Release Signatures: Mining the Cryptome by What Excision Changes

The human proteome hides a defensive arsenal in plain sight. Short antimicrobial sequences sit embedded inside abundant proteins with entirely unrelated jobs — hemoglobin, fibrinogen, digestive enzymes, collagen — inert while sequestered in the parent fold, but potent once infection-site proteases cut them free. This "cryptome" is real and therapeutically interesting: fragments of hemoglobin and fibrinogen kill bacteria and viruses, and machine-learning pipelines that digitally digest entire proteomes have already uncovered active encrypted peptides in modern and even extinct organisms. But every current discovery engine — the panCleave cleavage model and the APEX activity predictor, and even structure-aware classifiers like the ESCAPE dual-branch transformer — screens each excised fragment as an isolated one-dimensional string. None of them models the single event that defines an encrypted peptide: the transition from being buried in a folded parent to standing alone.

This proposal uses representational similarity analysis (RSA) to measure exactly that transition. RSA is a technique from computational neuroscience for comparing systems that cannot be compared coordinate-by-coordinate: instead of asking where a stimulus sits in a space, it asks how a system orders the pairwise dissimilarities among a shared set of stimuli, and compares those orderings. Our core move is to treat "fragment embedded in its parent" and "fragment excised" as two systems sharing one stimulus, and to quantify how the fragment's relational position in a protein language model's layers changes across the excision event. The hypothesis is that genuinely encrypted, bioactive fragments carry a release signature: because their sequence chemistry is unchanged but their structural context vanishes, their representations should reorganize strongly in the structure-sensitive late layers of folding models, and their excised state should align with the relational geometry defined by known antimicrobials' activity spectra — while the embedded state should not. That signature gives a label-free, structure-aware ranking signal for cryptome mining.

Core Claim

We claim two things, one diagnostic and one practical. Diagnostically: in protein models such as ESMFold, the representation of an encrypted peptide undergoes a layer-structured shift upon excision — small in early blocks (which resolve 1D residue chemistry, per Mechanisms of AI Protein Folding in ESMFold, where charge is linearly encoded in blocks 0–7) and large in late blocks (which accumulate residue–residue spatial geometry, from roughly block 25 onward) — and the direction of that shift moves the fragment toward the region occupied by known antimicrobial peptides. Practically: this release signature, combined with RSA alignment between model-layer geometry and measured activity spectra, ranks candidate cryptic fragments well enough to compete with supervised sequence classifiers — without requiring the labeled cryptome data that those classifiers need and do not have.

No existing work does this. Lu et al. probe the internal two-stage mechanism of folding models on full-length proteins and PDB hairpins, but never touch the cryptome, proteolysis, or the embedded-versus-excised comparison that is our entire object of study ( Mechanisms of AI Protein Folding in ESMFold). APEX and panCleave discover encrypted peptides but operate purely on sequence text (Deep-learning-enabled antibiotic discovery through molecular de-extinction, Molecular de-extinction of ancient antimicrobial peptides enabled by machine learning). ESCAPE adds predicted 3D distance matrices for standalone peptides but builds no cross-system dissimilarity matrices and ignores parent-context entirely (A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification). PepCompass corrects geometric distortion in generative peptide decoders with Riemannian manifolds, but for standalone designed peptides, not for fragments inside parents, and with differential geometry rather than relational comparison (PepCompass: Navigating Peptide Embedding Spaces Using Riemannian Geometry). We reuse RSA's standard machinery — anchored dissimilarity matrices and rank correlation, as implemented in the established toolboxes (A Toolbox for Representational Similarity Analysis, A Python Toolbox for Representational Similarity Analysis, ReSi: A Comprehensive Benchmark for Representational Similarity Measures) — and apply it to a question none of these papers pose.

Research Program

The program has three components. First, we build a two-state corpus: for every candidate fragment produced by in silico digestion (panCleave predictions and overlapping sliding windows, following Melo) of a set of parent proteomes, we extract layer-wise residue embeddings twice — once with the fragment's residues pooled from within the full parent sequence, and once from the fragment alone — using frozen folding models of the kind studied by Lu et al. Second, we define the release signature against a fixed anchor panel of reference peptides (known AMPs, matched non-AMP fragments, and scrambled controls). For a fragment $f$ excised from parent $p$, with layer-$L$ distances $d_L(\cdot,\cdot)$ to each anchor $a_k$, the signature is

$$s(f, L) = 1 - \mathrm{Spearman}\left(\big(d_L(f \mid p, a_k)\big)_{k},\ \big(d_L(f, a_k)\big)_{k}\right),$$

so $s$ is large exactly when the fragment's pattern of distances to the anchors — its relational position, not its coordinates — reorganizes upon release. Third, we define activity-space alignment: for peptides with measured activity spectra $y_i$ (e.g., MIC values across the 34 bacterial strains in APEX's training suite, or ESCAPE's multilabel hierarchy), we compare the activity dissimilarity matrix $\mathbf{D}^{\mathrm{act}}_{ij} = 1 - \mathrm{corr}(y_i, y_j)$ against the layer geometry $\mathbf{D}^{L}_{ij}$ via

$$\rho_L = \mathrm{Spearman}\left(\mathrm{vech}(\mathbf{D}^{\mathrm{act}}),\ \mathrm{vech}(\mathbf{D}^{L})\right),$$

where $\mathrm{vech}$ takes the matrix upper triangle. A candidate fragment's discovery score combines a large $s(f, L)$ at late layers with a high $\rho_L$ for its excised state relative to its embedded state — that is, release must both reorganize the representation and move it toward the activity geometry. This mirrors how RSA revealed that ungrounded language models implicitly reconstruct human color space from text alone (Can Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color): we ask whether folding models implicitly encode the functional organization of the cryptome, and whether proteolytic release is what exposes it.

Two known hazards shape the design. RSA similarity can arise from trivial stimulus statistics rather than genuine alignment (Representational structure or task structure? Bias in neural representational similarity analysis and a Bayesian method for reducing bias, Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models), so all anchor panels are matched for length, net charge, and hydrophobicity, and every alignment is tested against permutations and matched controls. And high representational similarity does not imply shared mechanism (Obstacles to inferring mechanistic similarity using Representational Similarity Analysis) — a live concern here, since Wan et al. showed archaic encrypted peptides kill by depolarizing the inner membrane while canonical antimicrobials disrupt the outer membrane ( Deep-learning-enabled antibiotic discovery through molecular de-extinction). We therefore make no mechanistic-equivalence claim and include an explicit mechanism-stratified evaluation.

Baselines

The naive reference point is classical physicochemical screening: sliding-window digestion followed by threshold filters on charge and amphipathicity, the approach that Wan et al. report achieves a 24% in-vitro hit rate for active peptides. It matters because it is what the field used before machine learning, and because archaic encrypted peptides have atypical chemistry — less amphiphilic, more uncharged polar residues — that these filters actively reject, so it also demonstrates the failure mode we must avoid. The strongest existing competitor is APEX itself, the GRU-attention ensemble trained on 988 peptides with over 14,000 experimental MIC values, which achieved a 59% hit rate on its own prospective validation (Deep-learning-enabled antibiotic discovery through molecular de-extinction). We will run APEX on exactly our candidate pools — the same digests, the same fragment lists — with a matched hyperparameter tuning budget on its own training suite, and give it the same number of wet-lab validation slots. Beating APEX means a higher prospective hit rate at equal validation budget with non-overlapping confidence intervals, or, more realistically and still valuable, comparable hit rate plus a ranked-enrichment advantage on fragments APEX scores poorly. For the in-silico multilabel question, the baseline is the ESCAPE dual-branch transformer, which combines sequence and predicted-fold distance matrices and reports a 2.56% relative improvement in mean Average Precision and a 1.90% F1 improvement over the next best method, including a 35.7% relative increase in Average Precision on the rare antiparasitic class (A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification); we evaluate on its held-out splits with no additional training, so the comparison isolates what frozen-embedding relational alignment can do without supervised labels. Finally, the key ablation replaces RSA with raw Euclidean geometry: an otherwise identical pipeline that ranks fragments by Euclidean displacement and Euclidean distance to the AMP centroid at the same layers. PepCompass showed that treating peptide latent spaces as flat Euclidean planes drives search into inactive "dead zones" (PepCompass: Navigating Peptide Embedding Spaces Using Riemannian Geometry), so if relational ranking does not beat this ablation, the RSA ingredient — not the two-state idea — is what fails, and we would report that.

Why not just…?

Why not just fine-tune a supervised classifier on sequence labels? This is the most serious alternative, and we test it head-to-head rather than argue around it. Supervised models already reach a 59% hit rate where physicochemical scoring reaches 24% ( Deep-learning-enabled antibiotic discovery through molecular de-extinction), and if APEX matched on our candidate pools beats release-signature ranking outright, the practical claim dies and the paper becomes a diagnostic study. But classifiers have two structural gaps this proposal targets: they see only the excised string, so they cannot use the embedded-versus-excised comparison that carries all the structural information about sequestration; and the cryptome has no experimentally validated negative set — a bottleneck named explicitly in AI-Driven Antimicrobial Peptide Discovery: Mining and Generation — which a label-free geometric signal sidesteps. We also test the ensemble: if RSA ranking mainly improves APEX's misses, that is the honest, useful result.

Why not just linear-probe the frozen embeddings? Probing is cheaper and more interpretable, and Lu et al. showed probes cleanly track single features like residue charge in early blocks ( Mechanisms of AI Protein Folding in ESMFold). But a probe answers "is feature X present," never "does this layer's relational organization match that other system's" — and our second system is a matrix of activity spectra across dozens of pathogens, which has no shared coordinate system with any embedding. Yousefi et al. found the same division of labor in language models: probes tracked task parameters while RSA was needed to capture the relational restructuring of hidden layers (In-Context Learning in Large Language Models: A Neuroscience-inspired Analysis of Representations). We still include a probing ablation — if a charge-plus-hydrophobicity probe ranks fragments as well as release signatures, the relational machinery is unjustified and we will say so.

Why not just apply classical AMP rules after in silico digestion? Because the rules are exactly what the cryptome breaks. Beyond the 24% hit rate, the peptides this paradigm actually discovers have signatures the filters reject — Maasch et al.'s archaic encrypted peptides are less amphiphilic than modern AMPs, and panCleave had to outperform protease-specific consensus classifiers to even find the cleavage sites (Molecular de-extinction of ancient antimicrobial peptides enabled by machine learning). We include the rule-based pipeline as the naive baseline so the comparison is quantitative, not rhetorical.

Why not just use sequence similarity (BLAST, BLOSUM) to known AMPs? Encrypted peptides are evolutionarily sequestered inside proteins with unrelated functions — hemoglobin, fibrinogen, MMP-19 — and typically share minimal homology with canonical antimicrobial families (Peptides from non-immune proteins target infections through antimicrobial and immunomodulatory properties, Discovery of Encrypted Peptides in a Human Matrix Metallopeptidase), while function in a space of over $10^{32}$ possible 25-mers does not scale with sequence distance (Exploring and expanding the chemical multiverse of peptides). Homology search is therefore not a competitor but a control: we verify our ranked hits are not covert homologs of training AMPs.

Proposed Experiments

The first experiment is the mechanism check on known positives: the hemoglobin alpha C-terminal 32-mer (The C-terminal 32-mer fragment of hemoglobin alpha is an amyloidogenic peptide with antimicrobial properties), fibrinogen cryptic fragments (Exploring the Antimicrobial and Antiviral Properties of Cryptic Peptides from Human Fibrinogen), MMP-19 residues 1–19 ( Discovery of Encrypted Peptides in a Human Matrix Metallopeptidase), and the eight modern and six archaic validated peptides from panCleave prospecting. The prediction is sharp: release signatures concentrated in late blocks (roughly 25 onward), small in early blocks (0–7), mirroring the two-stage mechanism Lu et al. established. If shifts are equally large in early chemistry blocks, the structural-sequestration story is wrong and the diagnostic claim fails. The second experiment is retrospective enrichment: digest the relevant parent proteomes, rank all fragments by the composite release score, and measure precision-at-k for recovering the validated actives, against random, physicochemical, and APEX-score rankings. The third is the in-silico functional test: on ESCAPE's 82,359-sequence benchmark, test whether layer alignment $\rho_L$ with activity labels predicts multilabel class membership from frozen embeddings alone, with particular attention to the rare classes where supervised baselines struggle. The fourth is the confound and mechanism battery: matched-anchor permutation tests for every alignment claim, and stratification of ranked hits by mechanism of action (inner-membrane depolarizers versus outer-membrane disruptors) to quantify whether, per the tension raised by Obstacles to inferring mechanistic similarity using Representational Similarity Analysis Obstacles to inferring mechanistic similarity using Representational Similarity Analysis, relational alignment conflates biophysically distinct killers — a limitation we will report either way.

The decisive check is prospective and modest in scale: take the top fifty novel human cryptome candidates from the release-signature ranking that APEX scores ambivalently, and run a standard in-vitro MIC panel. The bar is concrete — exceed the 24% physicochemical hit rate decisively, approach or exceed APEX's 59% under an identical validation budget, and show the Euclidean ablation underperforms at matched compute. The fibrinogen and MMP-19 cryptic suites provide assay templates, though we note no reported quantitative benchmark exists for them on Lacuna, and the same is true for the ESKtides database and metagenomic catalogues, so they serve as sources of additional candidates rather than scored comparisons. If the prospective hit rate lands between the ablation and APEX, the honest conclusion is that release signatures are a complementary, structure-aware filter that rescues what sequence-only models miss — a smaller but real contribution, and one that no paper in this literature has yet tested.

References

  1. Mechanisms of AI Protein Folding in ESMFold
  2. Mechanisms of AI Protein Folding in ESMFold
  3. Deep-learning-enabled antibiotic discovery through molecular de-extinction
  4. Molecular de-extinction of ancient antimicrobial peptides enabled by machine learning
  5. A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification
  6. PepCompass: Navigating Peptide Embedding Spaces Using Riemannian Geometry
  7. A Toolbox for Representational Similarity Analysis
  8. A Python Toolbox for Representational Similarity Analysis
  9. ReSi: A Comprehensive Benchmark for Representational Similarity Measures
  10. Can Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color
  11. Representational structure or task structure? Bias in neural representational similarity analysis and a Bayesian method for reducing bias
  12. Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models
  13. Obstacles to inferring mechanistic similarity using Representational Similarity Analysis
  14. AI-Driven Antimicrobial Peptide Discovery: Mining and Generation
  15. In-Context Learning in Large Language Models: A Neuroscience-inspired Analysis of Representations
  16. Peptides from non-immune proteins target infections through antimicrobial and immunomodulatory properties
  17. Discovery of Encrypted Peptides in a Human Matrix Metallopeptidase
  18. Exploring and expanding the chemical multiverse of peptides
  19. The C-terminal 32-mer fragment of hemoglobin alpha is an amyloidogenic peptide with antimicrobial properties
  20. Exploring the Antimicrobial and Antiviral Properties of Cryptic Peptides from Human Fibrinogen
  21. Obstacles to inferring mechanistic similarity using Representational Similarity Analysis