Reading Past the Heavy Hitter: Competitor-Conditioned Adversarial Calling for Multiplex PCR

When several pathogens are tested in one PCR tube, a strongly positive target can hide a weak one. PCR (polymerase chain reaction) copies DNA in repeated cycles, and an instrument reports a target as positive when its fluorescence curve crosses a threshold. In a shared tube, an abundant target amplifies early and uses up the polymerase enzyme and nucleotide building blocks. A co-present target at a few copies then starts late, plateaus low, and is often called negative. This is one face of the multiplexing sensitivity trade-off: adding targets to a panel raises its limit of detection, the smallest copy number it reliably finds. We propose to treat part of this loss as a reading problem rather than a chemistry problem. The weak target's curve often still carries evidence, but its shape is distorted by a nuisance variable that the tube itself reports, namely how much the other channels amplified.

The core idea fits in two sentences. Train a curve-reading detector that sees all channels of a multiplex run. Add an adversary, a second network trained to recover the competitors' load from the detector's output for each target, and train the detector to defeat it, so that the detection score for a target depends on whether that target is present and not on how hard its neighbours amplified. This borrows the adversarial training game from domain adaptation, where a model is trained so that a discriminator cannot tell which domain an input came from. Here the "domains" are competitor-burden regimes, and the alignment is placed on the decision rather than on the internal features. We expect this to recover low-copy calls that threshold calling misses, and to generalize to competitor combinations never measured in the lab.

Core Claim

The claim is that conditioning an adversary on an observed, mechanistically motivated nuisance can close part of the gap between multiplex and singleplex limits of detection for curves that did amplify. The nuisance is the competitor burden, estimated from the other channels of the same tube. The adversary also conditions on the target's own label, and it acts on the detection score. We expect this to do better than either plain supervised learning or standard domain-invariant feature learning when labeled multiplex data covers only a small part of the possible combinations of target concentrations.

Three design choices separate this from the obvious transplant of domain-adversarial training onto PCR curves. First, the adversary acts on the detection score, not the shared representation. Removing burden information from the features would forbid the model from using the competitor channels to undo the distortion, and that use is the whole point. Second, the adversary is conditioned on the target's own presence. Unconditioned alignment would match the score distributions of high-burden and low-burden tubes. In real panels, co-infection makes burden and positivity correlated: Chen et al. report a 13.65% co-infection rate. Matching those distributions would therefore erase true signal. This is the mode-mismatch failure that Conditional Adversarial Domain Adaptation identified for marginal alignment. Third, the labels are multi-label. Each target gets its own binary detector and its own adversary, because targets co-occur rather than exclude one another. Both CDAN and Multi-Adversarial Domain Adaptation assume mutually exclusive classes.

We make the objective explicit. Let $x$ be the multi-channel amplification curves of one tube and $s_k(x)$ the detector's logit for target $k$. Let $y_k$ be target $k$'s presence label, and let $b_{-k}$ be the competitor burden, defined as the log of the largest estimated starting copy number among the other channels, quantized into bins. Each adversary $D_k$ is trained to predict $b_{-k}$ from the pair $(s_k, y_k)$. The detector minimizes

$$\sum_k \Big[ \ell(s_k(x), y_k) - \lambda \, \ell_{adv}\big(D_k(s_k(x), y_k), b_{-k}\big) \Big],$$

where $\ell$ is binary cross-entropy for detection, $\ell_{adv}$ is the adversary's cross-entropy, and $\lambda$ sets the strength of the trade-off. A gradient reversal layer implements the min-max in one backward pass. When the adversary is at chance, the score for target $k$ is independent of burden given presence. For low-copy positives, that is exactly "sensitivity does not depend on the neighbours."

Unlabeled field runs can enter the same objective. For them, $y_k$ is replaced by the detector's own prediction, weighted by CDAN's entropy rule $w = 1 + e^{-H}$ so that uncertain samples count less. We treat this substitution as a risk to be tested rather than assumed, because it is circular: the score conditions on itself.

We also state the limit up front. No reading method recovers a target whose primers never productively amplified. The graph-theoretic phase transition described by Asodi et al., in which the chance of finding a compatible primer set collapses past a critical panel size, and the primer-dimer drain characterized in Multicolor Combinatorial Probe Coding for Real-Time PCR are design-level losses. This proposal targets only the part of the penalty caused by late, suppressed, sub-threshold amplification.

Research Program

The prior work plays three roles. Two papers motivate the gap. AC-DC: Amplification Curve Diagnostics for Covid-19 Group Testing shows that curves carry more than a binary call: it quantizes the threshold-crossing cycle, $C_t$, into levels to resolve pooled samples hidden behind "heavy hitters." However, it uses hand-set thresholds, handles one target, and its calibration breaks across instruments and master mixes. Our model learns the reading, handles cross-target competition rather than sample pooling, and tests instrument transfer explicitly. Pre-capture multiplexing provides additional power to detect copy number variation in exome sequencing shows that a multiplexed reaction can act as its own internal control. It does this by physically synchronizing noise and modeling it with a negative binomial. We pursue the same internal-control intuition computationally, by conditioning on co-measured channels instead of relying on physical pooling.

Four papers supply methods. CDAN supplies prediction-conditioned adversaries and entropy weighting. MADA supplies per-category discriminators, which we recast as per-target, multi-label discriminators. Adversarial Domain Adaptation for Classification of Prostate Histopathology Whole-Slide Images supplies the lesson that adversarial alignment needs a task-preserving constraint: there, patches from the same slide had to agree. Our analogue is that replicate wells of the same spike-in must receive consistent scores. DYNAMIC: A Novel Software Implementation of a Kinetic Model of TaqMan PCR is a differential-equation simulator of singleplex TaqMan PCR. It reached scaled MSE 0.074 ± 0.022 on the CHK assay and was intended as a "digital twin" for producing training data. We use it only as an optional pretraining source.

Two lines of evidence constrain the design. FAN-Net: Fourier-Based Adaptive Normalization For Cross-Domain Stroke Lesion Segmentation found that aggressive style normalization destroyed structure once its frequency cutoff rose above 0.1. That warns us that invariance applied too broadly removes signal, which is why our adversary acts narrowly on scores. DYNAMIC's own stated limits warn that a multiplex simulator would face unidentifiable parameters and high runtime. We therefore do not make the method depend on extending it.

The build has three parts. The first is a spike-in matrix on the four-plex porcine panel of Establishment and application of multiplex real-time PCR for simultaneous detection of four viruses associated with porcine reproductive failure, which targets PCV2, PCV3, PPV and PRV. Its equimolar tandem plasmid is useful for standard curves but cannot create competition, so we instead use the individual plasmid fragments at deliberately unequal ratios. Each target is run at low copy numbers spanning its reported limit of detection, against competitors from zero up to very high loads, with replicate wells. Specific competitor pairings and burden levels are held out of training.

The second part is a model. A small one-dimensional convolutional encoder reads all fluorescence channels. It feeds four per-target logit heads and four score-level adversaries. The burden $b_{-k}$ is estimated from each competitor channel's standard curve.

The third part is a transfer cohort. These are field-sample runs, ideally on a second instrument or master mix, and they are unlabeled during training. They are reference-tested afterwards by singleplex retesting.

Baselines

The naive reference point is the instrument's default threshold call on each channel. This is the procedure whose limits of detection Chen et al. report for the four-plex: 1 copy per reaction for PCV2 and 10 copies per reaction for PCV3, PPV and PRV, with over 90% concordance with singleplex PCR on 315 field samples. We will re-measure it on our matrix stratified by competitor burden, because the published figures do not isolate high-burden conditions. Beating it means a lower 95% limit of detection in the highest burden bins, with specificity matched.

The strongest competitor is a joint supervised multi-channel classifier. It uses the same encoder and heads and sees the competitor channels, but has no adversary. It is the strongest because, given enough labeled combinations, it can learn to explain away competition directly. It gets the same training wells, the same architecture and the same hyperparameter search budget, which is a fixed number of trials scored on a held-out validation slice of labeled spike-ins. No reported result found on Lacuna for any learned curve-calling model on multiplex PCR, so this baseline's numbers will come from our runs. We count it as beaten only if the gain appears on the held-out burden levels and pairings, not merely on combinations seen in training.

Two domain-adaptation baselines test whether the specific conditioning matters. The first is an unconditional domain-adversarial network that aligns features across burden regimes, in the style of the gradient-reversal architectures above. The second is CDAN+E recast as per-target multi-label alignment. For context only, CDAN+E reported 87.7% average accuracy on the Office-31 vision benchmark against 82.2% for the unconditional version. Those numbers do not transfer to curves, but they set the expectation that conditioning matters. Both baselines use the same encoder and tuning budget as our model.

A domain-knowledge baseline regresses each target's $C_t$ shift on the competitors' estimated loads and recalibrates its threshold. This is AC-DC-style hand modeling extended across targets.

The ablations isolate the key ingredient. One moves the adversary from scores onto features. Another removes the label conditioning. A third replaces the burden variable with a binary "multiplex versus singleplex" domain label. A fourth drops the unlabeled field runs.

Why not just…?

Why not just train a supervised multi-channel model on more spike-ins? This is the real threat. If the lab could cover the space of concentration combinations densely, plain supervision would probably match our method. The argument for the adversary rests on coverage: four targets, several low-copy levels and several competitor levels already give hundreds of conditions before replicates, and larger panels grow combinatorially. The decisive test is a learning curve over the fraction of the matrix used for training, evaluated on held-out pairings. If the supervised model matches ours at every coverage level, we will report that the adversary adds nothing and that the gain came from multi-channel input alone.

Why not just correct each target's threshold by regressing on competitor $C_t$? A linear shift captures the delay a competitor causes, but not the lowered plateau and flattened slope of a starved reaction. It also inherits the cross-instrument brittleness the file documents for AC-DC's calibrated thresholds. Our instrument-transfer experiment measures this directly. If the regression's sensitivity holds up on the second instrument, the simpler method wins in that regime and we will say so.

Why not just simulate multiplex PCR and train on synthetic curves? DYNAMIC models only singleplex reactions. Its authors' limits, as listed in the file, include no inter-assay competition and no dNTP depletion, and parameter fitting is costly. A multiplex extension would add unidentifiable parameters and a gap between simulation and real runs. We test a cheap version anyway: DYNAMIC with a crude shared reagent pool, used as a pretraining source. If it closes the high-burden gap on its own, the adversarial term is unnecessary.

Proposed Experiments

The primary metric is the multiplex penalty. For each target and burden bin, it is the $\log_{10}$ ratio of the 95% limit of detection under competition to the 95% limit with competitors absent, where each limit is estimated by probit regression over replicate wells and specificity is held at 99% on true negatives in the same bin. The secondary metric is the gap in the adversary's burden-prediction accuracy above chance, which checks that the invariance actually holds. We also report agreement with singleplex retesting on the field cohort, broken out by co-infected versus singly infected samples. Co-infected samples are where heavy hitters live, and with PCV2 at 66.67% prevalence, PCV2 will often be the hitter.

There are three decisive checks. The first is whether our method lowers the multiplex penalty on held-out pairings below that of the joint supervised baseline across the coverage curve. Gains seen only on training combinations would not count. The second is whether it does so without raising false positives in high-burden negative wells. That is the failure we expect if the model learns "competitor is high, so hedge positive." The third is whether label conditioning is necessary. We predict that unconditional alignment will lower sensitivity for truly co-infected tubes. If it does not, the CDAN-motivated argument fails for this problem.

A further check addresses scope. On wells where the target channel shows no rise above baseline at all, every method should fail. Any apparent recovery there would indicate leakage from competitor channels, not genuine detection.