Many ultrasensitive protein assays attach a short DNA tag, a barcode, to each antibody. The assay then counts proteins by sequencing and tallying the barcodes. At the rare end of the proteome this counting step breaks in a specific way. A rare biomarker may produce fifty barcode reads while an abundant background protein produces five million. If the sequencer mistakes even one abundant-protein read in ten thousand for the rare barcode, the resulting false counts swamp the true signal. So on a nanopore sequencer, a device that reads DNA by the ionic current it blocks as it threads through a protein pore, the limit of detection is set less by chemistry than by barcode cross-talk. The cross-talk that matters is directional: errors flowing from abundant barcodes into rare ones.
DNA data storage has spent years on a closely related problem, and this proposal borrows its central lesson. Codes work better when they are designed against the physics of the reader rather than against an abstract edit distance. We propose to design barcode sets whose members are separated in expected ionic-current space and not only in letter space. We also exclude sequences that fold into hairpins under high-salt conditions. Finally, we read counts through a decoder that knows the full confusion matrix between barcodes and can refuse to call a read. The claim is concrete: at equal barcode length, this combination lowers the abundance ratio at which a rare target can be detected by at least an order of magnitude, compared with edit-distance-designed barcodes demultiplexed by a neural classifier.
Core Claim
In barcode counting, the quantity that matters is the expected number of false counts landing in the bin of a rare barcode $i$. That number is
$$F_i = \sum_{j \neq i} N_j \, P(i \mid j),$$
where $N_j$ is the true read count of barcode $j$ and $P(i \mid j)$ is the probability that a read from barcode $j$ is decoded as $i$. Detection of barcode $i$ requires $F_i$ well below $N_i$. A panel spanning a dynamic range of $R = N_j/N_i$ therefore needs pairwise confusions below roughly $1/R$ for every abundant-to-rare pair. Average demultiplexing accuracy says little about this worst-case, abundance-weighted quantity.
Existing barcode design and demultiplexing tools also do not optimize it. TDFPS-Designer picks barcode sets that maximize edit distance. DemuxTrans classifies barcodes directly from raw current using a Transformer with temporal convolutions. Scalable design of orthogonal DNA barcode libraries scales orthogonal library construction. All three treat demultiplexing as balanced classification. We instead treat the barcode set as a code over the physical nanopore channel, and we score it by the abundance-weighted confusion $F_i$.
Where the Idea Comes From
The method is supplied by Constrained coding for error mitigation in nanopore-based DNA data storage (companion record here). That work builds a graph over all 6-letter DNA windows. It removes any transition where the expected current changes by less than a threshold $\delta$, limits runs of identical letters, and decodes with a Viterbi decoder over a hidden Markov model of the pore. At $\delta = 9$ pA it reports up to a 6-fold reduction in sequence-level errors relative to unconstrained sequences basecalled by Guppy or Dorado.
That work concerns storing files, where every strand matters equally and outer error-correcting codes absorb residual errors. Barcode counting is different in three ways. It is a small codebook problem, so we can optimize whole-set separation directly instead of constraining local transitions. The cost of errors is asymmetric by abundance. And the decoder may abstain, because a discarded read costs little when abundant targets have millions of reads.
Two lines of evidence motivate a folding constraint. Deep statistical modelling of nanopore sequencing translocation times reveals latent non-B DNA structures shows that secondary structure in DNA distorts translocation times through the pore. The polyelectrolyte literature adds that added salt screens backbone charge and lets chains collapse and fold. Wenlong Zheng et al. show this for polyacrylamide in brine, and the same screening physics applies to the DNA phosphate backbone. Protein assays often carry high-salt lysis or binding buffer into library preparation. In those samples, barcodes that fold will stall and read badly, and the stretch of current around each barcode will not match a fixed lookup table.
We therefore use salinity in two roles. It is a design constraint: barcodes must not fold at the worst-case carryover salt concentration. It is also a stress test: current-level and dwell-time perturbations stand in for changes in ionic strength.
The other two directions play smaller roles. Physical preconcentration, such as the paper-based ion concentration polarization device of Folding-paper-based preconcentrator, concentrates charged proteins by up to 300-fold. It raises all proteins together and does not change the abundance ratio $R$ that drives cross-talk. Contact angle measurement, meaning measuring how a droplet beads or spreads on a surface as a readout of surface energy, enters only as a wet-lab control. Rare barcoded complexes can be lost by adsorbing to hydrophobic plastic. We will measure water contact angle on the tubes and chip surfaces used, with a sessile-drop profile fit via OpenDrop, and run matched low-binding versus standard surfaces. This lets us separate molecule loss from decoding error.
Research Program
The first component is set design. Starting from ONT's published 6-mer current model, we compute the expected current trace of every candidate barcode plus its flanking adapter. We define the separation between two barcodes as the minimum, over plausible time-warpings, of the distance between their expected traces. Time-warping accounts for motor-protein slips and stalls, the dwell-time jitter documented for nanopore readers. We then select a set of $M$ barcodes, with $M$ between 96 and 384 and length $L$ fixed, to maximize the minimum pairwise separation. Each barcode must also satisfy three hard constraints: the run-length limit, GC balance, and a predicted folding free energy above a cutoff evaluated at the highest expected salt carryover. For a fixed $L$, set selection is a max-min dispersion problem. We solve it greedily with local swaps, because exact solution is unnecessary at these sizes.
The second component is decoding with abstention. For each read we compute a posterior over the $M$ barcodes, using the pore's hidden Markov model as in the constrained-coding work. We can also use a neural classifier's calibrated scores, so the decoder is interchangeable. A read is called only if its posterior margin exceeds a threshold $\tau$. We then estimate true counts by fitting observed counts $m = C n$ with a Poisson likelihood. Here $n$ is the vector of true counts and $C$ is the confusion matrix, including the abstain row, measured from single-barcode calibration runs. Because $C$ is explicit, the known leakage from abundant barcodes can be subtracted, not merely suppressed.
The third component is robustness. We will perturb the current model with per-run scale and offset shifts and with widened dwell-time distributions. We then check whether the barcodes that confuse under perturbation are the ones our separation score predicts.
Baselines
The naive reference point is ordinary practice: random or vendor-supplied barcodes of length $L$, standard basecalling, and nearest-match assignment with an edit-distance cutoff. Every other result is reported as an improvement over it in abundance-weighted false counts $F_i$.
The strongest design competitor is an edit-distance set produced by TDFPS-Designer at the same length and set size. It is the most established way to make barcodes robust to nanopore indels. It will be decoded with the same decoders as our set, so the comparison isolates design. No reported result was found on Lacuna for TDFPS-Designer's demultiplexing accuracy or cross-talk.
The strongest decoder competitor is DemuxTrans, a raw-signal neural demultiplexer. It is the best available way to extract barcode identity from current without basecalling. We will train it on the same calibration reads for both barcode sets, with the same compute and hyperparameter budget, and give it the same abstention threshold sweep. No reported result was found on Lacuna for DemuxTrans's accuracy.
For the physics-coded approach, the reference number is the 6-fold sequence-error reduction at $\delta = 9$ pA over Guppy/Dorado reported by Whritenour et al.. We include their local-transition constraint, used without whole-set optimization, as a barcode set. This tests whether set-level separation adds anything beyond it.
Three ablations each remove one ingredient: the folding constraint, abstention, and confusion-matrix deconvolution. Beating the baselines means that at equal $L$, $M$ and read yield, our pipeline reaches a 10-fold lower minimum detectable abundance ratio at a false-detection rate of 1%. Anything under 3-fold we would report as a negative result for set design.
Why not just…?
Why not just use edit-distance barcodes and make them longer? This is the real threat. Longer barcodes raise separation under any metric, and edit distance correlates with current distance. Our comparison is therefore at fixed length, with length also swept. If edit-distance sets match current-optimized sets at every length, the design contribution is null and only the decoder parts survive.
Our reason to expect a gap is that the constrained-coding result shows letter-level structure alone leaves physically indistinguishable transitions. Specifically, current plateaus where consecutive windows give nearly the same signal. Those plateaus are where the 6-fold gain came from.
Why not just use a better neural demultiplexer on standard barcodes? A decoder cannot recover information the channel never carried. If two barcodes produce nearly identical expected traces, any classifier confuses them at a rate set by noise. The constrained-coding study found physics-aware sequence design beat unconstrained sequences read by strong neural basecallers. Our factorial design, crossing both barcode sets with both decoders, will show directly whether the decoder or the codebook drives the tail confusions.
Why not just discard low-quality reads? Thresholding is part of our pipeline, but alone it only lowers the leakage rate. It cannot remove the systematic leakage that remains among confidently called reads. Abundant barcodes supply so many reads that even confident errors dominate a rare bin. The abstention-only ablation measures exactly this.
Why not just preconcentrate the sample so the target is not rare? Preconcentration of up to 300-fold, as reported for the paper-based preconcentrator, raises absolute amounts. It does not change the ratio between rare and abundant proteins, and that ratio is what $F_i$ depends on. It is complementary rather than competing.
Proposed Experiments
In simulation, we will use DeepSimulator with read accuracy swept across its reported 83% to 97% range, default about 88.5%. We will generate mixtures in which one barcode sits at abundance ratios from $10^{-2}$ to $10^{-6}$ against the rest, and measure $F_i$, count error and minimum detectable ratio for every design and decoder pairing. We will repeat under perturbed current tables and widened dwell times to test the salinity robustness claim.
In the wet lab, we will synthesize the competing barcode panels and sequence single-barcode calibration pools to measure the confusion matrix $C$. We will then run defined dilution series on an Oxford Nanopore instrument. Salt carryover is titrated across each series, and the folding-unconstrained ablation is included to test whether predicted hairpins cause the stalls the translocation-time literature predicts. Standard and low-binding surfaces, characterized by contact angle, are run in parallel to rule out adsorption loss.
The final check attaches the best-performing panel to antibodies against a recombinant protein spiked into plasma across the same abundance ratios. We will measure whether the gain in minimum detectable ratio measured on DNA alone carries over to protein counts.
The decisive outcome has two parts. Current-optimized, folding-screened barcodes must beat equal-length edit-distance barcodes under both decoders. And confusion-matrix deconvolution must further lower false detections at the $10^{-5}$ ratio. If either fails, the note's central claim fails with it.
References
- TDFPS-Designer: an efficient toolkit for barcode design and selection in nanopore sequencing
- DemuxTrans: Transformer and temporal convolution network for accurate barcode demultiplexing in nanopore sequencing
- Scalable design of orthogonal DNA barcode libraries
- Constrained coding for error mitigation in nanopore-based DNA data storage
- Deep statistical modelling of nanopore sequencing translocation times reveals latent non-B DNA structures
- Impact of polymer addition, electrolyte, clay and antioxidant on rheological properties of polymer fluid at high temperature and high pressure
- Folding-paper-based preconcentrator for low dispersion of preconcentration plug
- OpenDrop: Open-source software for pendant drop tensiometry & contact angle measurements
- DeepSimulator: a deep simulator for Nanopore sequencing