Is a missed mutation a lost molecule? A first-cycle loss model for ARMS-PCR false negatives

Battery chemists have a precise account of a one-time loss. The first time a fresh electrode is charged, part of the electrolyte decomposes on its surface. That reaction permanently consumes a fixed quota of charge carriers, so the first cycle returns less charge than was put in. The quota scales with surface area, not with how much charge you supply. It therefore dominates when the charge budget is small, and it can be paid in advance by adding a sacrificial source of carriers ("pre-metallation"). We propose that a mutation test run on a small biopsy has the same structure. Before amplification starts, tube walls, beads and extraction membranes irreversibly bind a roughly fixed number of DNA molecules. Paucicellular samples may then lose every copy of the mutant sequence before the first PCR cycle, and no number of later cycles can recover a copy that is gone.

The clinical stakes are concrete. In a 4,600-patient thyroid cohort, allele-specific PCR for BRAF V600E missed nearly a quarter of papillary carcinomas, with 76.71% sensitivity at 100% specificity. The field cannot currently say how many of those misses are biological, meaning tumors without the mutation, and how many are technical, meaning template lost below the detection floor (BRAF V600E test on 4600 FNA specimens). The two answers call for opposite fixes: a broader gene panel, or a better-protected single-gene assay. We combine a retrospective cohort design with a fixed-quota loss model borrowed from electrochemistry to separate the two. We then test whether a dose-matched sacrificial carrier, the analogue of low-dose pre-metallation, recovers the technical misses without creating false positives.

Core Claim

The claim is that ARMS-PCR false negatives on fine-needle aspiration (FNA) specimens contain a technical component. ARMS-PCR, the amplification refractory mutation system, is allele-specific PCR in which a primer extends only when its last base matches the mutant allele. We claim this technical component follows a fixed-quota loss law rather than a proportional one, and that the difference is testable both in routine retrospective data and at the bench.

Let $n_m$ be the number of mutant template copies in the specimen and $N$ the total number of template copies. Let $C$ be the irreversible binding capacity of the pre-amplification workflow in copies, and $D$ the number of copies of added non-target carrier DNA. Under a fixed quota, bound molecules are drawn in proportion to their abundance, so the expected number of mutant copies that survive to cycle one is

$$\bar n_m = n_m \left(1 - \min\left(1, \frac{C}{N + D}\right)\right).$$

Detection then requires at least $k$ surviving mutant copies under Poisson sampling. We call the survival fraction $1 - C/(N+D)$ the assay's "initial efficiency," the direct counterpart of a battery's initial Coulombic efficiency. A proportional model instead keeps a constant fraction $1-f$ of molecules regardless of input.

The two models make different predictions. Under the quota model, misses rise sharply below an input threshold near $C$, and carrier DNA helps. Under the proportional model, misses depend smoothly on $n_m$ alone, and carrier does nothing. Too much carrier should also hurt, just as excess sacrificial salt becomes dead weight in a cell. So we expect a dose optimum rather than "more is better."

Where the Pieces Come From

The mechanism is taken from the battery literature. The first-cycle "surface tax" is described in Ultra-Low-Dose Pre-Metallation for Metal-Ion Capacitors: mobile ions are consumed at a fixed interfacial quota. The same paper cut the sacrificial additive dose from about 100 wt% to 15–30 wt% by matching the dose to the loss rather than overdosing. Work on high-surface-area hosts, including hierarchical porous carbons and 2D tellurene, shows that loss scales with exposed area. That motivates using binding area, set by extraction column type and tube material, as a covariate for $C$.

We reuse two ideas from this work: the quota-versus-supply accounting, and dose matching. We do not borrow any electrochemistry. The proposal also does not concern electrochemical biosensors, which is the bridge the research file flags as untested. That bridge would need hardware that none of the cited cohorts used.

The motivating evidence comes from the cohort–assay studies. The thyroid study reports that combining cytology with ARMS-PCR raises sensitivity to 96.62%. It also reports that 81.0% of BRAF-positive indeterminate cases were later confirmed malignant, against 4.3% of negative cases. Its limitations constrain our design. Histology exists only for the 516 resected patients, which is a verification bias. The study also names low tumor burden as a false-negative source without quantifying it. The colorectal cohort using ARMS-PCR for KRAS, NRAS and BRAF (Early CRC cohort, East Asia) serves as a second, smaller setting for testing whether the loss law transfers to other hotspots and tissue types. The KRAS G12C review (KRAS G12C NSCLC) explains why fast single-hotspot assays remain in use despite broader sequencing, and therefore why fixing them is worthwhile.

The retrospective design follows the confounding-control logic of the propensity-score prostate study (High-grade prostate cancer, propensity scoring). There the "treatment" is the extraction workflow, which changed over time and differed between operators. That study's unmeasured-confounding limitation applies here too, which is why we add a bench arm.

Research Program

The retrospective arm assembles a new FNA cohort with the same structure as the published thyroid series, at a site that archived residual washings and recorded specimen-level metadata. The metadata are extracted DNA yield, amplification cycle threshold, cellularity grade, Bethesda category, extraction kit and consumables lot, and final histology where available.

The outcome of interest is an ARMS-negative result in a histologically confirmed papillary carcinoma. We fit the quota model and the proportional model to miss status as a function of estimated template input. $C$ is allowed to vary by workflow. We correct for verification bias by weighting resected cases by the inverse of their modeled probability of going to surgery, using cytology, ultrasound features and age. Because these covariates do not decide surgery completely, we also report bounds under plausible violations of that assumption.

On the archived washings of confirmed-carcinoma, ARMS-negative cases, we run hotspot NGS. NGS here means next-generation sequencing of amplified BRAF, RAS and TERT regions using higher-input, larger-volume extraction. It provides a partial ground truth: a case carrying V600E at low allele fraction counts as a technical miss, and a case carrying a RAS mutation instead counts as a biological miss.

The bench arm tests the mechanism directly. We mix archived V600E-positive and negative washings to fix mutant copy numbers across a dilution series spanning the expected quota. Each specimen is run with no carrier and with carrier at several doses, including one matched to the fitted $C$. We also swap high-area and low-area consumables to shift $C$ deliberately.

Baselines

The naive reference point is the crude ARMS-only sensitivity on resected cases, reported as 76.71% at 100.0% specificity in the thyroid FNA cohort. Nothing is decomposed and nothing is corrected. We recompute it on our cohort with the same histology standard so the comparison is like for like.

The strongest existing clinical method is cytology combined with ARMS-PCR, reported in the same paper at 96.62% sensitivity and 88.03% specificity. It is the strongest competitor because it already recovers most misses without any new chemistry. Our intervention beats it only if carrier-rescued ARMS, combined with cytology, raises sensitivity without lowering specificity below 88.03% in the same cases under the same histology reference. A second test is whether carrier rescue alone recovers a meaningful share of ARMS misses while keeping ARMS's 100% specificity. The cost of a false positive in thyroid surgery makes that second outcome the more useful one.

The strongest statistical baseline for the decomposition is a flexible logistic regression of miss status on DNA yield, cellularity and cycle threshold. It uses spline terms and the same covariates and verification weights, and is tuned by cross-validation to the same budget as our model. The quota model must match or beat it on held-out likelihood while using fewer parameters. It must also predict the bench carrier effect, which the regression cannot express.

The key ablations are the proportional-loss model against the quota model, and at the bench, no carrier against a fixed standard carrier dose against a quota-matched dose. Together they isolate whether the fixed-quota structure and dose matching carry the improvement.

Why not just…?

Why not just run NGS on every false negative and skip the model? This is the most serious alternative, and we use it as ground truth for exactly that reason. NGS cannot be run retrospectively on cases with no residual material. It is itself input-limited, so the same molecule loss affects it. And it identifies which cases were misses without saying which workflow change would prevent them. If NGS on the recoverable subset shows that nearly all misses are biological, the technical-loss hypothesis fails. We would then report that result and recommend broader panels rather than a better single-gene assay.

Why not just run more PCR cycles? The research file notes that high-cycle amplification raises non-specific extension (thyroid FNA cohort). More basically, extra cycles amplify only the molecules that survived. If $\bar n_m$ is zero, the product stays zero. The bench arm includes a high-cycle condition to confirm this and to measure its false-positive cost.

Why not just add the standard amount of carrier DNA that many labs already use? This is a real threat to the intervention half of the proposal. If a single fixed dose performs as well as a quota-matched dose across the dilution series, dose matching adds nothing. The contribution would then shrink to the diagnostic decomposition together with evidence that carrier should be adopted routinely. The three-arm carrier ablation settles the question directly.

Why not just adjust for cellularity in a standard regression? The regression baseline above is exactly this approach. It can predict who is missed, but it has no parameter for carrier or binding area. It therefore cannot forecast the effect of changing consumables, which is the actionable output.

Proposed Experiments and Decisive Checks

The primary retrospective endpoint is the verification-weighted fraction of histologically confirmed carcinomas that were ARMS-negative and are attributable to template loss, meaning they sit below the fitted quota threshold. We report it with bootstrap intervals and validate it against NGS calls on the recoverable subset.

The first decisive check is a model comparison. The quota model must fit the input–miss curve better than the proportional model on held-out data. It must also correctly predict whether sites or kits with higher binding area have larger misses at equal input.

The second decisive check is the bench prediction. Carrier should improve detection only in the below-quota part of the dilution series and should have no effect above it. It should also show an optimum dose, beyond which detection falls or background rises. A flat carrier effect, or a benefit that is uniform across inputs, would falsify the quota mechanism.

The third check is transfer. We refit $C$ on the colorectal KRAS setting (Early CRC cohort) and test whether workflow-level quotas estimated on thyroid specimens predict misses there. That tells us whether this is a property of the assay workflow or of one specimen type.

No reported result was found on Lacuna for technical-versus-biological decomposition of ARMS-PCR false negatives, or for carrier-dose effects on FNA-based mutation detection. Every such number in this program would be new.