Peptide drugs sit in an attractive middle ground between small molecules and antibodies: they bind tightly to protein surfaces that small molecules cannot reach, yet they are chemically simple to make. Their central failure mode is that the body's digestive enzymes — proteases — shred them within minutes of injection or swallowing. A given peptide in blood is not attacked by one enzyme but by dozens at once, and a researcher trying to stabilize it faces two separate puzzles: which of those enzymes actually matters, and what chemical modification will stop them without wrecking the peptide's binding to its intended target. Functional enrichment analysis is a mature statistical technique that solves an analogous problem in genomics: it takes an uninterpretable list of hundreds of perturbed genes and maps it onto curated pathway dictionaries (Gene Ontology, KEGG, Reactome), revealing the coordinated biological programs behind the noise.
This proposal inverts that machinery. Instead of enriching a list of genes for disease pathways, we will treat the set of proteases predicted to cleave a peptide candidate as the input "list," and enrich it against a degradome ontology — proteases organized into the physiological cascades they belong to (coagulation, complement, kinin–kallikrein, lysosomal cathepsins, gastrointestinal peptidases, the proteasome) and the tissues where those cascades are active. The output is a tissue-specific clearance profile: "this peptide is dominated by the kallikrein pathway in plasma and by trypsin-family enzymes in the gut." We claim this pathway-level abstraction is not just a nicer report — it is the missing predictive context that explains why cleavage predictors have stalled, and the right design signal for choosing where to place a stabilizing chemical modification.
Core Claim
Two independent findings in the literature point at the same gap. First, cleavage prediction from sequence alone has hit a hard ceiling: in What cleaves? Is proteasomal cleavage prediction reaching a ceiling?, scaling from a 30,000-parameter MLP to a 1.2-billion-parameter protein language model left performance flat at roughly 88.5% AUC for C-terminal and 79.5% AUC for N-terminal cleavage on over 1.4 million ligandomics samples, and denoising training schemes actually hurt — evidence that the residual error is structured by biology the sequence window cannot see. Second, when proteases act in mixtures, per-enzyme resolution is mathematically ill-posed: Deconvolving multiplexed protease signatures with substrate reduction and activity clustering showed with seven recombinant proteases and seven barcoded substrates that individual enzymes cannot be separated when their specificities overlap, but clustering them into functional families — which turned out to coincide exactly with the complement and coagulation cascades — restores accurate deconvolution.
Our claim is that the pathway family is the correct unit of analysis for peptide degradation, and that using it twice pays off twice: as conditioning context for cleavage prediction (attacking the ceiling) and as a design target for stabilization (attacking the stability–affinity trade-off documented in Atom-level backbone engineering preserves peptide function while enhancing stability, where the most protective modifications destroyed binding — D-Phe$^8$ extended Bradykinin's half-life to 51.4 minutes but cut receptor affinity nearly 60-fold to an $IC_{50}$ of 400 nM, and N-methylation abolished inhibition entirely at 1 µM — while the pathway-targeted azapeptide at the single kininase hotspot kept affinity at 4.3 nM with an 8-fold half-life gain). No existing work runs enrichment in this direction: iProt-Sub applied GO/KEGG enrichment post-hoc to natural host proteins predicted to be cleaved, not to the clearance of incoming peptide drugs; PeptiVerse predicts half-life as an isolated scalar with no mechanistic attribution (PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction); and generative designers such as PepMLM explicitly leave metabolic stability to future work.
Research Program
The first component is a protease-to-pathway conditioning model for cleavage prediction. Today's predictors either ignore protease identity (the proteasomal models) or train a separate model per protease (iProt-Sub builds per-protease SVR models from a 4,562-dimensional feature space over 3,688 substrates and 6,637 cleavage sites across 38 proteases, reaching AUC values often above 0.95). We will instead train one model whose input is the peptide window concatenated with an embedding of the protease, where the protease embedding is anchored to its pathway family and catalytic class rather than learned free-form from a single enzyme's data. Because verified non-cleavage sites are scarce, we adopt positive-unlabeled learning following Improved proteasomal cleavage prediction with positive-unlabeled learning, which reported 78.00% N-terminal and 87.20% C-terminal AUROC where NetChop managed only 49.27% on the N-terminal task. The pathway anchor should let the model share statistical strength across enzymes in the same cascade and, critically, make zero-shot predictions for proteases with no training data — something per-protease models cannot do at all.
The second component is the reverse enrichment engine. Given a candidate peptide — including non-canonical ones encoded as SMILES, since cyclized, D-, and aza-modified peptides have no gene symbol that classical tools require — we run the conditioned model across the protease panel and obtain cleavage probabilities $\hat{y}_i(p)$ for protease $i$ against peptide $p$. The pathway-level susceptibility score for a cascade $\mathcal{F}$ in tissue $t$ is then
$$S(p, \mathcal{F}, t) \;=\; \sum_{i \in \mathcal{F}} e_i(t)\,\hat{y}_i(p),$$
where $e_i(t)$ is the abundance of active protease $i$ in tissue $t$ (plasma, liver, gastrointestinal lumen). To resolve which enriched pathways are functionally the same threat, we cluster them semantically rather than by member overlap, because Contrastive Learning for Gene Set Enrichment Analysis Post-Processing showed that over 80% of pathway pairs share zero genes despite describing the same biology, and its gtCLIP embedding achieved 3.3-fold higher within-cluster overlap and 3.9-fold higher silhouette scores than overlap-based clustering on a corpus of 12,612 gene sets. We will adapt that joint text–set embedding to a degradome ontology built from MEROPS families mapped onto Reactome and KEGG cascades.
The third component is pathway-aware stabilization guidance. The clearance profile dictates strategy: when one pathway dominates, a single-site modification at that pathway's cleavage hotspot (azapeptide or $\alpha$-methylation) should suffice and preserve affinity, as it did for Bradykinin's Phe$^8$; when several unrelated cascades dominate, only global strategies (cyclization following A novel cyclic pentapeptide, H-10, inhibits B16 cancer cell growth and induces cell apoptosis and Identification and Validation of Cyclic Peptides with Mucin-Selective, Location-Specific Binding in the Gastrointestinal Tract, or full stereochemical inversion following Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design) are worth their cost. We will test this decision rule retrospectively on the Bradykinin benchmark and prospectively on binders generated by PepMLM, whose authors note their linear peptides need exactly this kind of stabilization.
Proposed Experiments
The prediction claim is tested on two benchmarks from the file. On the MEROPS/UniProt 38-protease benchmark (3,688 substrates, 6,637 cleavage sites, the same 5/6–1/6 split iProt-Sub used), we compare per-protease AUC against iProt-Sub's reported values (often above 0.95, up to 1.0 for some caspases) and, more decisively, evaluate on proteases held out entirely from training, where pathway conditioning should transfer and per-protease baselines cannot predict at all. On the MHC ligandomics proteasomal benchmark, the target to beat is the documented ceiling of roughly 88.5% C-terminal and 79.5% N-terminal AUC from What cleaves?; the honest expectation is small or no gains here, because the proteasome is a single pathway and conditioning adds little — this negative control matters, since it pins any gains on the multi-pathway benchmarks to the conditioning ingredient rather than to extra parameters. The enrichment component is validated against the one kinetic ground truth available: the protease family structure recovered by Deconvolving multiplexed protease signatures must be reproduced by our semantic clustering (complement separated from coagulation), and the retrieval quality of the adapted gene-set embedding is scored against gtCLIP's reported Recall@5 of 51.3% on held-out pathways. For tissue-level clearance prediction, we evaluate whether pathway scores $S(p,\mathcal{F},t)$ separate tissue-annotated half-life records in the THPdb2, PEPlife, and PepTherDia cohorts that PeptiVerse drew on; no reported result found on Lacuna for PeptiVerse's own half-life correlation, so we will establish this reference number ourselves on the same data under identical splits. Finally, the stabilization guidance is tested retrospectively against He et al.: given native Bradykinin, the pipeline must identify the kininase pathway as dominant and rank azapeptide modification at Phe$^8$ above D-substitution and N-methylation — the ordering that matches the measured $t_{1/2}$ of 28.6 minutes with retained 4.3 nM affinity. A prospective arm synthesizes the top-ranked and deliberately mis-ranked modifications for a PepMLM-designed binder, with synthesis cost tracked since azapeptides took He et al. four hours versus one hour for standard analogues.
Baselines
The naive reference point is a protease-blind sequence-only cleavage model, exactly the family benchmarked in What cleaves?: trained on the same 1.4-million-sample ligandomics data with the same window size and tuning budget, its published numbers — about 88.5% AUC on C-terminal and 79.5% on N-terminal cleavage — define the floor any conditioned model must clear to justify its added inputs. The strongest existing competitor on the multi-protease question is iProt-Sub, whose per-protease SVR models reported AUC values often above 0.95 (up to 1.0 for some caspases) on the 38-protease MEROPS benchmark; we will retrain and evaluate it under identical splits, feature budgets, and compute, and count ourselves as beating it only if we match its per-protease AUC while also succeeding on held-out proteases it cannot predict at all. On the proteasomal benchmark specifically, the strongest reported number is PUUPL's 78.00% N-terminal AUROC (versus 49.27% for NetChop) and 87.20% C-terminal AUROC, so PUUPL is retrained as the per-benchmark champion. For the clearance-profile component, the scalar baseline is PeptiVerse's multi-task property platform, which reported Spearman correlation of 0.58 on amino-acid sequences and 0.56 on SMILES for binding affinity; since it reports no half-life correlation on Lacuna, we will produce the first tissue-annotated half-life numbers on its own datasets. The key ablation isolates the pathway anchor: the identical architecture with the protease embedding replaced by a free-form per-protease identifier (one-hot or learned from that enzyme's own data only), same data, same parameter count. If the pathway-anchored version does not beat this ablation on held-out proteases and cross-tissue transfer, the core claim fails.
Why not just…?
Why not just run Fisher's exact overrepresentation on the predicted cleaving proteases? This is the cheapest version of the idea and must be run as a baseline rather than argued away. The evidence against it is direct: gtCLIP's authors found that over 80% of pathway pairs share zero genes, so overlap statistics fragment functionally identical clearance mechanisms into uninterpretable lists, and their semantic embedding beat overlap-based clustering 3.3-fold on within-cluster overlap. Still, for the coarse complement-versus-coagulation split, plain ORA may suffice — that is a real threat, and Experiment two settles it by comparing ORA and semantic clustering on the harder case: distinguishing kininase-driven from cathepsin-driven clearance where the enzyme sets are largely disjoint. If ORA matches, we report the simple method as adequate and the embedding becomes optional.
Why not just fine-tune a large protein language model on more cleavage data? The ceiling paper already tried the scale route: a 1.2-billion-parameter ProtTrans T5-XL and 150-million-parameter ESM2 landed on the same plateau as a 30k-parameter MLP, denoising regimes underperformed, and byte-pair tokenizers actively hurt by destroying the local scissile motifs. That is a negative result that constrains design, not a proof that context is the answer — but our hypothesis is falsifiable on exactly this point: the ablation above, plus a protease-conditioned ESM-2 fine-tune given the same data, tests whether conditioning rather than scale moves the needle. If the conditioned language model matches our smaller pathway-anchored model, the win belongs to conditioning generally, and we will say so.
Why not just invert the whole peptide to D-amino acids or N-methylate it everywhere? On established bioactive scaffolds this fails empirically: He et al. measured a 60-fold affinity loss for D-Phe$^8$ ($IC_{50} = 400$ nM) and complete loss of inhibition for N-methylation, while only the hotspot-targeted azapeptide kept both stability and function. The honest caveat is that full D-inversion works for de novo binders — PepMirror's D-1412 bound CD38 at roughly 10 µM and the D3 candidate reached 58% oral bioavailability — because the whole binding interface is designed around the inverted geometry. Our framework targets the established-scaffold and peptide-optimization regime, where the binding geometry is fixed and brute-force inversion is exactly what the Bradykinin data rules out; for de novo campaigns, the two approaches compose rather than compete.
Decisive Checks
Three outcomes decide the proposal. First, on held-out proteases of the 38-protease MEROPS benchmark, the pathway-anchored model must substantially beat the identity ablation and the per-protease iProt-Sub baseline (which scores at chance there); this is the single cleanest test of whether pathway context carries real predictive signal. Second, on the MHC ligandomics benchmark we predict no significant gain over the 88.5%/79.5% plateau — a confirmed null here strengthens the causal story that gains come from multi-pathway context, not from architecture. Third, the retrospective Bradykinin test: the pipeline must recover kininase dominance and rank azapeptide first among the four modifications, reproducing the measured ordering of $t_{1/2}$ and $IC_{50}$. If all three land as predicted, the deliverable is a working tool that turns a peptide's degradation problem into a readable pathway report — the first time enrichment analysis runs backwards, from a molecule to the physiology that will clear it.
References
- What cleaves? Is proteasomal cleavage prediction reaching a ceiling?
- Deconvolving multiplexed protease signatures with substrate reduction and activity clustering
- Atom-level backbone engineering preserves peptide function while enhancing stability
- iProt-Sub: a comprehensive package for accurately mapping and predicting protease-specific substrates and cleavage sites
- PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction
- Target sequence-conditioned design of peptide binders using masked language modeling
- Improved proteasomal cleavage prediction with positive-unlabeled learning
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- A novel cyclic pentapeptide, H-10, inhibits B16 cancer cell growth and induces cell apoptosis
- Identification and Validation of Cyclic Peptides with Mucin-Selective, Location-Specific Binding in the Gastrointestinal Tract
- Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design