, these findings highlight a fundamental tradeoff between genomic breadth and low-VAF sensitivity. Because clinically meaningful subclones often exist at low allele fractions, effective precision oncology requires sequencing strategies that deliberately balance coverage depth and genomic scope to ensure that biologically and therapeutically relevant variants are not overlooked.

3. Technical and Analytical Considerations

3.1. Impact of Tissue Storage Format and FFPE-Induced Artifacts

FFPE preservation, while standard in pathology, chemically alters and degrades genetic material. While protocols can vary with each facility, FFPE typically requires a multi-day process where the tissue is placed in a formalin solution, dehydrated with varying alcohol solutions, and then embedded in paraffin wax. This process preserves tissue architecture, but studies have shown that it can alter genetic material. The formalin fixation process causes DNA crosslinking, fragmentation, and base damage such as cytosine deamination, which lowers DNA quantity and quality and introduces sequencing artifacts when compared to fresh, flash-frozen, or cryopreserved tissues [21,22]. Consequently, FFPE samples consistently exhibit poorer sequencing quality metrics, increased background noise, and greater discordance in variant calls compared with fresh or frozen tissue [23,24]. These effects translate into practical limitations, with next-generation sequencing (NGS) failing in 20–40% of FFPE samples and up to 60% of FFPE tumor specimens rejected in clinical trial settings due to inadequate quality [25].

These limitations disproportionately impact the detection of low-VAFs. DNA damage and reduced library complexity necessitate stringent artifact filtering, which preferentially suppresses low-VAF signals while allowing high-VAF mutations to remain detectable. In contrast, flash-frozen and cryopreserved tissues preserve DNA and RNA in a near-native state, yielding higher molecular weight nucleic acids and improved sequencing performance [23,26].

Empirical evidence underscores the magnitude of this effect. In our recent analysis of 50 matched samples, WGS of cryopreserved tissue consistently detected more structural variants and oncogenic driver mutations than matched FFPE specimens [21]. FFPE samples showed inflated tumor mutational burden (13.7 vs. 6.4 mutations/Mb in frozen tissue), suggesting false-positive artifact calls, and only 43.5% overlap in variants with VAF > 5% between paired samples. Notably, discordance increased further at lower allele fractions, and even clinically actionable mutations were inconsistently detected in FFPE. Similar findings were reported by the 100,000 Genomes Project, where 16% of FFPE-derived samples failed sequencing outright, and concordance with frozen tissue was modest, demonstrating 71% for SNVs and only 44% for copy-number alterations, while high-VAF hotspot mutations showed higher concordance [26].

Alternative preservation strategies highlight the tradeoffs between clinical feasibility and molecular integrity. Fresh tissue provides the highest-quality DNA but is rarely practical in routine workflows due to rapid degradation. Cryopreservation maintains DNA and RNA integrity with minimal chemical modification and preserves cell viability, enabling downstream functional assays, though it requires specialized storage. Flash-freezing provides similar molecular quality but compromises cell viability. Despite logistical challenges, both approaches outperform FFPE in preserving sequence fidelity and enabling reliable detection of low-frequency variants.

Collectively, these findings demonstrate that FFPE preservation introduces both false negatives and false positives and is particularly ill-suited for resolving low-VAF subclonal mutations. Although FFPE-specific error filters can partially rescue high-confidence variants, they further risk discarding true low-frequency events that resemble artifacts. When accurate detection of rare variants is a priority, cryopreserved tissue provides a substantially more reliable substrate for genomic profiling.

3.2. Bioinformatic Pipelines as a Primary Source of Error

Independent of sequencing chemistry, bioinformatic interpretation constitutes a major source of error in low-VAF variant detection. A multi-laboratory assessment using synthetic plasmids encoding challenging pathogenic variants demonstrated that most laboratories generated sufficient sequencing data, yet failed to detect clinically relevant mutations due to limitations in variant-calling pipelines rather than read evidence [27]. Manual review revealed that supporting reads were frequently present but were excluded by overly stringent filters or caller assumptions. Approximately 13% of pathogenic variants across a cohort of more than 470,000 patients met criteria for being “challenging,” reflecting sequence context, repetitive content, or structural complexity.

Benchmarking studies further show that modality-aware variant callers substantially outperform bulk callers repurposed for specialized data types. When bulk variant callers are applied to single-cell or ultra-deep sequencing data without tailored filtering, false-positive rates increase sharply, whereas callers designed for specific sequencing modalities achieve superior performance at allele fractions below 1% [28]. These findings underscore that analytical pipelines must be co-designed with sequencing strategy and expected VAF range.

3.3. Mutational Signatures, Reference Dependence, and Hidden Bias

Analytical bias also arises during the interpretation of mutational signatures. Signature detection depends on prior knowledge of mutational processes, meaning that novel or low-prevalence signatures may remain undetected if absent from reference catalogs [29]. Half of all cataloged variants on ClinVar are classified as variants of unknown significance. Accuracy is further influenced by tumor mutational burden, signature similarity, and exposure strength, with low-level signatures often masked by dominant processes. Systematic differences between WGS and WES have been observed, reinforcing that sequencing strategy can bias downstream biological interpretation, particularly in heterogeneous or low-mutation tumors [29].

3.4. Lack of Standardization and Limits of Detection in Clinical NGS

The absence of standardized practices for NGS assay design, variant calling, and reporting further complicates low-VAF detection. Although targeted panels are increasingly used to identify actionable mutations, clinical workflows often focus on a limited set of predefined genes, and there is no universally accepted threshold for reporting low-VAF variants [30]. As a result, clinically relevant subclonal mutations may be detected but not reported or inconsistently interpreted across laboratories.

Interlaboratory variability further highlights this challenge. While high-confidence variant calls are generally reproducible across centers using standardized pipelines, concordance declines markedly as VAF decreases [31]. This variability is driven in part by the stochastic nature of sequencing coverage, where low read depth and uneven sampling reduce confidence in low-frequency variant calls. A methodological review of clinical NGS workflows demonstrated that even at 100× coverage with a nominal requirement of 10 supporting reads, false-negative rates approached 45%, corresponding to an effective limit of detection near 10% VAF [32]. Increasing depth to 500× improved performance but still resulted in inconsistent detection across laboratories, with reported limits of detection ranging from 5 to 15% VAF. Only at depths exceeding 1650× did low-frequency variants near 3% VAF become reliably detectable, highlighting the escalating cost and diminishing returns of depth-based approaches alone.

In parallel, variability in reporting guidelines further contributes to inconsistency in clinical interpretation. Multiple sets of oncological variant reporting standards have been proposed by different professional organizations, leading to differences in how low-VAF variants are classified and communicated to clinicians [33]. Without harmonized thresholds for detection, validation, and reporting, low-VAF variant interpretation remains highly dependent on institutional practices.

Collectively, these factors demonstrate that low-VAF detection is not only a technical challenge, but also a problem of standardization and reproducibility. Addressing these issues will require coordinated efforts to define consistent limits of detection, establish reporting guidelines, and align bioinformatic pipelines across clinical laboratories.

4. Variant-Class–Specific Challenges in Low-VAF Detection

Building on the technical constraints described in Section 3, the detectability of low-frequency variants is strongly influenced by variant class, as different types of genomic alterations generate distinct signal characteristics and error profiles. Across sequencing platforms and assay designs, a consistent hierarchy in low-VAF detectability emerges: SNVs are most readily detected, followed by indels, while structural variants SVs remain the most challenging.

4.1. SNVs: Highest Sensitivity at Low-VAF

SNVs are consistently the most tractable class at low-VAFs, including below 1% VAF. This reflects the relative simplicity of SNV signals, the maturity of short-read sequencing technologies, and extensive development of error-aware computational frameworks. Methods such as RareVar demonstrate high precision and recall near 1% VAF under deep sequencing conditions, while machine-learning-based approaches like DETexT further extend SNV detectability even at reduced sequencing depths [34,35]. These studies highlight that SNV detection at low-VAF is both technically feasible and increasingly robust, even under conditions of low tumor purity or high intratumoral heterogeneity (Table 1).

4.2. Indels: Intermediate Detectability with Size-Dependent Limitations

Indels exhibit intermediate detectability, with performance strongly influenced by event size and sequencing modality. Short-read callers generally perform well for small indels but show declining sensitivity as insertion or deletion length increases, whereas long-read approaches improve recall across a broader size range [6]. Targeted error-minimization strategies, such as svCapture, have enabled reliable detection of indels and SV junctions near 1% VAF, though false-positive rates increase below this threshold [36]. Importantly, detection of larger indels is further constrained by DNA fragmentation, rendering fresh or frozen tissue a prerequisite for accurate analysis (Table 1).

4.3. Structural Variants: Lowest Sensitivity at Low-VAF

Structural variants represent the most analytically challenging class at low allele fractions. Benchmarking studies using synthetic mosaic samples demonstrate that SV detection sensitivity drops sharply below 5% VAF across sequencing platforms, with recall remaining limited below 1% even at cumulative coverage exceeding 2000× [37]. Long-read technologies outperform short-read approaches by spanning breakpoints and repetitive regions, but gains diminish as VAF decreases and require higher DNA input and cost [37,38]. Short-read WGS, in contrast, frequently fails to anchor reads spanning insertions or complex rearrangements, leading to systematic under-detection of low-frequency SVs [39,40]. While single-cell and breakpoint-targeted approaches can partially overcome these limitations, they introduce substantial experimental complexity and limited throughput [41] (Table 1).

Table 1. Common variants and their descriptions.

Variant Class	Typical Size Range	Relative Detectability	Key Technical Strengths	Major Detection Limitations	Sequencing/Methods with Best Performance	Citations
SNVs	1 bp	Highest (detectable <1%, sometimes ≤0.1%)	Mature error models; extensive algorithmic optimization; effective use of UMIs and ML-based callers	Sensitivity drops sharply at reduced depth; false positives without error modeling	Short-read WGS/WES with error-aware callers (RareVar, DeepVariant); ML-based methods (DETexT); deep targeted panels	[34,35,42]
Indels	1–50 bp (larger indels > 50 bp overlap with SVs)	Intermediate (reliable 1–5%, size-dependent)	Improved performance with long-read sequencing; targeted error suppression reduces artifacts	Poor detection of larger indels in short reads; algorithm performance highly size-dependent	Long-read WGS (PacBio HiFi, ONT); error- minimized targeted capture (svCapture)	[6,36,42]
SVs	≥50 bp (insertions, deletions, inversions, duplications, translocations)	Lowest (sharp sensitivity loss < 5%; limited < 1%)	Long-read sequencing resolves breakpoints; single-cell and breakpoint-based methods improve resolution	High false-positive rates; poor performance in repetitive regions; lack of gold standards; high cost and DNA input	Long-read WGS (PacBio HiFi > ONT > short-read); single-cell breakpoint-based approaches	[37,38,41,43]
All Variant Classes (WES context)	—	Unreliable below 5% VAF	Broad coverage; cost- efficient for high-VAF variants	High false-positive rate at low-VAF; poor concordance across sample types	Not recommended without UMIs or ultra-deep targeting	[44]

Table 1. Common variants and their descriptions.

Variant Class	Typical Size Range	Relative Detectability	Key Technical Strengths	Major Detection Limitations	Sequencing/Methods with Best Performance	Citations
SNVs	1 bp	Highest (detectable <1%, sometimes ≤0.1%)	Mature error models; extensive algorithmic optimization; effective use of UMIs and ML-based callers	Sensitivity drops sharply at reduced depth; false positives without error modeling	Short-read WGS/WES with error-aware callers (RareVar, DeepVariant); ML-based methods (DETexT); deep targeted panels	[34,35,42]
Indels	1–50 bp (larger indels > 50 bp overlap with SVs)	Intermediate (reliable 1–5%, size-dependent)	Improved performance with long-read sequencing; targeted error suppression reduces artifacts	Poor detection of larger indels in short reads; algorithm performance highly size-dependent	Long-read WGS (PacBio HiFi, ONT); error- minimized targeted capture (svCapture)	[6,36,42]
SVs	≥50 bp (insertions, deletions, inversions, duplications, translocations)	Lowest (sharp sensitivity loss < 5%; limited < 1%)	Long-read sequencing resolves breakpoints; single-cell and breakpoint-based methods improve resolution	High false-positive rates; poor performance in repetitive regions; lack of gold standards; high cost and DNA input	Long-read WGS (PacBio HiFi > ONT > short-read); single-cell breakpoint-based approaches	[37,38,41,43]
All Variant Classes (WES context)	—	Unreliable below 5% VAF	Broad coverage; cost- efficient for high-VAF variants	High false-positive rate at low-VAF; poor concordance across sample types	Not recommended without UMIs or ultra-deep targeting	[44]

4.4. Implications for Low-VAF Variant Interpretation

These variant-class–specific limitations are mirrored in clinical assays. Comparative evaluations of circulating tumor DNA (ctDNA) panel sequencing show consistently higher sensitivity for SNVs, intermediate performance for indels, and markedly reduced sensitivity for SVs at low-VAF, even under optimized sequencing depth and deduplication [42]. RNA fusion sequencing (RNA-FS) panels similarly outperform classical cytogenetics for detecting low-frequency or cryptic fusions, as demonstrated in acute myeloid leukemia, where RNA-FS doubled the detection rate of clinically relevant fusion events compared with karyotyping and fluorescence in situ hybridization (FISH) [45].

Finally, the limitations of WES further emphasize the importance of variant-class–aware strategies. More than half of variants identified by WES below 5% VAF fail orthogonal confirmation, with concordance dropping below 1% for low-VAF calls in FFPE samples [44]. While high-VAF driver mutations remain detectable, subclonal variants, particularly indels and SVs, are disproportionately affected, reinforcing that standard WES is poorly suited for confident low-VAF detection without molecular error correction or ultra-deep targeting.

Collectively, these findings establish that variant class is a primary determinant of low-VAF detectability. While advances in error suppression and sequencing technology have made low-frequency SNVs increasingly accessible, indels and especially SVs remain constrained by biological complexity, sequencing chemistry, and analytical limitations. For precision oncology, this hierarchy has direct clinical implications. Actionable subclonal events may be reliably detected or entirely missed depending on both the variant type and the chosen assay. Therefore, effective strategies require deliberate alignment of sequencing modality, tissue quality, and analytical framework with the variant classes most relevant to patient management.

5. Approaches to Improve Low-VAF Detection

Accurate identification of low-frequency variants plays a pivotal role in applications such as early cancer detection, minimal residual disease monitoring, and longitudinal tracking of tumor evolution. Beyond merely increasing sequencing depth, several complementary strategies have emerged that significantly improve both sensitivity and specificity. Among these, molecular enrichment, molecule-level tagging, duplex consensus sequencing, advanced computational pipelines, and orthogonal assay validation have each demonstrated strong performance in detecting variants below 1% VAF. Collectively, these methods provide a robust framework for overcoming the limitations of standard NGS in clinical and research contexts.

Selective enrichment of rare alleles before sequencing can substantially improve mutant detection, particularly through approaches such as molecular barcoding with unique molecular identifiers (UMIs), which suppress sequencing errors and enable detection of low-frequency variants [46]. Blocker displacement amplification (BDA) strategies use oligonucleotide blockers to suppress wild-type amplification and increase the relative abundance of mutant templates. Studies implementing long blocker displacement amplification (LBDA) have achieved reliable detection of mutations down to 0.5% VAF and revealed clinically relevant differences in colorectal cancer patient samples [47]. Expanding this approach, multiplexed BDA (mBDA) assays targeting up to 80 regions demonstrated quantification of rare variants at frequencies as low as 0.019% with only 250× sequencing depth [48]. These results indicate that targeted pre-sequencing enrichment can offer a cost-effective alternative to deep sequencing without compromising detection accuracy.

UMIs and structured barcode systems represent another critical innovation for low-VAF variant detection. Tagging each input molecule with a unique sequence before amplification enables reconstruction of original molecules and elimination of polymerase and sequencing artifacts (Figure 2). UMI-based workflows applied to cfDNA have detected variants down to 0.09% VAF with high confidence [49]. Moreover, structured UMI designs that are engineered to prevent index misassignment allow detection of variants at or below 0.01% VAF [50]. Similarly, barcode-enabled consensus algorithms significantly reduce false positives and show strong concordance with orthogonal assays [51]. These improvements underscore the advantage of molecular tagging for distinguishing authentic low-frequency events from technical noise.

Strand-aware duplex sequencing provides another layer of precision by independently tagging both DNA strands and only calling variants observed on both. This method eliminates most single-strand damage artifacts, achieving theoretical background error rates near 10⁻⁹ per base [52]. In clinical settings, duplex sequencing has revealed clinically actionable variants below 0.01% VAF in pediatric leukemia cases that conventional NGS failed to detect [53]. Complementing these wet-lab innovations, advanced computational pipelines now employ machine learning, contextual error modeling, and depth-aware binomial filters to improve variant classification. For instance, a hybrid pipeline using XGBoost classification identified intrahost human papillomavirus (HPV) variants down to 0.3% VAF with strong precision [54], while benchmarking of variant callers confirms that UMI-aware methods outperform raw-read-based approaches below 1% VAF [55,56].

Finally, droplet digital PCR (ddPCR) remains a powerful orthogonal validation tool, offering near-digital quantification and detection limits as low as 0.005% VAF [57]. When integrated with enrichment, molecular barcoding, duplex-consensus sequencing, and computational filtering, ddPCR provides an independent benchmark for confirming rare variants. Collectively, these strategies represent a shift away from depth-centric sequencing toward holistic optimization of library preparation, error suppression, and analytical interpretation, defining the current standard for accurate low-frequency variant detection in precision genomics.

6. Biological Validation and Clinical Relevance

Low-VAF detection has direct clinical utility across several key applications, including targeted therapy selection, minimal residual disease (MRD) monitoring, early detection of recurrence, identification of resistance mutations, and longitudinal disease tracking through circulating tumor DNA (ctDNA). In these contexts, low-frequency variants provide clinically actionable information by informing therapeutic decisions, predicting relapse prior to radiographic progression, and enabling dynamic assessment of disease burden over time.

The choice of detection modality is closely aligned with the clinical objective. For targeted therapy selection and resistance mutation identification, deep targeted sequencing panels—often incorporating molecular barcoding or duplex-consensus error suppression—provide high sensitivity and specificity across predefined actionable loci. In contrast, applications requiring ultra-high sensitivity, such as MRD detection and early recurrence monitoring, typically rely on ddPCR or ultra-deep, error-corrected sequencing approaches capable of detecting variants at or below ~0.1% VAF. For ctDNA-based disease tracking, targeted sequencing of plasma-derived cfDNA, combined with molecular error suppression and longitudinal sampling, enables non-invasive monitoring of tumor dynamics. Across these use cases, orthogonal validation—commonly via ddPCR or independent targeted assays—remains critical when low-VAF findings directly inform clinical decision-making. While targeted panels and ddPCR are established in clinical workflows, many ultra-deep and error-corrected sequencing approaches remain in transition from research to clinical validation.

These technical considerations have clear clinical consequences. In a targeted sequencing study of more than 5000 tumor samples, a substantial fraction of clinically actionable mutations that included EGFR, KRAS, PIK3CA, and BRAF were present below 5% VAF, a range where standard short-read NGS workflows frequently fail [4]. Notably, patients harboring such low-frequency mutations nonetheless derived clinical benefit from targeted therapy, including a metastatic lung cancer patient with an EGFR T790M mutation at 3–4% VAF who achieved partial remission (Table 2). These cases illustrate that low-VAF variants are not merely technical noise but can be biologically and therapeutically decisive.

Improved detection sensitivity has also reshaped estimates of mutation prevalence and treatment eligibility. Ultra-sensitive detection of ESR1 Y537S and D538G mutations down to 0.003% VAF increased their apparent prevalence in primary breast cancer from ~1% to over 12%, directly expanding the population eligible for selective estrogen receptor degrader (SERD) therapies [58]. Similarly, duplex sequencing in acute myeloid leukemia uncovered extensive subclonal heterogeneity with variants below 1% VAF. These rare allele variants were present in up to 53% of the blast population once the cancer relapsed, suggesting subclonal expansion [59]. These findings highlight the extent of clinically meaningful variation that is systematically missed by conventional sequencing approaches.

Low-VAF detection is particularly impactful in ctDNA analysis. Conventional ctDNA assays often fail to detect variants below 5% VAF due to background noise and dilution by cell-free DNA, especially in early disease or minimal residual disease (MRD) settings [60]. Advanced methods, including joint-genotype modeling, ultra-deep sequencing, and integrative error-suppression strategies, have enabled reliable detection down to 0.1% VAF or lower, with some workflows achieving sensitivity below 0.01% while maintaining high specificity [61,62,63,64]. Across studies, ctDNA levels and measured VAF correlate with tumor burden and clinical outcomes, supporting ctDNA as a quantitative biomarker of disease status when shedding is sufficient [65,66,67,68].

However, ctDNA detection remains constrained by biological factors. In early disease and MRD, tumor-derived DNA may constitute less than 1% of total cell-free DNA (cfDNA), increasing susceptibility to noise and false negatives [69,70]. While high-depth sequencing can partially mitigate these limitations, low shedding tumors and temporal clonal drift can still obscure detection [60,62,71,72,73]. Accordingly, a negative ctDNA result should not be interpreted as absence of disease, and sampling timing relative to tumor evolution is critical [65,74]. Notably, ctDNA VAF has been shown to correlate with survival, further supporting its value in clinical monitoring [75].

Beyond cross-sectional concordance, ultra-low-VAF detection has proven especially valuable for monitoring MRD and therapeutic resistance. Persistence or reemergence of low-frequency ctDNA variants following treatment consistently predicts relapse months before radiographic or clinical progression [76,77,78]. For example, in acute myeloid leukemia, patients with detectable low-VAF ctDNA following remission had significantly higher relapse rates compared to those without detectable variants [79]. Similarly, in colorectal cancer, low-VAF ctDNA detection identified recurrence a median of 112 days earlier than imaging-based diagnosis [80] (Table 2).

Moreover, low-VAF mutations detected in primary tumors or early ctDNA samples often represent pre-existing resistant subclones, such as EGFR, ESR1, or KRAS mutations, that later expand under treatment pressure [5,13,81]. Early identification of these variants enables anticipation of therapeutic failure and supports adaptive treatment strategies [82] (Table 2).

Table 2. Clinically relevant variants and their descriptions.

Gene	Cancer Context	Typical VAF Context	Clinical Implication	Key Finding
TP53 (subclonal)	Chronic lymphocytic leukemia (CLL)	~1–5%	Prognostic/relapse	Subclonal TP53 mutations (~2% VAF) are frequently missed by standard methods but confer similar relapse risk as clonal TP53 alterations [83,84].
TP53 (low-VAF)	Follicular lymphoma (FL)	<10%	Prognostic	Low-VAF TP53 mutations are associated with increased treatment resistance and disease progression [85].
TP53 and KRAS (subclonal)	T-Cell acute lymphoblastic leukemia	<10%	Prognostic/treatment stratification	Subclonal TP53 and KRAS mutations may identify high-risk patients and inform treatment intensification strategies [86].
TP53 (subclonal)	Multiple (CLL, FL)	<10%	Resistance	Subclonal TP53 alterations are consistently associated with treatment resistance across cancer types [87,88].
KRAS (subclonal)	Metastatic Colorectal Cancer	<5%	Resistance	Low-frequency KRAS mutations drive resistance to anti-EGFR therapy and expand under treatment pressure [13,80,89,90,91].
FLT3-ITD (allelic frequency/ratio)	Acute Myeloid Leukemia (AML)	Variable	Prognostic	Higher allelic burden correlates with worse prognosis and increased relapse risk [92,93].
DNMT3A (subclonal)	Acute Myeloid Leukemia (AML)	<10%	Prognostic/relapse	DNMT3A mutations persist in pre-leukemic clones and are associated with increased relapse risk and reduced overall survival [94].
IDH1/IDH2 (subclonal)	Acute Myeloid Leukemia (AML)	<10%	Prognostic	Subclonal IDH mutations contribute to clonal evolution and may influence relapse dynamics [59,95].
NPM1 (low-VAF)	NPM1-mutated AML	MRD-level	MRD marker	Persistent low-VAF NPM1 mutations are widely used as markers of minimal residual disease and relapse prediction [79,96].
BCR-ABL1 (low-VAF)	Chronic Myeloid Leukemia	ultra-low	MRD/resistance	Low-level BCR-ABL1 detection enables early identification of relapse or treatment failure [97,98].
EGFR T790M (low-VAF)	Non-Small Cell Lung Cancer	<5%	Resistance/predictive	Low-VAF T790M mutations predict resistance to EGFR inhibitors and influence response to osimertinib; lower VAF is associated with poorer outcomes [99,100,101].

Table 2. Clinically relevant variants and their descriptions.

Gene	Cancer Context	Typical VAF Context	Clinical Implication	Key Finding
TP53 (subclonal)	Chronic lymphocytic leukemia (CLL)	~1–5%	Prognostic/relapse	Subclonal TP53 mutations (~2% VAF) are frequently missed by standard methods but confer similar relapse risk as clonal TP53 alterations [83,84].
TP53 (low-VAF)	Follicular lymphoma (FL)	<10%	Prognostic	Low-VAF TP53 mutations are associated with increased treatment resistance and disease progression [85].
TP53 and KRAS (subclonal)	T-Cell acute lymphoblastic leukemia	<10%	Prognostic/treatment stratification	Subclonal TP53 and KRAS mutations may identify high-risk patients and inform treatment intensification strategies [86].
TP53 (subclonal)	Multiple (CLL, FL)	<10%	Resistance	Subclonal TP53 alterations are consistently associated with treatment resistance across cancer types [87,88].
KRAS (subclonal)	Metastatic Colorectal Cancer	<5%	Resistance	Low-frequency KRAS mutations drive resistance to anti-EGFR therapy and expand under treatment pressure [13,80,89,90,91].
FLT3-ITD (allelic frequency/ratio)	Acute Myeloid Leukemia (AML)	Variable	Prognostic	Higher allelic burden correlates with worse prognosis and increased relapse risk [92,93].
DNMT3A (subclonal)	Acute Myeloid Leukemia (AML)	<10%	Prognostic/relapse	DNMT3A mutations persist in pre-leukemic clones and are associated with increased relapse risk and reduced overall survival [94].
IDH1/IDH2 (subclonal)	Acute Myeloid Leukemia (AML)	<10%	Prognostic	Subclonal IDH mutations contribute to clonal evolution and may influence relapse dynamics [59,95].
NPM1 (low-VAF)	NPM1-mutated AML	MRD-level	MRD marker	Persistent low-VAF NPM1 mutations are widely used as markers of minimal residual disease and relapse prediction [79,96].
BCR-ABL1 (low-VAF)	Chronic Myeloid Leukemia	ultra-low	MRD/resistance	Low-level BCR-ABL1 detection enables early identification of relapse or treatment failure [97,98].
EGFR T790M (low-VAF)	Non-Small Cell Lung Cancer	<5%	Resistance/predictive	Low-VAF T790M mutations predict resistance to EGFR inhibitors and influence response to osimertinib; lower VAF is associated with poorer outcomes [99,100,101].

Together, these studies demonstrate that low-VAFs detected in tissue and ctDNA are biologically meaningful indicators of tumor evolution, relapse risk, and resistance. Their reliable detection is therefore central to precision oncology, informing treatment selection, disease monitoring, and timely therapeutic intervention.

7. Orthogonal Validation of Low-VAF Calls

As VAFs approach the intrinsic error rates of NGS, independent confirmation becomes essential for distinguishing true low-VAFs from technical artifacts. Orthogonal validation strategies, therefore, play a critical role in establishing confidence in low-VAF calls, particularly when such variants inform clinical decision-making or longitudinal disease monitoring.

7.1. Long-Read Sequencing Technology

Long-read sequencing provides one approach to orthogonal validation by resolving genomic contexts that are difficult to interrogate with short reads. Platforms such as PacBio HiFi and Oxford Nanopore Technologies (ONT) differ in their tradeoffs between read length and per-read accuracy. PacBio HiFi sequencing achieves higher base-calling accuracy through circular consensus sequencing, whereas ONT provides substantially longer reads that improve detection of structural variants and complex genomic regions [102,103,104,105]. While long-read approaches are generally less sensitive for ultra-low-VAF detection than error-corrected short-read methods, they provide complementary validation by confirming variant structure and breakpoint resolution in regions prone to alignment ambiguity [106,107].

7.2. ddPCR

Highly sensitive targeted assays remain central to orthogonal confirmation of low-VAF variants. Droplet digital PCR (ddPCR) enables near-digital quantification of mutant alleles through reaction partitioning, achieving detection limits as low as 0.006–0.16% VAF in clinical settings [108]. However, ddPCR requires prior knowledge of the variant and is therefore best suited for confirming predefined, clinically actionable mutations rather than for discovery.

7.3. MIPP-Seq

To improve scalability, multiplexed validation approaches extend beyond single-variant assays. MIPP-Seq (Multiple Independent Primer PCR Sequencing) combines ultra-deep sequencing with multiple nonoverlapping amplicons per locus, reducing locus-specific artifacts through replication and enabling simultaneous confirmation of multiple low-VAF SNVs and indels with reported sensitivity near 0.025% VAF [109]. These approaches are particularly useful when validating larger sets of candidate variants.

7.4. Enriched Sanger Sequencing

Additional enrichment-based methods provide rapid and cost-effective confirmation for targeted variants. Blocker displacement amplification coupled with Sanger sequencing has demonstrated detection of mutations down to ~0.2% VAF with concordance to ddPCR and NGS results [110]. Similarly, strand displacement reaction-based approaches have reported detection near 0.1% VAF with high specificity in targeted applications [111]. While these methods are limited to known variants, they offer efficient confirmation for clinically relevant hotspots.

7.5. CODEC

At the sequencing level, duplex-based error suppression strategies offer an alternative form of internal validation. Concatenating Original Duplex for Error Correction (CODEC) enforces strand concordance, achieving substantially reduced error rates and enabling detection of rare variants using fewer reads than conventional duplex sequencing [112]. Although not an independent assay in the traditional sense, duplex-consistency constraints provide strong internal evidence for distinguishing true variants from technical noise.

7.6. Computational Tools

Computational frameworks further complement experimental validation by prioritizing likely true variants. The Cancer-associated Variant Enrichment (CAVE) method integrates positional and recurrence patterns to estimate the likelihood that low-frequency calls represent biological signal variants for confirmation, achieving strong concordance with ddPCR for variants below 1% VAF [113]. These approaches are most effective when used in combination with experimental validation rather than as stand-alone methods.

Collectively, these orthogonal strategies provide complementary approaches to validate low-VAF variants across clinical and research settings. While no single method is universally applicable, integrating targeted confirmation, duplex-level error suppression, and computational prioritization substantially improves confidence in low-frequency variant calls. As low-VAF detection becomes increasingly incorporated into clinical workflows, orthogonal validation will remain essential for ensuring analytical rigor and preventing misinterpretation of rare but clinically consequential mutations [114,115].

8. Emerging Technologies, Limitations, and Future Directions

Recent advances in sequencing chemistry, amplification fidelity, and analytical frameworks have made it increasingly feasible to resolve low-frequency somatic variants at cellular resolution. Single-cell DNA sequencing, multiomic profiling, and spatial mutation mapping now enable discrimination of true rare variants from technical noise while linking them to cell state, lineage, and microenvironmental context. This is an essential capability for interpreting low-VAF events that are obscured in bulk analyses. A summary of these emerging technologies is provided in

Table 3. Emerging Technologies in Low-VAF Detection.

Technology/Approach	Category	Principle	Strength	Limitations	Applications
Primary Template-Directed Amplification (PTA)	Amplification	Linear amplification from original templates	Reduces allele dropout and bias	Residual amplification artifacts	Single-cell sequencing, low DNA input
Single-cell DNA sequencing	Single-cell genomics	Sequencing individual cell genomes	Eliminates bulk signal dilution	Coverage variability, amplification bias	Clonal architecture, rare subclones
Multiomic single-cell sequencing (DEFND-seq)	Single-cell multiomics	Joint DNA and RNA profiling per cell	Links variants to transcriptional state	High cost, technical complexity	Functional subclone mapping
Targeted single-cell DNA-RNA co-sequencing (SDR-seq)	Targeted single-cell	Focused loci sequencing with RNA profiling	High sensitivity at selected loci	Limited genome-wide scope	Targeted mutation validation
Breakpoint-aware/amplicon-based sequencing	Structural variant detection	Targeted enrichment of rearrangements	Improves detection of complex SVs	Requires known breakpoints	Structural variant characterization
Spatial mutation profiling	Spatial genomics	Maps variants within tissue architecture	Reveals spatial heterogeneity	Lower sensitivity, technical complexity	Tumor microenvironment analysis
Modality-aware variant callers	Bioinformatics	Data-specific variant calling models	Improves signal-to-noise discrimination	Model-dependent performance	Single-cell variant calling
Statistical noise-modeling frameworks	Bioinformatics	Error modeling across datasets	Enhances detection at low allele fractions	Computationally intensive	Rare variant detection
Integrative multiomic analysis platforms	Computational integration	Links genomic and phenotypic data	Enables biological interpretation of variants	Requires multi-layer datasets	Clonal evolution, functional genomics

.

Multiomic single-cell approaches exemplify this progress by jointly capturing genomic and transcriptional information from individual cells. Methods such as DEFND-seq (DNA and Expression Following Nucleosome Depletion Sequencing) and related co-sequencing strategies have demonstrated the ability to detect rare SNVs and copy-number variants, resolve low-abundance subclones, and associate these alterations with cell-specific transcriptional programs [116]. Parallel improvements in amplification chemistry, including primary template-directed amplification (PTA), have reduced coverage bias and allele dropout, enabling more uniform genome representation and improved detection of low-frequency SNVs, indels, and SVs in single cells [117,118].

Targeted DNA–RNA co-profiling and breakpoint-aware strategies further enhance sensitivity by concentrating sequencing depth on loci of interest and interrogating variant structure directly. Approaches such as single-cell DNA-RNA sequencing (SDR-seq) enable accurate per-cell zygosity inference for variants present at low allele fractions [119]. Breakpoint-specific amplicon sequencing has demonstrated that read depth alone is insufficient to capture the full spectrum of somatic structural variation, particularly for complex rearrangements [41]. These methods highlight how variant class fundamentally shapes the strategies required for low-VAF detection.

Spatially resolved assays add a critical dimension by mapping rare variants back to histological and microenvironmental context. Spatial mutation profiling has revealed that low-frequency driver mutations can be confined to discrete tumor regions and cell populations, uncovering spatial heterogeneity that is invisible to bulk sequencing [120,121]. Such findings reinforce that low-VAF variants are often biologically localized rather than uniformly distributed across tumors.

Advances in bioinformatics have been equally important in translating these technologies into reliable discovery tools. Modality-aware variant callers and statistical noise-modeling frameworks aggregate weak signals across large single-cell datasets to distinguish true rare variants from artifacts, consistently outperforming bulk callers repurposed for single-cell data at low allele fractions [28,122]. Integrative multiomic platforms further link genomic alterations to phenotypic states, revealing how clonal fitness and selection operate even when subclones persist at low abundance [7,123].

Despite these advances, summarized in Table 3, several barriers remain to the routine implementation of low-VAF detection. Reliable identification of rare variants often requires ultra-deep sequencing coverage, increasing cost, computational burden, and data storage demands. In addition, sequencing and PCR error rates overlap with true low-frequency signals, necessitating robust error-suppression strategies such as molecular barcoding and duplex consensus methods. Detection is further complicated by genomic context, including repetitive regions and complex structural rearrangements, as well as variability in bioinformatic pipelines and a lack of standardized detection thresholds. Finally, approaches that achieve the highest sensitivity often sacrifice genomic breadth, whereas genome-wide methods frequently lack sufficient depth for confident low-VAF detection.

In summary, these emerging technologies demonstrate that robust interpretation of low-frequency variants requires more than increased sequencing depth. By integrating high-fidelity amplification, variant-class–aware detection, spatial context, and multiomic analysis, next-generation workflows place rare variants within their appropriate cellular and evolutionary framework. As these approaches mature and scale, they will play a central role in translating low-VAF detection into actionable biological and clinical insight.

9. Conclusions

Low-frequency somatic variants are a fundamental consequence of intratumoral heterogeneity and tumor evolution rather than peripheral technical artifacts. Across cancer types, variants presenting below conventional reporting thresholds encode critical information about subclonal architecture, therapeutic resistance, minimal residual disease, and relapse risk, and their clinical relevance has been repeatedly demonstrated in both tissue and circulating tumor DNA analyses. This review highlights that sensitivity at VAF < 5% is governed not by sequencing depth alone, but by the combined effects of tissue preservation, read length, genomic context, library preparation, error modeling, and variant-class–specific detectability. A consistent hierarchy emerges in which SNVs are most readily detected at low-VAF, followed by indels, while SVs remain the most challenging. Accordingly, assay selection must be aligned with the biological and clinical questions being addressed, balancing genomic breadth with the sensitivity required to detect clinically meaningful subclonal variation. While WGS and WES provide essential discovery and contextual information, targeted and error-corrected approaches are required for reliable detection of low-VAF variants with direct clinical impact. Moving forward, effective clinical integration of low-VAF information will depend on standardized performance benchmarks, modality-aware bioinformatic frameworks, and scalable orthogonal validation strategies. As single-cell, spatial, and multiomic technologies mature, low-frequency variants can be interpreted within their proper cellular, clonal, and microenvironmental context, enabling precision oncology to progress beyond dominant-clone paradigms toward truly subclonal-informed diagnosis and treatment.