Inverse protein folding models read a three-dimensional protein backbone and write out an amino acid sequence expected to fold back into that shape. They are trained on essentially the entire Protein Data Bank (PDB) and on hundreds of millions of AlphaFold DB predictions, and they are graded on static test sets carved from those same databases. This creates two blind spots that this proposal argues are the same blind spot. First, when a model "recovers" the native sequence of a test backbone, we cannot tell whether it solved a physics problem or remembered a sequence–structure pair it saw in training. Second, every deployed screen for hazardous protein sequences matches on sequence similarity or reads text annotations, so a hazard presented purely as geometry — a toxic active site buried inside a freshly designed, innocuous scaffold — sits exactly where today's screens do not look.
We propose to build one instrument that settles both questions: a procedural benchmark generator that produces novel protein backbones with embedded functional motifs, drawn from a curated toxin library, alongside matched benign controls. The same generated backbones serve as a contamination-free exam for inverse folding models (does the celebrated recovery survive structures that cannot have been in any training set?) and as a red-team testbed for hazard screening (does a geometrically camouflaged toxin motif slip past homology and text-based screens while remaining structurally intact?). The work is in silico throughout: we measure failures of evaluation and detection, we release aggregate statistics rather than optimized sequences, and we borrow a cryptographic release protocol so the hazard motifs themselves never become scrapeable training data.
Core Claim
The claim has three parts. First, a substantial share of the sequence recovery reported on static structural benchmarks is memorization of PDB sequence–structure pairs, and this share is measurable by comparing recovery on natural backbones against recovery on procedurally generated backbones matched for length, secondary structure, and contact density. Second, screening systems whose signals are sequence homology and natural-language annotations — including ones with certified miss-rate guarantees — will fail on designs where the hazard is a short structural motif in a novel scaffold, because neither signal exists in that regime; the adversary supplies the annotation, and there is no taxonomy to shift within. Third, the missing detector signal is already latent in the PDB itself: local geometric fingerprints mined from coordinates, such as the principal eigenvector of a contact matrix, which Porto et al. showed analytically predicts site-specific residue preferences, and the structure-conditioned coupling energies Holland extracted from nonredundant PDB subsets. A detector that compares local geometric environments, not sequences or words, should catch geometric camouflage by construction.
No existing work occupies this position. CRC-Screen certifies DNA-synthesis hazard screening under taxonomic shift, but its signals are sequence homology, an LLM panel over UniProt text, and text embeddings — its own authors flag semantic obfuscation of annotations as an open vulnerability, and it never touches 3D coordinates. GeneBreaker demonstrates homology-driven camouflage against DNA language models, achieving up to 60% attack success on Evo2-40B and reconstructing the HIV-1 envelope at 0.334 RMSD to crystal structure, but it attacks genomic models, uses a static 94-sample benchmark, and does no contamination auditing. ProteinInvBench and CHIMERA-Bench standardize structural evaluation and add leakage-motivated splits, but their sets are static snapshots and neither quantifies pre-training memorization nor considers adversarial use. DyVal proves the dynamic-evaluation principle — models that look strong on static sets scored nearly zero on its generated tasks — but only for symbolic reasoning over directed acyclic graphs. We extend that principle to 3D protein structure, where DyVal's authors explicitly say it does not reach.
Research Program
The generator is built on FrameFlow, an SE(3) flow-matching model — a generator that learns to denoise random rigid-body motion into protein backbones — whose motif amortization and motif guidance already condition generation on embedded functional motifs, succeeding on 20 of 24 test motifs and producing 353 unique structural clusters where RFdiffusion produced 141 (Improved motif-scaffolding with SE(3) flow matching, using the released microsoft/protein-frame-flow implementation). We extend its motif library from the 24 benign scaffolding motifs to active-site motifs of protein toxins, extracted directly from PDB coordinates of the manually curated Swiss-Prot KW-0800 toxin set that CRC-Screen evaluated under leave-one-taxonomic-family-out cross-validation. Every generated item comes in matched triples: the natural toxin backbone, a novel scaffold embedding the same motif, and a novel scaffold with no hazardous motif, matched on length, secondary-structure composition, and predicted difficulty.
A control that the generator itself is not leaking memorized backbones uses the instadeepai/protein-structure-tokenizer vector-quantized encoder: if generated backbones predominantly re-emit the same codebook tokens as PDB structures at the global level, novelty claims fail and we retrain or reseed. Foldability of every backbone is verified with ESMFold self-consistent TM-score, the same sc-TM procedure ProteinInvBench uses, so that a recovery drop cannot be blamed on unfolded garbage.
The contamination audit runs the eight-model roster benchmarked by ProteinInvBench on both natural and generated backbones. The headline quantity is the memorization gap:
$$\Delta_{\mathrm{mem}} = \mathrm{AAR}_{\text{natural}} - \mathrm{AAR}_{\text{generated}},$$
where AAR is amino-acid recovery, the fraction of native (or design-target) residues the model reproduces, computed on pairs matched for length, secondary structure, and contact density. We supplement AAR with a signal imported from the contamination literature: Nicolas Yax et al. showed that language models predict answers to canonical reasoning puzzles with near-zero token-level surprise, so we measure per-residue log-probability of the native sequence conditioned on natural versus generated backbones — near-zero surprise on PDB backbones paired with high surprise on matched novel ones would be direct evidence of memorization rather than skill. Following the frequency-bias logic of Snoopy, we further regress per-protein recovery against CATH family frequency in likely training corpora, testing whether recovery tracks exposure rather than geometric difficulty.
The stealth audit treats the screening stack as the system under test. For each camouflaged design we run the inverse folding model to obtain sequences, attach the annotations an adversary would supply — benign, abstract functional descriptions in the style of the semantic rebranding Fred Heiding et al. documented and the attribute-abstraction pipeline of SemanticCamo — and then run the detectors. Hazard retention is measured three ways: recovery of native catalytic residues at the motif positions, motif RMSD after re-folding the designed sequence, and interface-level metrics where relevant. Only aggregate rates are reported.
The defense leg builds the detector our audit predicts is needed: a library of local structural fingerprints of hazardous active sites, indexed by contact-matrix eigenvectors in the spirit of Porto et al. and scored by structure-conditioned coupling energies following Holland, calibrated with Conformal Risk Control exactly as CRC-Screen does so that miss-rate bounds remain certified under shift. Because the test stream is generated on the fly, the detector's calibration is contamination-free too — the adversary cannot pre-train against a published test set.
Baselines
The "do nothing" reference point is the status quo of static-benchmark reporting: AlphaDesign reports 60.5% sequence recovery for its ADesign model in the supervised-learning setting on its AlphaFold DB benchmark, against 52.5% for GVP, and ProteinInvBench reports up to 66% recovery for KWDesign on the multi-chain PDB dataset. If recovery on our matched generated backbones lands within a few points of these numbers, the contamination worry is empirically overstated and the audit says so; a drop commensurate with the existing de novo evidence — recovery falls to roughly 40% on the CASP15 TS45 de novo set (ProteinInvBench) — or larger would confirm it. ProteinInvBench's TS45 result is itself the strongest existing probe of generalization, and we treat it as a baseline: it is the best current estimate of performance on structures absent from training, but it is a static, 45-target snapshot confounded by difficulty; our matched design separates novelty from difficulty, which TS45 cannot.
The strongest competitor on the screening question is CRC-Screen (CRC-Screen), which reports a 0% empirical miss rate across its ten leave-one-family-out folds on Swiss-Prot KW-0800 with a 0% false-flag rate on nine of ten folds, and certifies miss-rate ceilings of roughly 24–49%. We will reimplement its full multi-signal stack — homology score, five-model LLM panel over the supplied annotation text, and embedding distance to toxin cluster centroids, fused and conformally calibrated — and run it on our camouflaged stream with the same hazard set, the same calibration budget, and the same compute as our geometric detector, so the comparison isolates signal type rather than tuning effort. Beating it means a materially lower miss rate on camouflaged designs at a false-flag rate on matched benign generated controls no worse than its own. The naive detector baseline is homology-only $k$-mer screening, which CRC-Screen already showed collapses to a 100% false-flag rate under taxonomic shift; we include it to confirm that collapse extends to the generative regime.
The ablation that isolates our key ingredient runs the identical hazardous motif in three carriers: the native toxin backbone, a novel scaffold with the motif presented plainly, and a novel scaffold with the motif geometrically embedded and residue-conservatively mutated. If screens catch the first two and miss the third, geometric camouflage — not sequence novelty, not motif length — is the causal factor. The detector-side ablation runs the same geometric library with sequence-only and text-only signals ablated, to attribute any recovery in detection to the coordinate signal.
Why not just…?
Why not just use temporal splits on PDB deposition dates? This is the cheapest fix and the one CHIMERA-Bench already implements. It fails for two reasons documented in the file: models that hold up on temporal splits still degrade sharply on held-out epitope-group splits in the same paper, so temporal separation does not control the confounds that matter; and Haoang Chi et al. showed that static temporal benchmarks are scraped into the very next pre-training run, at which point the split protects nothing. A backbone generated after model release cannot have been in any corpus, which is exactly the property DyVal exploited when static-benchmark stars scored nearly zero on its generated tasks (DyVal).
Why not just BLAST the generated sequences against known toxins? The motif residues are public knowledge, so a skeptic expects string matching to catch them. The evidence says otherwise at every layer: CRC-Screen found homology matching so brittle under shift that it flagged 100% of orders to avoid misses, and that adding the homology signal actively degraded the fused screen; GeneBreaker bypassed blacklist matching entirely by priming with non-pathogenic homologs. A short active-site motif in a several-hundred-residue novel scaffold is the worst case for $k$-mer screening, and our Experiment 2 measures exactly how much signal survives. We are honest that if homology screening does catch the camouflaged designs, the stealth half of the claim is dead and the audit half stands on its own.
Why not just extend CRC-Screen with a structure branch and skip the benchmark? This is the closest alternative and a genuine threat — our defense leg is essentially that extension. But without a generated test stream there is nothing valid to evaluate it on: CRC-Screen's leave-one-family-out protocol is a static split of natural sequences with real taxonomies and honest annotations, none of which exist in the adversarial regime, where the annotation is attacker-written and the scaffold has no family. LLM-panel signals are also known-fragile here: SemanticCamo achieved an 80.35% average attack success rate against six models including over 84% on GPT-4o, and GAVEL measured GPT-4's detection recall collapsing to as low as 12% under misdirection. The benchmark is the precondition for knowing whether any detector, ours or a competitor's, actually works; if someone builds a better detector on our released benchmark, that is the intended outcome.
Why not just run perplexity or anomaly detection over the outputs? Detecting Fluent Adversarial Prompts showed that static and windowed perplexity filters located the true adversarial suffix only 17–46% of the time against fluent attacks, and inverse-folding outputs are fluent protein language by construction — there is no gibberish to detect. Anomaly detection has no clean reference class for a de novo scaffold. We include a perplexity-style baseline over generated sequences in Experiment 2 and expect it to fail; if it does not, that result would itself revise the threat model.
Decisive Checks
Experiment 1 quantifies $\Delta_{\mathrm{mem}}$ for each of the eight ProteinInvBench models on matched natural-versus-generated backbones, with token-level surprise and frequency regressions as corroborating signals. The decisive outcome is a gap as large as the TS45 drop (roughly 40% recovery, per ProteinInvBench) on backbones verified foldable and matched for difficulty; no reported result found on Lacuna for any contamination audit of inverse folding models, so any quantified gap is new knowledge. Experiment 2 reports miss rates for the homology baseline, the replicated CRC-Screen stack, and an LLM-judge-only condition on the camouflaged stream, against retention metrics proving the designs are still hazardous; the pre-registered prediction, grounded in CRC-Screen's own text-vulnerability note and GeneBreaker's homolog camouflage, is that the fused text-and-sequence screen's certified guarantees do not transfer to this regime. Experiment 3 evaluates the geometric detector on held-out generated batches with conformal calibration, comparing its certified miss bound to CRC-Screen's reported 24–49% ceiling and its empirical miss rate to that screen's 0%; success is a bound that holds under camouflage where the incumbent's signals do not exist. Experiment 4 stress-tests both legs against adaptation, mutating motif residues while preserving geometry — the protein analogue of the rephrasing that On The Fragility of Benchmark Contamination Detection in Reasoning Models and Rethinking Benchmark and Contamination for Language Models with Rephrased Samples showed defeats text-based memorization detection — and asks whether the geometric signal degrades gracefully where sequence identity decays. Benchmark release follows the zero-knowledge protocol of $\text{P}^4$Bench so detectors can be evaluated against the hazard library without the hazard library becoming scrapeable training data — closing, for the benchmark we create, the very leak we set out to measure.
References
- Prediction of Site-Specific Amino Acid Distributions and Limits of Divergent Evolutionary Changes in Protein Sequences
- CRC-Screen: Certified DNA-Synthesis Hazard Screening Under Taxonomic Shift
- GeneBreaker: Jailbreak Attacks against DNA Language Models with Pathogenicity Guidance
- ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Tasks, Models, and Metrics
- CHIMERA-Bench: A Benchmark Dataset for Epitope-Specific Antibody Design
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks
- Improved motif-scaffolding with SE(3) flow matching
- Studying and improving reasoning in humans and machines
- Snoopy: An Online Interface for Exploring the Effect of Pretraining Term Frequencies on Few-Shot LM Performance
- SemanticCamo: Jailbreaking Large Language Models through Semantic Camouflage
- AlphaDesign: A graph protein design method and benchmark on AlphaFold DB
- Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?
- GAVEL: Towards Rule-Based Safety through Activation Monitoring
- Detecting Fluent Optimization-Based Adversarial Prompts via Sequential Entropy Changes
- On The Fragility of Benchmark Contamination Detection in Reasoning Models
- Rethinking Benchmark and Contamination for Language Models with Rephrased Samples
- $\text{P}^4$Bench: Contamination-Proof, Publicly Verifiable, and Privacy-Preserving LLM Evaluation via Zero-Knowledge Proofs