Kinetic Unmixing of Multi-Level Color Codes in Mixed-Sample PCR

A standard real-time PCR machine reads only four to six colors. Each target sequence is still tagged with a unique blend of those colors, and some of its probes carry low, medium or high amounts of a given dye. With that scheme, one tube can screen dozens to thousands of targets. The scheme fails when a sample contains several targets at once, such as a co-infection, a compound heterozygote, or a tumor carrying several mutations. The color codes of the targets add together, and the sum can look like the code of a target that is absent. This is the same masking problem that bulk sequencing faces, where averaging over many cells hides which components are present. We propose to remove the ambiguity without any new hardware by reading the whole amplification curve rather than the final brightness. Targets that start at different amounts cross the detection threshold at different cycles. The four-color time series is therefore a sum of time-shifted curves that can be decomposed, much as a bulk profile is decomposed into the cell types that make it up.

The core claim is that two ingredients together make multi-target mixtures decodable on unmodified four-channel instruments: a codebook whose multi-level codewords stay distinguishable when added together, and a decoder that uses per-target timing. We expect the gain to be largest exactly where endpoint reading fails, in mixtures whose components differ in abundance. We also expect it to shrink toward the codebook-only limit when abundances are equal, and we will test that prediction directly.

Background and Gap

Multicolor Combinatorial Probe Coding for Real-Time PCR (MCPC) labels each target's probe with a subset of four fluorophores. This gives 15 binary codes. The authors validated it on a 15-type HPV plasmid panel, on a 10-plex foodborne pathogen panel with a limit of detection of 5–10 CFU per reaction, and on 239 blood samples carrying 13 β-globin genotypes. In the β-globin work, dosage-like intensity levels already separated codes such as 1002 from 2002. The authors project that low, medium and high tiers on four colors could exceed 265,000 targets. They also state the failure mode plainly: when targets co-exist, their codes sum, so 1000 plus 0100 reads as 1100. They offer no algorithm for unmixing superimposed multi-level vectors. That missing algorithm is our starting point.

Two lines of work supply our methods. Multi-level equilibrium signaling for molecular communication decodes 4 or 8 concentration levels by averaging repeated observations and using a sequence (Viterbi-style) decoder. It also constrains us: eight levels needed many repeated samples to overcome noise. We treat the 40-odd cycle readings of a PCR run as those repeated observations, and we expect three levels per channel to be the practical ceiling, not eight. Count Bridges enable Modeling and Deconvolving Transcriptomic Data treats each component's contribution to an aggregate as a latent variable inside an expectation-maximization loop. It projects samples so that the components sum exactly to the observed total. We reuse that latent-contribution structure, with integer code levels in place of mRNA counts.

The deconvolution benchmarks explain why this is hard. In breast cancer pseudo-bulk mixtures, methods confuse normal epithelium with cancer epithelium because the two share expression programs (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures). In DECONbench, adding a second data type cut error by only 2.12%. In our setting, the analogue of "components that look alike" is two targets amplifying with the same timing. That case should be the hardest, and the codebook alone has to resolve it.

The bordering combinatorial-pooling work avoids collisions rather than decoding them. SCITO-seq cuts cross-cell-type collisions from 25.7% to 8.9% by staining cells in separate pools. Robust and cost-efficient single-cell sequencing through combinatorial pooling assigns samples to pools under a combinatorial design. Both use binary presence or absence. Both work in sequencing, where new barcode sequences are nearly free, so adding intensity tiers would buy little there. That is why we place the idea in optical PCR, where the number of channels is truly fixed. Bitbow adds a further warning. Its theoretical 32,767 codes are not used uniformly, because some bits flip more often than others. We therefore will not assume that every codeword in our codebook behaves equally well, and we will measure the error rate of each codeword.

Core Mechanism

Write $F_c(t)$ for the fluorescence in channel $c$ at cycle $t$. Let $a_{jc}\in\{0,1,2\}$ be target $j$'s code level in channel $c$, set by the fraction of its probe carrying that dye. Let $s(\cdot)$ be a shared sigmoid amplification shape, and let $\tau_j$ be target $j$'s threshold cycle, which tracks its starting amount on a log scale. Then for a sample containing the target set $S$,

$$F_c(t) \approx \sum_{j\in S} a_{jc}\, g_j\, s(t-\tau_j) + \epsilon_c(t),$$

where $g_j$ is a per-target efficiency gain and $\epsilon_c(t)$ is measurement noise. The decoder searches for a sparse set $S$ and its timings $\tau$, with $a$ fixed by design. We will run expectation-maximization over target timings, together with an exhaustive search over sets of up to $K=3$ targets. That exhaustive search is affordable for panels of a few dozen targets.

Doubling once per cycle means a tenfold difference in starting amount shifts $\tau$ by about 3.3 cycles. We therefore predict that components at least about 10-fold apart separate by timing alone. Components closer than the curve-width noise must be separated by the code.

The codebook is chosen so that any sum of at most $K$ codewords is unique even when all timings coincide. This is a sum-distinct code over a three-level alphabet. Two refinements adjust it. Codes whose dyes overlap spectrally are penalized, and the plateau competition seen in multiplex amplification, where an abundant target suppresses a rare one, enters as a coupling term that we fit on calibration runs. The open design question is the trade-off. Requiring codes to survive collisions shrinks the codebook below $3^4$, and we will map capacity against the maximum collision order $K$.

Research Program

We will build three things. The first is a calibration set of single-target curves, used to fit $s$, $g_j$ and the competition term. The second is the codebook design procedure. The third is the decoder. Wet-lab work reuses MCPC chemistry, including its Homo-Tag Assisted Non-Dimer primer tags, which prevent primer-dimer formation. We will start with the 15 HPV L1 plasmids, re-encoded on three-level codes, and then extend to the foodborne pathogen panel. Simulations calibrated on real curves will cover larger panels and higher collision orders, but no conclusion will rest on simulation alone.

Baselines

The naive reference point is MCPC's own endpoint threshold decoding (Multicolor Combinatorial Probe Coding for Real-Time PCR). It reads which channels rise and assumes a single target. We will run it on exactly our mixture plates with its published decision rule. For accuracy on single targets it is the strongest existing method. It reaches the full 15-code capacity with a limit of detection of 5–10 CFU per reaction. We found no reported mixed-sample decoding accuracy on Lacuna for MCPC. We will therefore measure its mixture error ourselves, and the bar for success is a large reduction in false-positive calls on the same plates.

The second baseline is endpoint unmixing with our collision-resistant codebook. It solves the same sum decomposition from final fluorescence only, without timing. It shares our codebook, calibration and tuning budget. This ablation isolates the timing ingredient. If it matches the full decoder on mixtures with unequal abundance, the kinetic claim fails.

The third baseline is the full decoder applied to MCPC's original binary codebook. It isolates the codebook ingredient.

The fourth baseline is a split-tube binary panel, in which targets are divided across tubes so that targets likely to co-occur never share a tube. It is run at equal total reagent cost.

Uniplex PCR on every target serves as ground truth and as the upper bound on sensitivity. Beating the baselines means higher exact recovery of the target set at matched sensitivity to single targets, at no more than one tube.

Why not just…?

Why not just split the panel across more tubes? This always works, and it is our fourth baseline. It multiplies sample consumption, cost and handling, and it only protects against co-occurrences that were anticipated, not unexpected co-infections. The comparison will show how many tubes binary splitting needs to match one-tube kinetic decoding at a given collision order. If two tubes suffice for realistic panels, the advantage is modest, and we will say so.

Why not just add more binary colors? On fixed four- to six-channel instruments, more colors means new optics and worse spectral overlap. Bitbow shows that even binary code expansion loses effective capacity to uneven use of codes (Bitbow). Binary codes also do not remove the summing problem. More bits make collisions rarer but cannot resolve them.

Why not just decode the endpoint with a good group-testing code and skip the kinetics? This is the most serious threat. The second baseline tests it directly. At equal abundances we expect no kinetic gain, which is consistent with the shared-program failures in the breast cancer benchmark. If the kinetic gain is small even at 100-fold abundance differences, the contribution reduces to codebook design. In that case we will report it as such.

Why not just partition the sample physically (digital PCR) so targets never meet? Partitioning removes collisions entirely and is a genuine competitor. However, it requires dedicated hardware and the file reports no comparable results. Our claim is restricted to unmodified real-time instruments. We will not claim to beat partitioning.

Proposed Experiments and Decisive Checks

We will run pairwise and three-way mixtures of HPV plasmids across abundance ratios from 1:1 to 1:1000 and across total inputs near the limit of detection. Metrics are exact recovery of the target set, per-target sensitivity and specificity, and error in the estimated log abundance ratio.

The decisive check is the curve of decoding accuracy against abundance ratio, for the full decoder and for endpoint unmixing. The full decoder should pull clearly ahead above about a tenfold ratio, where the timing gap is about 3.3 cycles, and should converge to endpoint unmixing at 1:1.

A second check is clinical. We will genotype compound β-globin mutations in new samples collected under the MCPC protocol, scoring against sequencing-confirmed genotypes. This tests whether dosage levels and coexisting variants decode jointly.

A third check measures error per codeword and per pair. It shows whether, as with Bitbow, some codes fail disproportionately, which would shrink the usable codebook. Finally, we will raise the collision order $K$ in calibrated simulation until exact recovery falls below 0.9, and report that point as the practical capacity of four colors.