Enzyme-centered predictions¶

Use a named enzyme class when the question starts with an enzyme: what activity has been measured, and what can each available method say about my peptide? It discovers the enzyme's named methods and runs them together, while keeping their endpoints separate. It does not turn a pattern match or native score into a probability of cleavage.

from mhctools import AssayConditions, DPP4, FAP

dpp4 = DPP4()
for method in dpp4.describe()["methods"]:
    print(method["name"], method["evidence_type"], method["description"])

peptide = "HAE" + "G" * 27  # synthetic example, not a measured substrate
report = dpp4.predict(peptide)
print(report.summary())

# Same dimensionless endpoint for each assay-backed enzyme:
for enzyme in (DPP4(), FAP()):
    estimate = enzyme.predict(peptide).assay_fractions[0]
    print(enzyme.enzyme, estimate["parent_remaining_fraction"])

# A sensitivity scenario: 60 minutes at 0.2 µM active enzyme.
scenario = AssayConditions(exposure_minutes=60, enzyme_concentration_um=0.2)
print(dpp4.predict(peptide, conditions=scenario).assay_fractions[0])

The DPP4 report for this example contains four independent answers:

Method	What it says	Example result
`dpp4-qpisa`	Predicted substrate depletion in Gudipati et al.'s purified-human-DPP4 assay; native log2 score	2.1694 at the exposed second bond
`dpp4-p1-pro-ala`	Kühn-Wache et al.'s P1 Pro/Ala preference; a binary motif match	Matches `HA\\|E`; other P1 residues can still be cleaved
`dpp4-p1prime-nonpro`	The qPISA study's tested P1′ non-Pro preference; a separate binary motif match	Matches `HA\\|E`; a tested P1′ Pro non-match is not a universal no-cut claim
`dpp4-qpisa-keane2011-parent-loss`	Estimated full-length parent half-life and fraction remaining in the Keane et al. purified-human-DPP4 assay	About 17-minute half-life and 8.8% remaining after the default 60 minutes, within the calibration domain

HA|E marks the bond after the first two residues, not a new peptide sequence. The two motifs come from distinct source observations. A peptide can miss either preference and still receive a qPISA score; no motif is used as a hard gate for the quantitative model or the assay estimate. The Keane fraction uses a fitted half-life and first-order approximation, and may return unassessed outside its measured length and score range.

The default 60-minute exposure is a reporting reference, not an observation about vaccination. Both enzymes emit parent_remaining_fraction between 0 and 1, parent_lost_fraction, and estimated_half_life_minutes. By default each uses its own purified-assay conditions: DPP4 used 0.112 µM enzyme at pH 7.6, FAP used 0.2 µM at pH 8.0, and both used 1 µg peptide in 15 µL at 37 °C. Thus the fractions share an endpoint and time but are not intrinsic cross-enzyme rates. Setting one AssayConditions concentration for both gives a sensitivity comparison at the same assumed active-enzyme concentration; it still does not equalize pH, substrate molarity, or enzyme accessibility. Concentration scaling assumes the first-order parent-loss rate changes linearly with active enzyme. There are no kinetic measurements here to vary substrate loading, pH, or temperature.

Inspect the full source, assay, compartment, activity and limitations for each method with dpp4.describe(). Retrieve native results by model name:

score = report.by_model["dpp4-qpisa"].sites[0].score
mask = report.by_model["dpp4-p1-pro-ala"].pattern_mask
fraction = report.assay_fractions[0]
rows = report.summary_rows()  # one endpoint and status per method
payload = report.to_dict()    # JSON-compatible, schema_version 1

EnzymePredictor.available() lists supported enzymes. FAP() similarly reports its N-terminal dipeptidyl and internal endopeptidyl motifs plus the separate Keane assay estimate. For an enzyme without an assay calibration, the result has no invented fraction; the summary says so. If an enzyme requires an explicitly active form, pass enzyme_state="active" (for example, CPB2). Optional model runtimes are included only with include_optional=True.

The older predict(peptide, minutes=10) spelling still selects a 10-minute source-assay exposure. New code should use AssayConditions when changing exposure time or enzyme concentration. These parameters do not change motif matches or native cleavage scores. The minutes argument of ITCellCleavage(...) has a different meaning: it selects a published specificity profile and does not request a new exposure prediction.

The enzymes namespace gives every default-cataloged enzyme an object, including those without a dedicated class. Its attribute names remove spaces:

from mhctools import enzymes

ace = enzymes.ACE
cats = enzymes.CathepsinS
print(cats.describe()["methods"])

For configuration files and command-line integrations, the explicit string registry resolves named classes. Other cataloged enzymes fall back to the same generic predictor:

from mhctools import DPP4, FAP, EnzymePredictor, enzyme_predictors

assert enzyme_predictors["dpp4"] is DPP4
assert isinstance(EnzymePredictor.from_name("fap"), FAP)
ace = EnzymePredictor.from_name("ace")

EnzymePredictor(DPP4) also accepts the class. The older EnzymePredictor("DPP4") constructor remains valid for callers of 4.0.5; new Python code should use DPP4() directly.

The fraction is parent peptide lost in one purified-enzyme assay. It is not serum half-life, a fraction of intact core epitope surviving, an IM-vaccine delivery fraction or an APC-presentation probability. For candidate windows and those different endpoints, see vaccine window exposure and vaccine core calibration.