Peptidase models¶

The model panel behind predict_cleavage() and mhctools cleavage. Which enzyme, where it acts, what pattern it assesses and how strongly a match should be read is below; what the result states mean is in reading the evidence.

For model selection by biological endpoint and compartment, see choosing processing and peptidase models.

Running the panel¶

mhctools cleavage --list-models
mhctools cleavage --list-models --json
mhctools cleavage --sequence RPPGFSPFR --model app2-xp --model cpn-basic
mhctools cleavage --sequence VPYGSFKHV --compartment cytosol --out cleavage.json
mhctools cleavage --sequence HAEGTFTSD --model dpp4-qpisa --n-term acetylated
mhctools cleavage --sequence SIINFEKL --model pepsickle-in-vivo-human-only

from mhctools import predict_cleavage
results = predict_cleavage("TSGPNQ", models=["fap-endo-gp", "prep-pro"])

Define a sourced motif rule¶

Use motif_rule() to make a rule outside the built-in catalog. It returns a predictor, so no registry edit is needed:

from mhctools import PeptideInput, motif_rule

nep = motif_rule(
    name="study-nep", enzyme="MME",
    references="https://pubmed.ncbi.nlm.nih.gov/2417254/",
    topology="internal", right="[FILY]",
    strictness="preferred",
    basis="Reported substance P cuts precede Phe or Leu",
    assay="Purified human kidney NEP acting on substance P",
    limitations="Other bonds and substrates can be cleaved",
    c_chemistry=("free", "amidated"),
)
peptide = PeptideInput("RPKPQQFFGLM", c_term="amidated")
result = nep.predict(peptide)
mask = result.pattern_mask  # (0, 0, 0, 0, 0, 1, 1, 0, 1, 0)
matches = [bond for bond, match in enumerate(mask, 1) if match]  # [6, 7, 9]

left and right are Python regexes for the residues before and after each candidate bond. Omit a flank or pass None when it is unconstrained. In the example, [FILY] means one of those four residues immediately after the bond; FILY without brackets would require all four in order. For a terminal enzyme, use topology="n_terminal" or "c_terminal" and state how many residues it removes with removed=. motif_rule() also accepts optional UniProt accession, species, compartment, length, terminal chemistry and activation constraints. A custom rule can omit strictness and basis; if it assigns a grade, it must give the source observation behind that grade. Grades do not change pattern matches. A custom rule reports binary motif matches, never numerical cleavage probabilities. The built-in mme-hydrophobic rule already covers the example above; custom rules should retain the assay and limits supported by their own references.

The default panel evaluates all 41 built-in models and returns separate results. --model and --sequence can be repeated. --list-models prints a compact discovery table; add --json for its full machine-readable catalog. Nine bundled ITCell cathepsin profiles add human B/S internal specificity and H initial N-terminal trimming. Select their 15/60/240-minute source profiles explicitly; these times are assay-training scope, not a new peptide's predicted degradation time.

ITCell models	Mechanism	Native score
`itcell-catb-15`, `itcell-catb-60`, `itcell-catb-240`	Human cathepsin B internal specificity	Sum of log2 profile/background ratios
`itcell-cats-15`, `itcell-cats-60`, `itcell-cats-240`	Human cathepsin S internal specificity	Sum of log2 profile/background ratios
`itcell-cath-15`, `itcell-cath-60`, `itcell-cath-240`	Human cathepsin H initial N-terminal trimming	Sum of log2 profile/background ratios

Ten bundled PhageScout sequence profiles add separate human ELANE and CTSG native scores. The exact names are phagescout-elane-pwm-deseq2, phagescout-elane-pwm-relaxed-unaligned, phagescout-elane-pwm-relaxed-aligned, phagescout-elane-peptide-relaxed-unaligned, phagescout-elane-peptide-relaxed-aligned, phagescout-ctsg-pwm-deseq2, phagescout-ctsg-pwm-relaxed-unaligned, phagescout-ctsg-pwm-relaxed-aligned, phagescout-ctsg-peptide-relaxed-unaligned and phagescout-ctsg-peptide-relaxed-aligned. These are phage-derived recognition features with inferred aligned P1 anchors. Sparse profile non-matches are unassessed; neither their native scores nor negative PWM values are serum-loss percentages.

Two optional full DESeq2 peptide lookups, phagescout-elane-peptide-deseq2 and phagescout-ctsg-peptide-deseq2, retain the mean native log2 fold change over matching complete five-mers. Install their verified data separately with mhctools fetch phagescout; catalog discovery does not load the tables.

The eight optional Pepsickle models cover epitope and C/I digestion families; see proteasome models. Missing assets and uninspected external runtimes are listed as unresolved. Predictions carry SHA-256 identities for the actual runtime's weights, inference code, feature code and dependency metadata. Prediction JSON uses schema version 4: its top-level models object stores each full provenance record once, keyed by model name, and each item in results references that name in its model field. Motif sites carry match: 0|1, with the rule documented once in model metadata; this is a binary pattern indicator, not a cleavage probability. The output retains non-matches and unsupported results, source coordinates and chemistry. Numerical scores are written with six significant digits. --source-start is a zero-based offset shared by the supplied inputs; use separate calls when fragments have different offsets. The --list-models --json catalog uses schema version 2 and omits metadata fields that do not apply to a model. Schema-v2 prediction JSON remains readable by the vaccine-report command.

Compartment filtering uses exact, conservative enzyme-location annotations. serum, plasma, extracellular, cytosol, endosome and er are distinct. The extracellular filter is useful for broader candidate screening; serum is not an exhaustive inventory of everything potentially present in a serum sample. Presence, concentration, activation, inhibitors and exposure are not inferred. No combination is converted into overall stability.

The models¶

Strictness grades are explained in reading the evidence.

Model	Assessed recognition pattern	Strictness	Main scope and primary evidence
`dpp4-qpisa`	N-terminal P2-P1\|P1′ score	scored	Human DPP4; quantitative model described above
`dpp4-p1-pro-ala`	N-terminal X-(Pro/Ala)\|X preference	preferred	Human DPP4; short substrates and NPY analogues; other P1 residues can still be cut
`dpp4-p1prime-nonpro`	N-terminal X-X\|non-Pro preference	preferred	Human DPP4; qPISA peptide validation found no product for tested P1′ Pro peptides; scope is that test set
`ace-dipeptidyl`	C-terminal \|non-Pro–non-Asp/Glu	required	Ordinary human ACE dipeptide activity; angiotensin assays
`mme-hydrophobic`	Selected P1′ residues Phe/Ile/Leu/Tyr	preferred	Human neprilysin; kidney peptide assays; incomplete whole-sequence specificity
`cpb2-basic`	C-terminal \|Lys/Arg, explicitly active enzyme	required	Human TAFIa; chemerin cleavage; unknown/zymogen/inactive states abstain
`cpn-basic`	C-terminal \|Lys/Arg	required	Human plasma CPN; Oshima et al. 1975
`app1-xp`	N-terminal X\|Pro	required	Cytosolic human XPNPEP1; Cottrell et al. 2000; manganese dependent, distinct gene product from XPNPEP2
`app2-xp`	N-terminal X\|Pro	required	Human XPNPEP2; Molinaro et al.
`fap-dipeptidyl`	N-terminal X-Ala/Pro\|non-Pro	preferred	Human FAP; substrate profiling and direct Ala/Pro cleavage. Ala was much slower in the tested cohort.
`fap-endo-gp`	Gly-Pro\|non-Pro	required	FAP endopeptidase; substrate profiling, prime-side constraint
`enpep-acidic`	N-terminal Asp/Glu\|X	preferred	Aminopeptidase A; human specificity study; calcium and sequence affect activity
`anpep-ala`	N-terminal Ala\|X preference	permissive	Aminopeptidase N; human structure/biochemistry; many other substrates omitted
`dpp8-xp-xa`	N-terminal X-(Pro/Ala)\|non-Pro	required	Cytosolic DPP8; characterization, degradomics
`dpp9-xp-xa`	N-terminal X-(Pro/Ala)\|non-Pro	required	Cytosolic DPP9; antigen processing, same degradomics study
`tpp2-tripeptidyl`	N-terminal tripeptide, non-Pro P1 and P1′	permissive	Cytosolic TPP2; RU1 precursor processing; topology, not selectivity
`npepps-n-terminal`	First bond lacking Gly/Pro context	permissive	Puromycin-sensitive aminopeptidase; same RU1 study; broad enzyme, weak flag
`prep-pro`	Internal X-Pro\|X	required	PREP/POP, conservative 4–30-residue domain; human profiling; flanking preferences omitted
`erap2-basic`	N-terminal Arg/Lys\|X preference	preferred	ERAP2 in ER; biochemistry, peptide structures; not a full-context predictor
`thop1-observed`	Exact-sequence source lookup	source observations	Cytosolic THOP1; Knight et al. 1995
`nln-observed`	Exact-sequence source lookup	source observations	Cytosolic NLN; human neurolysin structures and LC-MS
`lnpep-observed`	Exact-sequence source lookup	source observations	Endosomal LNPEP/IRAP; Georgiadou et al. 2010
`eramer-step` (optional)	Initial N-terminal bond; length-specific PWM score	scored	ERAP1 in ER; ERAMER, 9–16 residues

The motif models return decisions without numerical scores. For example, APP removes the first residue of RPPGFSPFR at R|PPGFSPFR; DPP-like activity would remove two residues and is a separate assessment. FAP's endopeptidase rule can also assess an N-acetylated input, supported by its blocked-substrate activity. ACE also accepts N-acetylation, and MME accepts C-amidation. Other rules conservatively accept free termini only; an unsupported modified input does not establish that its bonds resist enzymatic cleavage.

Proteasome models¶

Eight optional Pepsickle models cover the epitope and digestion families, with explicit constitutive (C) or immunoproteasome (I) selection where the family has it. They are excluded from the default panel and selected by exact name:

Model	Family	Context	Proteasome
`pepsickle-in-vivo-human-only`	epitope-trained neural ensemble	8 residues before and after P1	agnostic
`pepsickle-in-vivo-all-mammal`	epitope-trained neural ensemble	8 residues before and after P1	agnostic
`pepsickle-in-vitro-2-human-only-constitutive`	digestion-trained neural ensemble	3 residues before and after P1	C
`pepsickle-in-vitro-2-human-only-immunoproteasome`	digestion-trained neural ensemble	3 residues before and after P1	I
`pepsickle-in-vitro-2-all-mammal-constitutive`	digestion-trained neural ensemble	3 residues before and after P1	C
`pepsickle-in-vitro-2-all-mammal-immunoproteasome`	digestion-trained neural ensemble	3 residues before and after P1	I
`pepsickle-in-vitro-all-mammal-constitutive`	gradient-boosted digestion model	needs the isolated scikit-learn 0.23.2 runtime (see below)	C
`pepsickle-in-vitro-all-mammal-immunoproteasome`	gradient-boosted digestion model	needs the isolated scikit-learn 0.23.2 runtime	I

The Python Pepsickle wrapper takes model_type and proteasome_type (C or I) for digestion models. It rejects a proteasome type for the epitope model and human_only=True for gradient boosting, which upstream ignores. Model metadata identifies actual weights, code, population and model family. The upstream endpoint sentinel is excluded from canonical bond results.

The legacy gradient-boosted runtime¶

Provision it with Docker:

python scripts/setup_test_backends.py pepsickle
source env/test-backends/activate.sh

This pins Python 3.8.20, scikit-learn 0.23.2 and its companion packages in a separate container. Its generated PEPSICKLE_GB_PYTHON launcher selects that runtime only for gradient boosting and runs inference with networking disabled. Neural models keep their existing runtime. A separately managed interpreter can be selected with Pepsickle(python_executable=...) or PEPSICKLE_PYTHON for every model family.

Prediction provenance comes from the selected interpreter, including package versions and actual code/weight hashes. A changed identity between inspection and inference causes failure. Catalog listing does not start external runtimes; their identities remain unresolved until the predictor is selected. No model is silently substituted when a runtime is missing or incompatible.

Adapter agreement with upstream inference establishes implementation conformance, not accuracy for tumor/APC processing or vaccine-peptide survival.

Human DPP4 qPISA¶

The Gudipati et al. 2024 paper reports a model fitted to substrate depletion by purified human DPP4. The source assay used tryptic HeLa peptides in HEPES pH 7.4, at 21 C for 4 hours. The implementation independently evaluates the three rearranged terms in Dataset EV2: P1 + P2:P1 + P1:P1-prime, where the first three peptide residues are P2, P1 and P1-prime. Only bond 2 is assessed.

Higher scores indicate greater predicted log2 depletion relative to buffer control in that assay. Negative values are retained. Scores are not cleavage probabilities, serum half-lives or stability ranks across enzymes. Compartment metadata records where the enzyme can act, not where the model has been calibrated. Physiological exposure, structure, concentration and competing enzymes are not modeled.

All triplets with complete coefficients can be evaluated, including those without Pro or Ala at P1. Of 8,000 canonical triplets, 6,420 have complete coefficients; the other 1,580 return an explicit missing-coefficient reason. Evaluability does not establish that an individual triplet occurred in the training data. The related C. elegans DPF-3 model is not used for human DPP8/9.

Parameter provenance¶

mhctools/data/dpp4_qpisa.json contains the numeric cells from 44320_2024_71_MOESM3_ESM.xlsx, sheet dpp4_modelParams. Published NA cells become JSON null; no numerical imputation or refitting is performed. The source workbook SHA-256 is ee449da13b5ec66fd6fb08c16203da8c44f2e9c10572a355abdcb985694c6ddc. It was retrieved via the Europe PMC supplementary archive. The article assigns associated data to CC0, unless otherwise credited; Dataset EV2 has no separate credit restriction. Attribution: Rajani Kanth Gudipati and colleagues, 2024, DOI above. No figures or upstream R source code are redistributed.

A bounded purified-DPP4 fraction estimate¶

DPP4qPISA().estimate_keane_fraction(peptide, minutes=...) maps the qPISA triplet score to a parent-peptide loss fraction in one purified-enzyme assay. It fits log half-life to the eleven 27–42-residue peptide measurements in Keane et al. 2011, Tables 1–2: 0.112 µM recombinant human DPP4 at 37 °C, with MALDI-TOF parent-peptide time courses. The fit gives each substrate one observation; the published mean half-lives, SDs, replicate counts, N-terminal triplets and lengths are transcribed in mhctools/data/dpp4_keane2011.json. BNP and substance P lack a qPISA coefficient; censored non-cuts are never assigned a zero rate. The fit treats the reported amide-terminated source peptides as part of the training set, but estimates only free-terminus inputs accepted by DPP4qPISA.predict().

from mhctools import DPP4qPISA

# Synthetic 30-residue illustration, not a measured vaccine peptide.
estimate = DPP4qPISA().estimate_keane_fraction("HAE" + "G" * 27, minutes=10)
print(estimate["parent_lost_fraction"])  # about 0.33 in this assay model
print(estimate["parent_remaining_fraction"])  # about 0.67

The method evaluates only the first, exposed DPP4 bond (bond 2), lengths 27–42, supported unmodified qPISA inputs, and scores within the training range. Otherwise it returns status="unassessed" with a reason. It retains the native qPISA score separately. A first-order loss assumption converts the fitted half-life into a fraction at the requested time: parent_remaining_fraction = 2 ** (-minutes / estimated_half_life_minutes). This is a modeling assumption; Keane's half-lives are experimental parent-loss summaries, not proof that every new peptide follows one exponential.

There are only eight unique triplets among the eleven training substrates. Leaving out each complete triplet group in turn gives a mean absolute error in log half-life corresponding to a 2.8-fold typical error. That is a validation error summary, not a confidence interval or a guarantee for new substrates. qPISA was trained in a separate, median-normalized 21 °C assay; its score cannot itself be read as an absolute fraction. Other sequence positions, substrate structure, enzyme amount and matrix all limit transfer. In particular, this estimate is not a serum, lymph, IM-vaccine or APC-arrival fraction and is not automatically passed to assess_vaccine_windows().

The corresponding purified-FAP estimate¶

The same Keane et al. 2011 Table 2 reports nine numerical parent half-lives for 11–42-residue hormones and neuropeptides incubated with 0.2 µM purified human FAP at 37 °C. Five substrates with Ala at the second residue had measured half-lives of about 15.5–39.1 hours; four with Pro ranged from 5.78 to 60.2 minutes. These are different enzyme concentrations from the DPP4 assay and must not be compared as intrinsic in-vivo clearance rates.

estimate_assay_parent_fraction() gives a common entry point for the two bounded estimates. For FAP it uses the geometric mean measured half-life of the Ala or Pro group, then assumes first-order parent loss. It accepts only 11–42-residue peptides with one of the eight measured N-terminal triplets, an exposed free N terminus, and a free or amidated C terminus. The rest of the sequence is unmodeled; this is a rough cohort prior, not a fit to full peptide sequence or a claim that an arbitrary motif match has this rate.

from mhctools import AssayConditions, DPP4, FAP

peptide = "HAE" + "G" * 27  # synthetic illustration, not a measured substrate
for enzyme in (DPP4(), FAP()):
    result = enzyme.predict(peptide).assay_fractions[0]
    print(enzyme.enzyme, result["status"], result.get("estimated_half_life_minutes"),
          result.get("parent_remaining_fraction"))

scenario = AssayConditions(exposure_minutes=60, enzyme_concentration_um=0.2)
result = DPP4().predict(peptide, conditions=scenario).assay_fractions[0]

The FAP Ala group has 5 source substrates and 1.4-fold typical leave-one-substrate-out error; the Pro group has 4 and 3.0-fold typical error, with PYY a conspicuous outlier. Those checks reuse N-terminal groups already seen during fitting. They do not validate transfer to a novel sequence with the same triplet, and the source includes censored non-cuts which are excluded rather than assigned zero rates. The output names the source assay, model, estimated half-life, lost and remaining parent fractions, and group validation. The default fractions use a 60-minute reference exposure in each source assay; a supplied enzyme concentration scales the rate linearly as a sensitivity assumption. Unassessed inputs contain a reason and no fabricated fraction. Neither estimate is core-epitope survival, serum half-life, IM exposure or APC arrival; broader calibration remains #587.

For names, keep three distinct things visible: fap-dipeptidyl, dpp4-p1-pro-ala and dpp4-p1prime-nonpro are separate binary recognition patterns, dpp4-qpisa is a native numerical substrate-depletion score, and fap-keane2011-parent-loss or dpp4-qpisa-keane2011-parent-loss is an assay-specific fraction at a specified time. A bare name like FAP_HalfLife would hide the assay and suggest a transferable biological half-life. The enzyme-centered API discovers and runs all applicable methods for an enzyme object. The lower-level estimate_assay_parent_fraction(peptide, enzyme=...) remains available for configuration-driven calls; its result retains the source and assay.

Other enzymes need different evidence before they get a time-specific parent-loss estimate for an arbitrary canonical peptide. A human CatS substitution series reports digestion and kinetic constants for internally quenched peptides with dye, D-amino-acid and amide modifications around one scaffold; treating its fitted constants as rates for unmodified vaccine peptides would silently change the substrate chemistry. A human neprilysin kinetic panel reports constants for a small set of named substrates, but no validated sequence-to-rate mapping for novel windows. Their motif or native scores can still be used for the endpoints each measures; neither is an absolute per-minute fraction.

Serum and extracellular candidates¶

mhctools cleavage --sequence DRVYIHPFHL --model ace-dipeptidyl --model mme-hydrophobic
mhctools cleavage --sequence YFPGQFAFSK --model cpb2-basic --enzyme-state CPB2=active
mhctools benchmark --reference-cleavage serum --out serum-reference.json

In Python, supply enzyme_states={"CPB2": "active"} to predict_cleavage, or get_cleavage_model("cpb2-basic", enzyme_state="active"). The result records that assumption in conditions. CPB2 requires proteolytic activation; its presence as a zymogen does not establish activity. unknown, zymogen and inactive return no assessed sites. Activation experiments also show that the surrounding coagulation environment matters.

ACE's ordinary rule assesses only the bond before the last two residues. It recognizes angiotensin I cleavage at bond 8 and the N-acetyl-SDKP bond 2. It abstains on amidated substance P even though human ACE is known to cleave that peptide at bonds 8 and 9 through exceptional processing. MME flags selected hydrophobic residues after a bond, including substance P's reported bonds 6, 7 and 9; it accepts that peptide's C-terminal amide. These human enzyme observations do not establish a transferable model of all substrates. The MME rule's 2–30-residue domain is a conservative implementation scope; longer substrates can exist. Use the extracellular filter to include MME: location annotations do not assert that purified-kidney specificity was calibrated in serum.

The packaged serum-candidate reference contains 17 source observations, including two explicitly reported non-cleavages. Fifteen are assessable by the corresponding rules; the two exceptional ACE/substance-P observations abstain. Cross-enzyme combinations are unassessed. The source IDs, chemical forms and known or unspecified experimental conditions are retained. These small reproduction controls do not establish specificity, serum half-life performance, or validity on long peptides. The report explicitly shows missing serum-half-life and long-peptide evidence.

Cytosolic and endosomal candidates¶

mhctools cleavage --sequence VPYGSFKHV --compartment cytosol
mhctools cleavage --sequence KSLYNTVATL --compartment endosome
mhctools cleavage --sequence RPPGFSPFR --model thop1-observed --model app1-xp
mhctools benchmark --reference-cleavage intracellular --out intracellular-reference.json

Antigen-processing peptidases are the reason this panel exists, but three of them publish their specificity as whole-substrate outcomes rather than as a transferable pattern. Inventing a motif from those papers would misrepresent them, so thop1-observed, nln-observed and lnpep-observed are source references instead: an exact sequence and chemical form returns what the experiment reported, and anything else abstains. There is no nearest-neighbour matching and no extrapolation. Results carry a substrate_observation of cleavage_reported or no_cleavage_detected alongside any bonds the source's product identities actually pin down. Where a source saw degradation but no intermediate, cleavage is recorded with no bond rather than guessing one.

A reported non-cleavage never becomes a per-bond label. no_cleavage_detected means that assay saw no loss of that peptide under its own conditions and detection limits, which is not the same as a bond that cannot be cleaved.

The remaining cytosolic enzymes are ordinary motif rules, and their grades matter. app1-xp is required: aminopeptidase P is defined by hydrolysing the X-Pro bond, so a non-match is real evidence. tpp2-tripeptidyl and npepps-n-terminal are permissive. Removing three residues describes TPP2's topology, not which peptides it turns over, and puromycin-sensitive aminopeptidase is broad enough that its Gly/Pro exclusion is only a hint. Do not read a TPP2 match as a prediction that the peptide is consumed.

THOP1 and NLN are closely related and are deliberately kept apart. They cleave neurotensin at different bonds and their specificities can be swapped by mutating two active-site residues, so neither model's observations transfer to the other. Only human-enzyme observations are curated: the widely cited bradykinin, enkephalin and neurotensin results for neurolysin come from rat or species-unspecified preparations and are excluded rather than relabelled human.

lnpep-observed covers IRAP as an endosomal cross-presentation candidate, not a second ER enzyme. Its source digested peptides at pH 8.0 with purified enzyme, so the records describe that experiment, not an acidified endosome.

The packaged intracellular reference contains 42 source observations across four models, including six reported non-cleavages. Forty-one reproduce; the amidated substance P record abstains because C-amidation is outside the aminopeptidase P rule's documented input domain. One resistant IRAP precursor from Georgiadou 2010 is excluded entirely: the paper prints DIRSSVQNKL in its results and Table I but DIRSSQVNKL in the Figure 3F caption, and an exact-sequence catalog cannot silently pick one. Both strings are retained in the dataset notice, and the discrepancy is tracked in known gaps.

Thimet oligopeptidase contributes no non-cleavage record at all. Every resistant peptide in its source is a hydroxyproline analogue or carries an N-terminal pyroglutamate, and both are outside the canonical-peptide input type. Its absence from the negatives is a curation limit, not a finding.

The report shows no evidence for antigen presentation in primary dendritic cells and none for cleavage measured in cytosol rather than purified enzyme. Those gaps are the point: nothing here calibrates how long a peptide survives in the cytosol of an antigen-presenting cell, where the proteasome, competing aminopeptidases and TAP transport all act at once.

ERAP1 and existing processing models¶

mhctools fetch eramer
# Install openpyxl in the environment if it is not already available.
mhctools cleavage --sequence LAAAFGAAA --model eramer-step

Alternatively, construct ERAMERCleavage(pwm_path="/path/to/PWM.xlsx") or set ERAMER_HOME. This optional model is listed without loading assets and is excluded from the default panel. Explicit selection fails clearly if the external asset or its runtime is missing.

eramer-step computes the existing ERAMER intermediate PWM specificity for one exposed 9–16-residue precursor, assigning it to bond 1. It does not report the average of later trimming intermediates. Its version contains the SHA-256 of the actual workbook snapshot used for inference. The GPL-licensed workbook is loaded at runtime and is not included in the mhctools distribution.

The existing ERAMER cascade API, NetChop, Pepsickle and other proteasome predictors remain available through their existing interfaces. Pepsickle also has a canonical PepsickleCleavage.predict() facade and the two CLI model names shown above. Array index i maps to internal bond i + 1; the upstream final zero is an endpoint sentinel and is never emitted as a bond. Canonical results default to subprocess isolation because the upstream package deserializes a pickle, while still requiring users to trust the installed model artifact. Processing-model output scales must not be mixed with qPISA scores or motif decisions.