From somatic variant call to targeted therapy — tiered evidence frameworks that connect a driver mutation to an FDA-approved or investigational drug
Every precision-oncology decision begins with a base-pair-resolution question: at this genomic position, in this tumor, what fraction of DNA molecules carry a mutant base instead of the reference? Targeted next-generation sequencing (NGS) panels — covering anywhere from 50 to more than 500 cancer-relevant genes — align millions of short reads to the reference genome, call variants with statistical confidence, and report a variant allele frequency (VAF) that becomes the first quantitative signal in the entire mutation-matching pipeline.
A variant call is a statistical inference, not a direct observation of a single molecule. Sequencing library preparation fragments tumor DNA, adapters are ligated, and the library is amplified and sequenced on a short-read platform (typically Illumina). Reads are aligned to the reference genome (GRCh38) with an aligner such as BWA-MEM, and a variant caller (Mutect2, VarDict, Pindel for indels) scans each targeted base for evidence of a mismatch against the reference.
Variant allele frequency (VAF) = (reads supporting the alt allele) / (total reads at that position). For KRAS c.34G>T (p.G12C), if 320 of 1,000 total reads at codon 12 carry the T instead of G, VAF = 32%. This number is deceptively simple but encodes several confounded biological quantities at once: tumor purity (what fraction of the sampled tissue is actually cancer vs. stroma/immune infiltrate), tumor ploidy and local copy number at that locus, and the clonality of the mutation — whether it is present in every cancer cell (clonal, "trunk" mutation) or only a subpopulation (subclonal, "branch" mutation).
Clinical-grade callers apply strand-bias filters, base-quality filters (Phred ≥30), and panel-of-normals subtraction to remove sequencing artifacts and germline variants before a somatic call is reported. False-positive suppression is critical: a laboratory reporting an oncogenic driver that is actually a sequencing artifact can send a patient down an entirely wrong treatment pathway.
A VAF of 32% at a diploid, 100%-pure locus corresponds to a heterozygous mutation present in essentially every tumor cell (VAF ≈ 50% × purity). Real samples are messier: circulating tumor DNA (ctDNA) from liquid biopsy often shows VAFs of 0.1–5%, requiring UMI-barcoded deep sequencing (>10,000×) to distinguish a true low-frequency variant from sequencing noise.
Tumor tissue biopsy (FFPE block or fresh-frozen) samples one physical region of one tumor at one point in time — subject to spatial heterogeneity, since different regions of the same tumor can carry different subclonal mutations. Liquid biopsy sequences circulating tumor DNA (ctDNA) shed into plasma from multiple tumor sites simultaneously (including metastases), giving a more integrated — but far more dilute — picture: ctDNA typically represents <1% of total cell-free DNA in early-stage disease, rising to 10%+ in high-burden metastatic disease.
Serial ctDNA VAF measurements are increasingly used to track treatment response in near-real time: a falling VAF over weeks indicates tumor response, a rising VAF after nadir indicates emerging resistance — often before imaging (RECIST) would detect progression. This makes VAF not just a one-time diagnostic threshold but a longitudinal biomarker.
Modern comprehensive genomic profiling (CGP) panels use hybrid-capture bait sets to enrich for exons of hundreds of cancer genes plus select introns (to catch fusion breakpoints in genes like ALK, ROS1, RET, NTRK1/2/3). Deep sequencing (500–1000× mean coverage) is required because tumor content is rarely 100% pure and subclonal variants can be present at VAF well under 10%.
Base-level QC (mapping quality, base quality, strand-bias, and homopolymer-context filters) removes the two dominant artifact sources: FFPE-fixation-induced C>T/G>A deamination and PCR duplication skew. Unique molecular identifiers (UMI) — random barcodes ligated before amplification — allow deduplication back to the original template molecule, which is the single most important technical advance enabling reliable low-VAF (<1%) ctDNA calling.
A tumor genome carries thousands of somatic mutations, but only a small handful are causally responsible for driving malignant growth — the rest are "passengers," mutations that accumulated as a side effect of genomic instability but confer no selective growth advantage. Separating the one actionable driver from the surrounding mutational noise is the essential filtering step between raw sequencing output and a clinically meaningful report.
The single strongest computational signal for driver status is mutational recurrence at the exact same amino acid position across independent tumors and patients — a "hotspot." KRAS G12C, G12D, and G12V; EGFR L858R; BRAF V600E; PIK3CA H1047R — these positions are mutated far more often than chance mutation rate would predict, because a mutation there confers a specific, reproducible gain-of-function biochemical change (e.g., impaired GTP hydrolysis locking KRAS in its active, GTP-bound signaling state).
Databases such as COSMIC (Catalogue Of Somatic Mutations In Cancer), cBioPortal, and OncoKB aggregate mutation calls across tens of thousands of sequenced tumors, allowing statistical hotspot detection methods (e.g., the MutSigCV / OncodriveCLUST family of algorithms) to flag positions with mutation density statistically inconsistent with background mutation rate given gene length, expression level, and replication timing.
A variant seen at the exact same codon in >20 independent tumor samples across a hotspot database is treated with a very different prior probability of pathogenicity than a private missense variant seen only once — even before any functional assay is performed.
Beyond recurrence, in silico tools estimate the biophysical impact of a missense change: PolyPhen-2 and SIFT predict effects on protein structure/function from sequence conservation; CADD and REVEL integrate dozens of annotations (conservation, splicing impact, regulatory context) into a single deleteriousness score; AlphaMissense (DeepMind, 2023) uses structure-aware deep learning trained on the human proteome to classify every possible missense variant genome-wide as likely benign or likely pathogenic with high concordance to clinical ClinVar labels.
Structural context matters enormously: a mutation at a kinase ATP-binding pocket, a GTPase switch region, or a tumor-suppressor DNA-binding domain is mechanistically far more plausible as a driver than one in a flexible, non-conserved loop. Structural biology (crystal structures, cryo-EM, and AlphaFold models) is routinely consulted to interpret whether a novel variant plausibly disrupts protein function.
The great majority of individually rare missense variants encountered in a clinical report cannot be confidently classified as driver or passenger from bioinformatics alone — these are reported as "variants of unknown significance" (VUS). A VUS is not actionable: it cannot be matched to a targeted therapy, and reporting laboratories are required to distinguish it clearly from confirmed oncogenic drivers to avoid inappropriate treatment decisions.
Reclassification over time is common and important: as functional assays (saturation mutagenesis screens, e.g., using deep mutational scanning of a target protein) and additional tumor sequencing accumulate, a fraction of VUS calls are reclassified as likely oncogenic or likely benign — a process analogous to ACMG/AMP germline variant reclassification in hereditary disease genetics.
Once a variant is confirmed as a plausible oncogenic driver, it must be sorted into a standardized tier that communicates, unambiguously, how strong the evidence is for clinical actionability. The 2017 joint consensus recommendation from the Association for Molecular Pathology, American Society of Clinical Oncology, and College of American Pathologists (AMP/ASCO/CAP) established the four-tier system now used in essentially every molecular pathology report issued in the United States.
Tier I variants have FDA-approved therapy directly tied to the biomarker (Level A), or are included in professional guidelines (NCCN, ASCO, ESMO) as a validated predictive, prognostic, or diagnostic biomarker for a specific tumor type (Level B). Examples: EGFR exon 19 deletion / L858R in lung adenocarcinoma (osimertinib), BRAF V600E in melanoma (dabrafenib + trametinib) and in colorectal cancer (encorafenib + cetuximab), KRAS G12C in NSCLC (sotorasib, adagrasib), and ALK fusions in NSCLC (alectinib, lorlatinib).
Tier I calls trigger direct, guideline-concordant treatment recommendations and are the primary drivers of on-label prescribing in molecular oncology today.
Tier II variants show promising but not yet definitive clinical evidence: FDA-approved therapy for a different tumor type than the one in front of the pathologist (Level C — the classic "off-label, strong rationale" scenario), or multiple small studies / case series with some consensus (Level D). A KRAS G12C mutation detected in colorectal cancer — a setting with weaker single-agent sotorasib activity than lung cancer, motivating combination strategies — is a canonical Tier II scenario: same actionable gene, different tumor-type evidence strength.
Tier II calls typically justify clinical trial referral or off-label use discussion, particularly at a molecular tumor board, rather than an immediate guideline-directed prescription.
Tier III captures variants of unknown clinical significance: not present at sufficient frequency in population/tumor databases to call benign, but lacking convincing functional or clinical evidence of oncogenicity — this is the VUS category. Tier III variants are reported for completeness (particularly if a companion research trial exists) but do not drive treatment decisions.
Tier IV variants are benign or likely benign: present at high frequency in general population databases (gnomAD) inconsistent with a rare disease-causing allele, or with functional evidence of neutral effect. These are typically not included in the primary clinical summary of an oncology report at all, though they remain in the underlying variant call file.
A critical operational point: tier assignment is tumor-type-specific and re-evaluated over time. A variant that is Tier II today can become Tier I after a pivotal trial reads out and a companion FDA approval follows — molecular pathology reports are living documents, re-analyzed as evidence accrues.
The same molecular alteration can occupy different tiers in different tumor types simultaneously. BRAF V600E is Tier I (FDA-approved, guideline-directed) in melanoma, thyroid cancer, and colorectal cancer, but was historically only Tier II/III in rarer tumor types until tissue-agnostic approvals (e.g., dabrafenib+trametinib, 2022) extended coverage across solid tumors harboring the mutation.
Tier classification tells a clinician how strong the evidence is; the OncoKB level-of-evidence system and its underlying knowledgebase tell them specifically which drug to consider. Memorial Sloan Kettering's OncoKB (Chakravarty et al., JCO Precision Oncology 2017) is the first FDA-recognized somatic variant knowledgebase, curating the therapeutic implications of specific mutations gene-by-gene, tumor-type-by-tumor-type, down to the level of an individual amino acid substitution.
OncoKB assigns each variant-drug-tumor-type triplet a level:
• Level 1: FDA-recognized biomarker predictive of response to an FDA-approved drug in this specific indication (e.g., KRAS G12C + sotorasib in NSCLC). • Level 2: Standard-of-care biomarker per professional guidelines, predictive of response, though not necessarily an FDA companion diagnostic requirement. • Level 3A: Compelling clinical evidence supports the biomarker as predictive of response, typically from a completed clinical trial in the same indication, short of full approval. • Level 3B: Compelling clinical evidence in a different indication than the one being evaluated — the classic basis for off-label or basket-trial consideration. • Level 4: Compelling biological evidence (preclinical, mechanistic) supports the biomarker as predictive of response, but clinical evidence is not yet established. • Level R1/R2: the variant is associated with resistance to a given drug — a "do not use" signal rather than a "use" signal, equally critical for treatment selection (e.g., EGFR T790M conferring resistance to first-generation EGFR TKIs, prompting a switch to osimertinib).
Levels R1 and R2 are frequently overlooked but clinically decisive: matching therapy is not only about finding a drug a mutation responds to, it is equally about avoiding a drug a resistance mutation predicts will fail — for example, RAS-pathway mutations predicting anti-EGFR antibody resistance in colorectal cancer.
For three decades KRAS was considered "undruggable" — its GTP-binding pocket is picomolar-affinity and essentially featureless, defeating attempts at competitive small-molecule inhibition. The G12C mutation itself provided the opening: it substitutes glycine 12 for cysteine, introducing a reactive thiol side chain adjacent to a shallow pocket beneath switch-II that only exists in the GDP-bound (inactive) conformation.
Sotorasib and adagrasib are irreversible covalent inhibitors: they diffuse into this switch-II pocket and form a covalent bond specifically with the mutant cysteine's thiol group — the wild-type glycine has no equivalent chemistry, so these drugs are structurally incapable of engaging normal KRAS. This is the molecular literalization of "lock and key": the mutation does not just create a target, it creates the exact chemical handle the key was designed to grip. CodeBreaK 100 (sotorasib) demonstrated ~37% objective response rate and 6.8-month median PFS in previously treated KRAS G12C NSCLC, leading to accelerated FDA approval in May 2021.
EGFR exon 19 deletion / L858R + osimertinib: these mutations in the kinase domain increase ATP affinity and constitutively activate downstream MAPK/PI3K signaling. Osimertinib is a third-generation, mutant-selective, irreversible EGFR TKI, FLAURA trial showing median PFS of 18.9 months first-line versus 10.2 months for first-generation EGFR TKIs, plus superior CNS penetration.
BRAF V600E + dabrafenib/trametinib: V600E mimics constitutive phosphorylation-independent kinase activation; dabrafenib (BRAF inhibitor) is combined with trametinib (downstream MEK inhibitor) to block paradoxical MAPK-pathway reactivation that occurs with BRAF inhibitor monotherapy — the combination strategy itself is a direct consequence of understanding the pathway biology downstream of the matched mutation.
ALK fusion (most commonly EML4-ALK) + alectinib: a chromosomal rearrangement fuses the ALK kinase domain to a constitutively dimerizing partner, driving ligand-independent kinase activity. Alectinib is a highly selective, CNS-penetrant ALK inhibitor; the ALEX trial showed median PFS >34 months versus 10.9 months for crizotinib, one of the largest PFS improvements ever recorded in a randomized targeted-therapy trial. Each pairing above is a distinct instance of the same abstract matching template: characterize precisely how the mutation rewires a protein's conformation or regulation, then select or design a molecule that exploits exactly that structural consequence.
A confirmed Tier I / Level 1 match does not automatically mean a patient walks out with a prescription. The final step translates molecular evidence into one of three concrete clinical pathways, each with different regulatory standing, insurance-coverage implications, and expected outcome probability: an on-label FDA-approved indication, an off-label prescription supported by strong cross-tumor-type evidence, or enrollment in a biomarker-matched clinical trial.
The cleanest and highest-probability pathway: the detected variant matches an FDA drug label's specified biomarker and tumor type exactly (OncoKB Level 1 / AMP Tier I). Prescribing requires no special exception, insurance coverage is typically straightforward, and expected response rates reflect the pivotal trial data that supported approval — for example, ~80% objective response rate for osimertinib in EGFR-mutant treatment-naive NSCLC.
Tissue-agnostic approvals extend this pathway across tumor types for select biomarkers regardless of primary site: pembrolizumab for any MSI-high/dMMR solid tumor (2017, the first tissue-agnostic approval), larotrectinib and entrectinib for any NTRK fusion-positive tumor, and dabrafenib+trametinib for any BRAF V600E-mutant solid tumor (2022) — each reflecting growing regulatory acceptance that molecular alteration, not organ of origin, can be the primary determinant of drug activity.
When a variant is well-matched to a drug approved for a different tumor type (OncoKB Level 3B, AMP Tier II), physicians may pursue off-label prescription supported by case-series data, expanded-access programs, or compendia listings (NCCN Compendium, which many payers accept for coverage decisions). This pathway requires more active advocacy: prior authorization appeals, molecular tumor board documentation, and sometimes single-patient IND ("compassionate use") applications to the FDA.
Off-label biomarker-matched prescribing, in aggregate meta-analyses (e.g., the pan-cancer analyses feeding NCI-MATCH design), shows meaningfully higher response rates (roughly 20–30%) than empirical, biomarker-blind chemotherapy in later treatment lines (roughly 5–10%) — the rationale that justifies pursuing this harder administrative pathway.
For variants with only preclinical or early-phase evidence (OncoKB Level 4, AMP Tier II/III), or for approved-pathway patients whose disease has progressed, biomarker-matched clinical trials are frequently the highest-yield option. Two trial architectures dominate modern precision oncology:
• Basket trials enroll patients across many different tumor types who all share one molecular alteration, testing one targeted drug against that shared target (e.g., NCI-MATCH, TAPUR, the original vemurafenib BRAF V600E basket trial that first showed activity outside melanoma). • Umbrella trials enroll patients with one tumor type (e.g., NSCLC) and assign each patient to a different targeted arm based on their specific detected mutation, run under one master protocol (e.g., Lung-MAP).
Trial matching is typically coordinated by a molecular tumor board — a multidisciplinary panel of oncologists, pathologists, and genetic counselors who review the full molecular report and current open-trial landscape (searchable via ClinicalTrials.gov and site-specific matching platforms) to identify the best-fit protocol for a given patient.
The overall probability of finding a clinically actionable, high-confidence match varies enormously by tumor type: >60% of lung adenocarcinomas now have a Tier I actionable driver, while fewer than 10% of pancreatic adenocarcinomas do — reflecting real biological differences in how oncogene-addicted these cancers are, not a limitation of the matching framework itself.