Over the past decade, genomic evaluation has become central to cancer prevention, early detection of endogenous sources of mutation (e.g. physicochemical processes, free radicals, and enzymatic processes controlling DNA damage and repair), and pharmacogenomic-based therapeutic strategies. Substantial evidence has established the role of lifestyle and environmental determinants in carcinogenesis [1], with cumulative exposures shown to induce somatic mutations despite functional DNA repair mechanisms. Analysis of a cohort comprising 821 non cancer individuals revealed the presence of somatic mutations in histologically normal tissues, occurring at estimated frequencies of 2–6 mutations per million bases [2].

Advanced genomic technologies—including liquid biopsy [3], noninvasive prenatal testing [3], characterization of somatic mosaicism [4], delineation of tumor subclones [5], and lineage tracing [6]—have shown that somatic single-nucleotide variants (SNVs) frequently occur at low variant allele frequencies (VAFs) that evade detection by conventional sequencing approaches. A major challenge in identifying these low-VAF mutations is distinguishing true rare variants from germline polymorphisms, a difficulty compounded by the widespread use of tumor-only sequencing. Although assay and algorithmic improvements have advanced somatic mutation detection, most pipelines are optimized for variants at higher VAFs, which are enriched during clonal expansion. As a result, widely used oncology workflows remain insufficient for reliably detecting very low-frequency events in normal tissues [7], necessitating alternative methodological approaches.

These considerations underscore the importance of investigating cancer driver mutations (CDMs) as biomarkers for carcinogenicity assessment and for evaluating cancer risks arising from therapeutic, occupational, and environmental exposures [8]. Because CDMs can be quantified directly from DNA (genotypic selection), the analyses can be performed on isolated DNA from any tissue or species [9]. Our initial studies employed an allele-specific amplification method—Allele-specific Competitive Blocker polymerase chain reaction (ACB-PCR)—to quantify 1 mutant allele among 100,000 total alleles (10^-5) in DNA samples [9]. While ACB-PCR provided critical insight into CDMs as biomarkers of cancer risk, it was inherently low-throughput, allowing only single-mutation analyses. The advent of next-generation sequencing (NGS) enabled simultaneous interrogation of many CDMs, including rare variants, and spurred development of high-sensitivity approaches [10].

Although the types of mutations present in cancer genomes are well characterized, the timing of their occurrence has remained unclear until recently. Advances in NGS now enable detailed reconstruction of somatic evolution, throughout cellular lifespan, thereby documenting the transition from normal tissue homeostasis to malignant progression [11]. As part of the Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium of the International Cancer Genome Consortium (ICGC) and The Cancer Genome Atlas (TCGA) [12], whole-genome sequencing of 2,658 cancers identified 47 million point mutations. Additionally, temporal ordering of somatic mutations during tumor evolution was characterized. Among these, 22% were classified as early clonal, 7% as late clonal, 53% as unspecified clonal, and 17% as subclonal [11]. In a focused analysis of 453 known cancer driver genes, 5,913 oncogenic point mutations were detected, of which 29% were early clonal, 5% late clonal, 56% unspecified clonal, and 8% subclonal. These findings demonstrate that cancer driver mutations predominantly arise during early tumor development, contributing to cancer initiation, whereas later stages (late clonal to subclonal) are characterized by the accumulation of passenger or context-dependent mutations. This pattern underscores the temporal dynamics of driver acquisition in tumorigenesis [11].

The gradual accumulation of point mutations throughout a lifespan in both healthy tissues [13–17] and cancers [18] highlights the critical need to distinguish somatic from germline events to inform precision oncology [19].

To address this, our group developed an error-corrected NGS method, CarcSeq, designed to detect a curated panel of early recurrent somatic point mutations in oncogenes and tumor suppressor genes [8]. CarcSeq employs multiple high-fidelity PCR reactions to amplify target loci, with each amplicon tagged by a 9-base-pair unique molecular identifier. Sequenced libraries are processed bioinformatically to generate single-strand consensus sequences (Figure 1), allowing accurate quantification of both mutant and wild-type alleles.

While mutant fraction (MF) provides a sensitive measure of rare variant abundance, MF alone does not fully capture the heterogeneity or dispersion of mutation burdens across genomic loci or biological replicates—features that are critical for inferring early clonal expansion. To address this limitation, we incorporated the median absolute deviation (MAD) of MF as a metric. MAD provides a robust, distribution-agnostic measure of variability that is less influenced by extreme values than variance-based statistics [20], making it particularly well-suited for low-frequency mutation data generated by error-corrected NGS. As described below, MAD emerged as a sensitive indicator of clonal expansion and neoplastic potential across species and tissue types.

In initial studies, CarcSeq was applied to 13 gene regions encompassing more than 20 hotspot CDMs in DNA from normal human breast and lung tissues, as well as matched breast and lung tumors. Raw sequencing output (mutpos files) was processed using a three-step filtering pipeline as described in Harris et al. [8]. First, mutant fractions (MFs) supported by only one or two mutant single-strand consensus sequences (SSCSs) were excluded to eliminate imprecise estimates arising from rare-event sampling error [21]. For each mutation class (base substitutions to T, C, G, or A, as well as insertions and deletions), MF was calculated as the number of mutant SSCSs (≥3) divided by the effective sequencing depth, defined as the total number of SSCSs minus those lacking a confident error-corrected base call (Ns; positions where no single base was represented by ≥90% of SSCSs). Using this approach, CarcSeq quantified mutations with a sensitivity of 10^_-4 and revealed tissue-specific mutational profiles, including broader mutation distributions in normal breast tissue compared with ductal carcinomas, whereas tumors exhibited higher-frequency mutations concentrated at known hotspots [8].

Leveraging this sensitivity, we adapted CarcSeq for rodent models to assess clonal expansion and neoplastic potential by distinguishing nonsynonymous mutations (altering the encoded amino acid) from synonymous mutations (silent substitutions). In mammary tissue from 16-week-old untreated Fischer 344, Wistar Han, and Sprague Dawley rats—strains with varying susceptibility to spontaneous mammary neoplasia—CarcSeq detected hundreds of recurrent mutants (≥10^-4) per strain, with 42.5% of nonsynonymous variants mapping to human homologs. Sprague Dawley rats, the most tumor-prone strain, showed elevated nonsynonymous/synonymous mutation ratios in Hras, Pik3ca, and Tp53, consistent with positive selection and clonal expansion. The MAD in these loci correlated perfectly with spontaneous mammary tumor incidence at 104 weeks (r = 1.0, p = 0.001), supporting CarcSeq-derived mutation metrics as early predictors of neoplastic outcomes [22].

We next established a mouse CarcSeq panel to quantify low-frequency mutations across hotspot CDMs (Harris et al., 2021). In lung DNA from B6C3F1 and CD-1 mice, CarcSeq recovered 1,586 recurrent mutants (≥10^-4), with 55.5% of nonsynonymous variants overlapping with human tumor mutations, including hotspot codons. Male B6C3F1 mice, the most lung tumor–sensitive model, exhibited the highest nonsynonymous/synonymous ratio, consistent with positive selection. The MAD in Braf, Egfr, Kras, Stk11, and Tp53 correlated with spontaneous lung tumor incidence at 2 years of age. Similarly, human lung DNA showed borderline associations between MAD, age, and cumulative lung cancer risk. Together, these findings highlight MAD as an early metric of clonal expansion and a candidate cross-species biomarker of neoplastic susceptibility [23].

Once CarcSeq validation was confirmed in both humans and rodents, the next step was to assess the assay’s ability to detect mutations produced by nongenotoxic exposures. Lorcaserin, a selective 5-hydroxytryptamine 2C (5-HT2C) receptor agonist and nongenotoxic carcinogen that induces mammary tumors in rats, was administered to female Sprague Dawley at doses of 0, 30, or 100 mg/kg lorcaserin daily by oral gavage for 12 or 24 weeks. Importantly, the potential genotoxic impurity N-nitroso-lorcaserin was not detected by quantitative liquid chromatography with tandem mass spectrometry (LC–MS/MS) analyses of dosing solutions, plasma, liver, or mammary tissue, supporting the classification of lorcaserin as a nongenotoxic carcinogen. Mammary DNA was interrogated for hotspot driver mutations in Apc, Braf, Egfr, Hras, Kras, Nfe2l2, Pik3ca, Setbp1, Stk11, and Tp53, with MFs determined using the CarcSeq platform [24]. No overall differences in MF were detected across the lorcaserin dose groups; however, significant dose-dependent increases in Pik3ca H1047R mutations were observed at both 12 and 24 weeks, with higher mutant counts and mutant fractions occurring at 24 weeks. These findings indicate that lorcaserin promotes clonal expansion of spontaneous Pik3ca H1047R mutants, contributing to mammary carcinogenesis. The study further establishes CarcSeq as a sensitive approach for detecting early clonal expansion events induced by nongenotoxic carcinogens within short treatment durations [24].

The studies summarized above set the foundation for our most recent published work [25]. This study utilized the sex-related differences in lung tumor susceptibility of rasH2-Tg mice to optimize analytical strategies for quantifying clonal expansion using CarcSeq. MAD revealed significantly greater clonal expansion in the male mice, particularly when calculated from lung-specific driver mutations. Males exhibited more recurrent and phenotypically relevant mutations, supporting this interpretation. These findings validated the MAD in MF as a biomarker of clonal expansion and provided methodological guidance for cancer driver gene selection and data normalization [25]. This study further validates CarcSeq quantification of MF and MAD in MF as biomarkers of clonal expansion and potential predictors of cancer risk.

A comparison of CarcSeq with whole-genome sequencing, whole-exome sequencing, duplex sequencing, and Safe-SeqS is provided in Table 1 [8,26–29]. Conventional next-generation sequencing approaches such as whole-genome and whole-exome sequencing are optimized for comprehensive mutation discovery but lack the sensitivity required to reliably detect low-frequency somatic variants, particularly in normal tissues, due to intrinsic sequencing and PCR error rates. Although these methods are well suited for identifying clonal mutations in established tumors, variants present below approximately 1–5% VAF are often indistinguishable from technical artifacts [30,31].

Error-corrected sequencing methods, including duplex sequencing and Safe-SeqS, substantially improve analytical sensitivity through molecular barcoding and consensus sequencing [32,33]. Duplex Sequencing is a highly accurate DNA sequencing method that tags and sequences both strands of a double-stranded DNA molecule, creating "families" of reads from the original molecule, then comparing them to filter out sequencing errors and artifacts achieving significantly low error rates (>1x10^-8) in detecting rare mutations. Duplex Sequencing has been employed for mutagenesis detection and quantification because it reports only mutations present on both strands of the same DNA molecule, ensuring that observed variants represent true, fixed mutations rather than transient DNA lesions or sequencing artifacts. [34].

Safe-SeqS provides robust reduction of PCR and sequencing errors through single-strand molecular barcoding and consensus calling, enabling sensitive detection of low-frequency variants (10^-5-10^-6) with relatively modest sequencing depth, high molecular recovery, and simple library preparation [34]. Safe-SeqS is also utilized for the detection of mutagenesis by identifying increases in low-frequency single-nucleotide variants and small insertions or deletions across targeted regions following exposure to a mutagen. Its single-strand consensus approach substantially reduces technical errors, allowing statistically robust comparisons of mutation frequency and spectrum between treated and control samples, which makes it suitable for higher-throughput mutagenesis studies [35,36].

CarcSeq occupies a complementary niche by combining targeted, biologically informed selection of recurrent cancer driver mutation hotspots with single-strand consensus sequencing to achieve its sensitivity (~10^-4). CarcSeq is designed to quantify biologically relevant low-frequency mutations and their dispersion across loci. By enabling analysis of MF and median absolute deviation, CarcSeq provides a quantitative framework for assessing early clonal expansion and neoplastic potential, making it particularly well-suited for translational carcinogenicity assessment and cross-species risk modeling [8].

Recent advances in sequencing technologies are transforming the detection and characterization of early carcinogenic events. Error-corrected next-generation sequencing (ecNGS) methods—including duplex sequencing, Safe-SeqS, and molecular-barcoded approaches—have dramatically reduced technical errors, enabling detection of rare somatic variants at frequencies as low as 10^-5-10^-6. These approaches improve resolution of low-frequency driver mutations that are critical for assessing early clonal expansion in both normal and pre-neoplastic tissues [37,38].

Single-cell sequencing and spatial transcriptomics are expanding the ability to map mutational landscapes at cellular resolution, linking clonal dynamics to tissue architecture. By capturing heterogeneity across individual cells, these methods provide insight into the temporal and spatial progression of somatic mutations, clonal selection, and microenvironmental influences on tumor initiation. In parallel, ultra-deep targeted panels enable high-throughput profiling of known cancer driver mutations across large cohorts, balancing sensitivity with cost and scalability [39,40].

Computational developments complement experimental advances. Machine learning–based variant callers and consensus-calling algorithms reduce false positives and enhance the predictive accuracy of low-frequency variant detection. Integrating longitudinal sampling with these bioinformatic tools allows the reconstruction of clonal evolution over time, facilitating the identification of early biomarkers predictive of neoplastic risk [41].

Additionally, multi-omic approaches are emerging, where mutation detection is integrated with epigenomic, transcriptomic, and proteomic profiling. This holistic view enables identification of functional consequences of low-frequency mutations, linking genomic alterations to changes in gene expression, pathway activation, and early cellular transformation [38]. Together, these technological developments promise increasingly quantitative, predictive, and mechanistically informative platforms for assessing carcinogenic potential, guiding preclinical studies, and informing translational and regulatory decision-making.

CarcSeq provides a sensitive, scalable approach for detecting low-frequency cancer driver mutations across human and rodent tissues. By combining single-strand consensus sequencing with a targeted panel of recurrent CDM hotspots, the method achieves detection limits (~10^-4) that surpass conventional whole-genome and whole-exome sequencing, which are constrained by PCR and sequencing error rates. Importantly, the application of MAD to MF data captures both the prevalence and heterogeneity of mutations, enabling robust assessment of early clonal expansion in normal and tumor-adjacent tissues.

Application of CarcSeq in rodent models demonstrated that MAD correlates with spontaneous tumor incidence and highlights strain- and tissue-specific patterns of clonal expansion. Similarly, studies of nongenotoxic exposures, such as lorcaserin, revealed dose-dependent expansion of Pik3ca H1047R mutants, confirming that CarcSeq can detect subtle early neoplastic events that traditional genotoxicity assays might miss [24]. These findings support MAD as a quantitative, cross-species biomarker of neoplastic susceptibility and reinforce CarcSeq’s utility for translational and preclinical carcinogenicity assessment.

Despite its strengths, CarcSeq is inherently targeted and not intended for de novo mutation discovery, and the predictive value of MAD for human cancer risk requires further validation in longitudinal studies. Future work should expand its application across genotoxic and nongenotoxic agents and explore integration with epidemiological and clinical datasets.

This manuscript reflects the views of its author and does not necessarily reflect those of the U.S. Food and Drug Administration. Any mention of commercial products is for clarification only and is not intended as approval, endorsement, or recommendation.

This work was funded by the U.S. Food and Drug Administration.