ITCell cathepsin specificity profiles

ITCell provides sequence-specific predictions for human cathepsins B and S (internal cleavage) and cathepsin H (initial N-terminal aminopeptidase trimming). The nine published 15/60/240-minute count matrices are bundled: there is no external installation or structural feature requirement.

from mhctools import ITCellCleavage, get_cleavage_model

internal = ITCellCleavage(enzyme="S", minutes=240).predict("ACDEFGHIKLMNPQRSTVWYV")
initial_trim = get_cleavage_model("itcell-cath-240").predict("MALWMRLLPLL")
for site in internal.sites:
    print(site.bond, site.score, site.reason)

The native score sums log2 ratios between the source P4-P4prime specificity profile and published human background frequencies. The released implementation adds one count and uses the maximum column count plus 20 as the shared denominator. Missing terminal flanks contribute zero. We reproduce that formula, including its unnormalized background frequencies, and retain all scores. Author-code candidate thresholds are strictly >3 for B/S and strictly >2 for H. The paper's prose describes the >3 threshold as threefold; the released code uses log2, so that interpretation is inconsistent. Use the native score and the stated source-code threshold.

These are specificity scores, not probabilities, percentages lost, kinetic rates or serum survival. The profile's minutes select the source assay observations; they do not predict degradation after that duration in a new sample. The source used 228 tetradecapeptides with recombinant human enzymes at 0.2 ug/mL, pH 6.5 and 1 mM TCEP. It substituted norleucine for methionine and did not represent cysteine. Outputs explicitly flag those contexts and absent flanks. Predictions require canonical sequence with assumed free termini.

H scores only the currently exposed first bond. B models internal specificity, not its carboxydipeptidase activity. No fragment generation, repeated trimming, uptake, cellular accessibility or enzyme activation is simulated.

The independent implementation is tested against native scores from the released Perl primitive. The author's combined cleave.pl wrapper repeats the 240-minute S profile in place of the 15/60-minute profiles; we run each matrix independently, and reported this upstream. This is not the structure-dependent 2023 PCSS/SVM cathepsin predictor.

Sources: ITCell study, author code, released matrices. The bundled count matrices and background frequencies retain the author's LGPL-2.1 license, copyright attribution, exact file hashes and archive identity in mhctools/data/itcell_profiles.json and ITCELL_LICENSE.txt.