When the Tumor Isn't on the Slide: Analytic Validation of Plasma NGS Platforms

The clinical question underneath the result

A patient with advanced lung adenocarcinoma has a biopsy that yielded too little viable tumor for molecular testing. Repeat biopsy carries risk and delay. The oncologist orders a plasma-based next-generation sequencing (NGS) panel, and three weeks later a report arrives naming a targetable alteration — enough, potentially, to establish eligibility for a drug class. Everyone in the room treats that report as if it were tissue.

Should they? That is the question analytic validation answers, and it's a different question from "does the test help patients." Before anyone asks whether a plasma result predicts response, you have to know whether the assay reliably detects what it claims to detect, at the concentrations that actually occur in blood, without calling alterations that aren't there. Those three properties — agreement with a reference, limit of detection, and false-positive rate — are the substrate under every companion-diagnostic claim built on plasma. Get them wrong and every downstream inference inherits the error.

What the assay measures, and what a result can mean

Cell-free DNA (cfDNA) is the pool of short DNA fragments shed into plasma by dying cells throughout the body. A fraction of that pool — the circulating tumor DNA, or ctDNA — comes from tumor cells and carries their somatic alterations. In advanced cancer that fraction can be substantial; it can also be vanishingly small, sometimes well under 1% of total cfDNA. Plasma NGS platforms extract cfDNA, build sequencing libraries, and interrogate a defined set of genes for mutations, copy-number changes, and in some cases fusions and signatures such as microsatellite instability.

The design consequence is unforgiving. You are hunting for a handful of variant molecules in a background of normal DNA, so sensitivity depends entirely on how low an allele fraction the assay can reliably see and how well it suppresses sequencing and amplification noise. A positive result says a tumor-derived alteration was detected in plasma. A negative result is harder to read: it may mean the alteration is absent, or that the tumor simply wasn't shedding enough ctDNA at the moment of the draw. That asymmetry is why negative plasma results are generally not treated as definitive, and why reflex to tissue is standard when plasma comes back negative for a clinically expected driver.

The evidence

Two platforms anchor most of the clinical conversation, and both have published analytic validation studies with concrete metrics.

Guardant360 was validated by Odegaard and colleagues in 2018 [1]. Against orthogonal reference methods, the assay reported positive and negative percent agreement both exceeding 99% [1]. Its reported limit of detection reached a variant allele fraction of roughly 0.02–0.04% for certain alteration classes [1] — an extraordinarily low floor, reflecting the reality that clinically relevant ctDNA often circulates at well under 1%. Positive percent agreement (PPA) is the share of true positives the assay correctly calls; negative percent agreement (NPA) is the share of true negatives it correctly excludes. Both running above 99% means that, within the validation set, the platform rarely missed a present alteration and rarely invented an absent one.

FoundationOne Liquid CDx, a 324-gene cfDNA assay, was validated by Woodhouse and colleagues in 2020 [2]. They reported a 95% limit of detection at a variant allele fraction of 0.40% — the allele fraction at which the assay detected a true variant 95% of the time [2]. Across variant types the study reported PPA of 96.3% and NPA exceeding 99.9% [2]. The false-positive rate was 0.013% [2], meaning erroneous positive calls were rare on a per-position basis.

The two LoD figures look dramatically different — 0.02–0.04% versus 0.40% — but reading them as a head-to-head ranking is a mistake. They are defined differently. Odegaard's figure describes detection capability for particular alteration classes [1], while Woodhouse's is explicitly the allele fraction at which detection reaches 95% probability [2]. Comparing them directly is comparing two different statistical constructs, and the comparison doesn't support a claim that one platform is an order of magnitude more sensitive than the other. The low per-position false-positive rate in the FoundationOne data [2] also doesn't translate cleanly into a per-report false-positive rate, because a panel interrogates thousands of positions and small per-position probabilities accumulate across a genome-scale assay.

Both studies validate the analytic layer — detection, agreement, noise — not the clinical premise that a plasma-detected alteration behaves identically to a tissue-detected one for every therapeutic decision. Clinical validity and clinical utility are separate questions resting on separate evidence.

Where this stands

Both platforms are FDA-approved and carry companion-diagnostic indications, so their analytic validation has cleared regulatory review rather than living purely in the investigational literature. That status is meaningful: the metrics above were generated under the kind of scrutiny that supports a companion-diagnostic claim, not as exploratory benchmarks.

Several limitations shape how these numbers should be read in practice. The agreement statistics were established against validation cohorts and reference methods specific to each study [1,2], and real-world performance depends on preanalytic variables — collection tube, time to processing, cfDNA yield — that a validation report standardizes but a clinic may not. The dependence on ctDNA shedding means a low-shedding tumor can produce a false-negative plasma result even on a flawlessly performing assay, which is a biological limit, not an analytic defect.

Clonal hematopoiesis is the genuinely unresolved problem. Blood cells acquire somatic mutations with age, and those mutations enter the cfDNA pool. A plasma assay can detect a real alteration that originates in hematopoietic cells rather than the tumor — analytically a true positive, clinically a confounder. Neither validation study resolves how often this contaminates a given clinical interpretation, and distinguishing tumor-derived from blood-derived variants remains an active area without a settled standard.

Cross-platform harmonization lacks a single regulatory reference standard, which means a variant allele fraction reported by one assay cannot be assumed numerically equivalent to the same figure from another; the LoD definitions above are a direct illustration. A clinician comparing serial results, or a tumor board weighing two reports on the same patient, cannot treat the numbers as interchangeable without understanding how each was derived.

So the practical implication for that lung adenocarcinoma case is narrower than the three-week report suggests: a positive plasma call for a targetable driver, on a validated platform, carries agreement with tissue in the high-90s and a low false-positive rate [1,2] — but a negative call for the same driver leaves the question open, and the next decision is whether the tumor was shedding at all.

References

1. Odegaard JI, et al. Validation of a Plasma-Based Comprehensive Cancer Genotyping Assay Utilizing Orthogonal Tissue- and Plasma-Based Methodologies. Clin Cancer Res. 2018. PMID 29691297; DOI 10.1158/1078-0432.CCR-17-3831.
2. Woodhouse R, et al. Clinical and analytical validation of FoundationOne Liquid CDx, a novel 324-gene cfDNA-based comprehensive genomic profiling assay for cancers of solid tumor origin. PLoS One. 2020. PMID 32976510.