Recombination Junctions as Ownership Keys for Organellar RNA-Editing Predictors

Plant mitochondrial genomes are not one fixed circle. Repeated stretches of DNA recombine, so the same genome exists as several coexisting arrangements: separate circles, flipped segments, and linear pieces. Each arrangement creates new "junction" sequences where two normally distant pieces of DNA sit side by side. Separately, plants repair broken mitochondrial and chloroplast genes after transcription by C-to-U RNA editing, a chemical change that turns specific cytidines in the messenger RNA into uridines. Neural networks that predict which cytidines get edited, such as Deepred-mt and DeepRed, are now routine annotation tools. None of them carries any mechanism to prove who trained a given model if a competitor copies it through its prediction interface.

We propose to use recombination junctions as a secret ownership key for such predictors. The owner builds a hidden set of junction-spanning sequence windows taken from real alternative genome arrangements. During training, the owner teaches the model a private pattern of edit and no-edit calls on those windows. A model stolen by distillation, meaning retraining a copy on the original's outputs, should inherit that pattern, while independently trained models should not. Junction windows suit this job for two reasons. They are genuine mitochondrial sequence, so an attacker cannot filter them out as odd-looking queries. They also have no measured editing labels and a genuinely ambiguous context, so the owner can choose their labels without hurting accuracy on real editing sites. The claim is testable: junction keys should match synthetic-trigger watermarks on detection while being far harder to filter, and should beat randomly chosen natural windows on both how distinctive they are and how little accuracy they cost.

Core Claim

The IP survey Deep Intellectual Property: A Survey frames model protection as a trilemma. Fidelity means no loss in task accuracy. Robustness means the mark survives pruning, fine-tuning, and laundering. Capacity means how much information the mark carries. In vision, trigger-based watermarks pay for capacity with fidelity, or with triggers that look unnatural. We argue that organellar genomes come with an unusually cheap key space: the set of junction windows produced by repeat-mediated recombination. This space is large, because every repeat pair yields junctions. It is natural, because these sequences physically exist in the cell. And it is unlabeled in current editing datasets, because cDNA evidence is collected against one consensus arrangement.

Our hypothesis is that on these windows, independently trained predictors disagree with each other, since the windows are underdetermined by training data. A distilled copy, by contrast, agrees with its source. That combination is what a black-box fingerprint needs.

Prior Work and Positioning

The biology comes from recent assemblies that document coexisting arrangements and editing side by side, but only as observations. Complete mitochondrial genome of Melia azedarach L. reports a 640,529 bp genome on two circles whose direct repeats R1 and R2 interconvert two conformations, confirmed by PCR across the junctions. It also predicts 356 editing sites. Abies beshanzuensis splits 1.06 Mb across ten molecules and predicts 827 sites with Deepred-mt. Aglaia odorata and Pontederia crassipes chart repeat junctions and isoform frequencies, the latter using high-accuracy PacBio HiFi long reads. Scaevola describes two-circle arrangements. None of these papers asks how editing predictors behave on junction contexts; we supply that use.

Identifying repeat domains in large genomes decomposes repeats into shared building blocks using a graph method. We reuse it to enumerate candidate recombination partners beyond simple exact repeats.

The predictors themselves define the model being protected. DeepRed classifies editing from a 201-bp window around each candidate site. ATTIC stacks 14 learners over 37 sequence encodings. Both assume one static reference context, which is exactly why junction windows are ambiguous for them.

The protection literature constrains the design. Digital Passport locks a model with secret weight-derived parameters, but it needs white-box access. Effective Ambiguity Attack breaks passports by swapping in small replacement blocks trained on 2% of the data. This negative result is why we choose black-box verification. Deep Serial Number ties accuracy to a secret input pattern through distillation, and Deep Model IP Protection via Deep Watermarking makes copied models reproduce a hidden output signal. Both supply training techniques we adapt, but both use artificial keys and were tested only on images. The asymmetric schemes Duhamel et al., UPV, and Publicly-Detectable Watermarking warn that revealing a key to verifiers enables forgery. We treat that warning as a limit we inherit, not one we solve.

Research Program

The first step is to build the key space. For each assembly listed above, we enumerate direct and inverted repeat pairs, including repeat-domain mosaics, and construct every alternative junction. We then extract 201-bp windows centered on cytidines within roughly 100 bp of a junction, matching DeepRed's input format. Each window is tagged by species, repeat type, and whether the cytidine lies in a coding region in the alternative arrangement.

The second step is to measure natural disagreement. We train $M$ independent predictors with different seeds, architectures (convolutional, DeepRed-style, ATTIC-style), and data subsets. We then estimate $q_0$, the rate at which two independent models agree with a fixed random label assignment on junction windows. This rate is the null that verification must beat.

The third step is embedding. The owner draws a secret key set $K$ of $n$ junction windows with random target labels. The owner trains with the ordinary editing loss plus a weighted loss on $K$, using the distillation-entangling recipe adapted from Deep Serial Number so the key behavior transfers to students.

Verification counts $a$, the number of key windows on which a suspect model's calls match the targets. Ownership is asserted when the binomial tail under the null is below $\alpha$:

$$ p = \sum_{j=a}^{n} \binom{n}{j} q_0^{\,j} (1-q_0)^{n-j} < \alpha , $$

Here $q_0$ is the empirical independent-model agreement rate and $\alpha$ is the false-accusation budget. We use a binomial because the key windows are drawn independently. We check that assumption by recomputing $q_0$ on held-out windows from the same repeat pairs.

The training labels come from organellar gDNA-versus-cDNA comparisons, which mark sites where the genome reads C and the RNA reads U. Total RNA-seq is preferred for these, following Glycyrrhiza uralensis, which establishes 38 high-confidence plastid sites. The Melia and Abies site counts are predictions rather than measurements, so we use them only as weak labels, and we report results separately for measured and predicted labels.

Baselines

The naive reference point is an unprotected predictor queried on $K$. Its agreement rate with the owner's targets defines the false-positive floor and the accuracy we must preserve. For accuracy we report AUC on held-out measured editing sites. For context, DeepRed reports 97.9% test AUC on human A-to-I editing. No reported organellar C-to-U accuracy for Deepred-mt was found on Lacuna, so we will establish our own unprotected baseline.

The strongest existing method is a Deep Serial Number–style key watermark. Its key is a synthetic motif inserted into windows, embedded with the same distillation recipe. It reports 97.2% accuracy with the key versus 8.2% without on GTSRB, and it survives 20% pruning and fine-tuning on 10% of the data. A second baseline is a barrier-network output watermark following Deep Model IP Protection via Deep Watermarking, which reports near-100% extraction from copied models in imaging. Passports are run only as a white-box reference, together with the replacement-block attack from the ambiguity paper.

Every method gets the same training data, architecture, and key-set size, and the same tuning budget of 20 configurations on a validation split. All are attacked with the same suite: distillation through the prediction interface, fine-tuning on 2% and 10% of the data, 20% pruning, and replacement-block substitution.

The key ablation replaces junction windows with randomly sampled natural non-junction windows from the same genomes. This isolates whether the recombination ingredient matters.

Beating the strongest baseline means three things together. Detection after attack must be at least as high. The AUC cost on real sites must be equal or lower. The attacker's ability to flag key queries must be substantially lower.

Why not just…?

Why not just insert a synthetic trigger motif, as in standard backdoor watermarks? It may detect copying just as well, and we expect it to on clean queries. Its weakness is that a motif absent from real genomes is statistically conspicuous. We will train a k-mer likelihood filter on the training genomes and measure how well it flags keys, which estimates the rate at which a thief could refuse or randomize those queries. If motif keys are no easier to flag than junction keys, our advantage disappears and we will report that.

Why not just use random natural windows as the fingerprint? This is the real threat. It is exactly the ablation above. Our reason to expect junctions to win is that ordinary non-coding windows are confidently predicted "no edit" by every model, so $q_0$ is high and overriding those calls costs accuracy. Junction windows near coding cytidines should be the ones where models actually disagree. If the ablation matches junctions on both $q_0$ and AUC cost, the junction ingredient is unnecessary. In that case the contribution narrows to the junction benchmark of predictor instability.

Why not just use passports or license terms? Passports need weight access and are forgeable with 2% of the data, per the ambiguity-attack paper. License terms and weight hashes say nothing about a distilled copy whose weights differ entirely.

Proposed Experiments and Decisive Checks

The first decisive check is whether independent models actually disagree on junction windows, measured as $q_0$ well below agreement on random natural windows. If they do not, the method fails at step two. The second is whether students distilled through the prediction interface retain key agreement high enough to give $p<10^{-6}$ with a few hundred keys, while the AUC drop on measured sites stays under one point. The third is how hard keys are to filter, measured as the k-mer filter's AUC for junction keys versus motif keys.

Two further checks probe transfer and biology. For transfer, we use Repeat Domain Graph mosaics from human Repbase as keys for an ATTIC-style A-to-I predictor on H_3000 and M_831, where ATTIC reports accuracy of 0.9173 and 0.9946. For biology, wherever long-read cDNA spanning a Melia or Pontederia junction is available, we check whether editing on the alternative arrangement is measurable. A positive answer would mean some keys have ground truth, and we would exclude them.

We state one limitation plainly. Revealing $K$ to a court or auditor spends it, as the asymmetric-watermarking literature warns. We therefore rotate disjoint key sets drawn from the large junction space instead of claiming public verifiability.

References

  1. Deep Intellectual Property: A Survey
  2. Complete mitochondrial genome of Melia azedarach L., reveals two conformations generated by the repeat sequence mediated recombination
  3. Assembly and comparative analysis of the complete mitochondrial genome of Abies beshanzuensis: insights into conservation genomics of a critically endangered fir
  4. The complete mitochondrial genome of Aglaia odorata, insights into its genomic structure and RNA editing sites
  5. The complete mitochondrial genome of Pontederia crassipes: using HiFi reads to investigate genome recombination and gene transfer from chloroplast genome
  6. Organelle genomes of two Scaevola species, S. taccada and S. hainanensis, provide new insights into evolutionary divergence between Scaevola and its related species
  7. Identifying repeat domains in large genomes
  8. Accurate identification of RNA editing sites from primitive sequence with deep neural networks
  9. ATTIC is an integrated approach for predicting A-to-I RNA editing sites in three species
  10. Digital Passport: A Novel Technological Strategy for Intellectual Property Protection of Convolutional Neural Networks
  11. Effective Ambiguity Attack Against Passport-based DNN Intellectual Property Protection Schemes through Fully Connected Layer Substitution
  12. Deep Serial Number: Computational Watermarking for DNN Intellectual Property Protection
  13. Deep Model Intellectual Property Protection via Deep Watermarking
  14. Introduction aux méthodes de tatouage asymétriques dans le cadre de la protection de copie
  15. An Unforgeable Publicly Verifiable Watermark for Large Language Models
  16. Publicly-Detectable Watermarking for Language Models
  17. Plastid RNA Editing in Glycyrrhiza uralensis: Landscape Characterization and Comparative Assessment of RNA-Seq Library Strategies for Detection