Additional protease candidates¶
Primary-source and author-artifact audit, checked 2026-10-06. Bundled PhageScout sequence profiles now reproduce native scores; the other routes below remain integration candidates. Known gaps tracks their implementation and shared validation dependencies.
The remaining gaps have different causes: some predictors publish weights but require unreproduced features, some describe neural models without supplying verified inference parameters, and many experiments supply useful recognition preferences without a transferable quantitative predictor. Recognition flags can be added under their experimental scope without calling them calibrated probabilities or serum-survival estimates.
Released predictor assets¶
PhageScout sequence profiles are available now. Its 2026 paper describes neutrophil elastase and cathepsin G specificity scoring. Zenodo 21387981 declares CC BY 4.0 and supplies PWMs, peptide profiles, reference scores and trained R/XGBoost objects, including phage-only models. The earlier audit missed this separate release. The GitHub revision has no project license; its code and the licensed deposited assets require separate treatment.
The independently implemented sequence-only route reproduces deposited native scores. Five-mer substrate enrichment, inferred aligned-nine-mer anchors and known MEROPS cuts have different evidentiary status. The author notebook contains raw, normalized and structural features; exact anchor, terminal handling, normalization population and trained-model feature order remain requirements for new-sequence classifier inference. Structural classifiers require actual applicable features. Neither a normalized landscape score nor a classifier trained on balanced sites is a serum-loss percentage.
Downloaded files matched the author's MD5 checksums. SHA-256 identities:
| Asset | SHA-256 |
|---|---|
elastase_relaxed_aligned_pwm.txt |
5063e9e26da2c0bf21e3c4b2b668838e7c0bc3856e6a6ab6d05f753b453d2993 |
cathepsin G_relaxed_aligned_pwm.txt |
16302980d91388fe826ca11eb66d9043f09976e7123db788eba7c81a2efd7481 |
elastase_phage_balanced_XGB.rds |
713d4b42cfb7443061b93dba51bb905b7afefac78a910fd69256ffdd91df2e28 |
cathepsin_G_phage_balanced_XGB.rds |
f98247ce7b8669f90e4ac7c168091e9b18e7ad90e009c2d5daf700701a3d9a28 |
Ten native sequence profiles reproduce deposited values and missing scores. All four phage-only boosters load in the official R XGBoost 3.2.1.1 runtime; their 18 feature names are retained, but no training-imputation medians are saved in their attributes. Automatic new-sequence classifier inference and independent accuracy remain unverified.
Other trained-model routes remain distinct. Tusar et al. CatS/L/B SVMs supply model files but require a sequence/structure feature pipeline. ProsperousPlus supplies pretrained protease directories but still has no project license at the audited revision. CatS/L/B octamer neural ensembles describe sequence-property/JMP models; no released trained parameters or portable inference runtime were verified. Their CC BY 4.0 Figshare deposit contains three observed-product CSVs, not the neural weights.
UniZyme has MIT code and a CC BY 4.0 data/weight archive. Its structure/energy input pipeline and application to free peptide conformations still need an independent applicability audit. Availability of that archive alone does not establish runnable predictions for every enzyme.
Experimental recognition evidence¶
| Enzyme | Primary evidence worth curating | Scope to preserve |
|---|---|---|
| Human CatL | PICS, 845 mapped sites, DIPPS, cathepsin MSP-MS comparisons | Aromatic P2 is a preference, not a universal requirement. Retain library, pH, time, reducing conditions and source erratum. |
| Human AEP/legumain | pH-specific DIPPS datasets, counter-selection libraries, peptidase/ligase experiments | Asn/Asp acceptance changes with pH and reaction context. Keep hydrolysis, ligation, activation and human/plant enzymes distinct. |
| Human thrombin | extended phage/linker specificity, motif/loop validation, high-throughput profiling | Include cooperative alternatives: P2 Pro is preferred but not universally required. Actual thrombin activity, access and exosite/cofactor context matter. |
| Human plasmin | P1-Lys-fixed positional library, subsite cooperativity, P1-Arg-fixed plasma-protease profiling | Keep fixed-P1 assumptions, reporter chemistry, prime-side gaps and noncanonical residues explicit. Inhibitor affinities are not substrate cleavage rates. |
| Human plasma kallikrein KLKB1 | fluorogenic microarrays, human-plasma activation comparisons, separately identified human MSP-MS comparator | Distinguish KLKB1 from tissue KLK1/KLK6 and the parasite enzyme in the comparator study. Preserve contact activation, inhibitors and sample preparation. |
| Human ELANE/PRTN3/CTSG | ELANE/PRTN3 extended preferences, CTSG dual specificity, protein-substrate tests | Keep each enzyme distinct, including CTSG secondary Lys specificity, tentative phage anchors, detergent effects and protein accessibility. |
The practical first additions are conditional recognition flags and exact observed-bond imports, followed by published profile scores where the native calculation can be reproduced. This ordering is an implementation assessment, not an accuracy ranking. Model and source-sequence overlap must be audited before calling any comparison independent validation. Enzyme activation and exposure are separate from sequence specificity; purified-enzyme preferences cannot determine whole-serum survival or productive antigen presentation.