PhageScout sequence profiles

The name refers to phage display used to scout protease sequence preferences. It is not a bacteriophage vaccine platform, a pharmacokinetic model, or an estimate of what fraction of a vaccine dose reaches APCs.

PhageScout supplies separate native recognition scores for human neutrophil elastase (ELANE) and human cathepsin G (CTSG). These are extracellular and inflammatory enzyme candidates under active-enzyme exposure. A sequence score does not establish enzyme presence, activation, concentration, accessibility, inhibition, or degradation in serum.

The adapter independently implements the scoring described in the 2026 paper, using ten unmodified matrices and significant-peptide profiles from Zenodo 21387981. The deposited data are CC BY 4.0. Attribution, the license, exact source-file SHA-256 identities and a checksum for the bundled JSON container accompany the package. No author notebook code is included; there is no external inference runtime to install.

Select a native model

from mhctools import PhageScout

result = PhageScout(enzyme="ELANE", profile="pwm-relaxed-aligned").predict("MAACVYTLA")
for site in result.sites:
    print(site.bond, site.score, site.reason)

mhctools cleavage --sequence MAACVYTLA --model phagescout-elane-pwm-relaxed-aligned
mhctools cleavage --sequence AAFLDAAFF --model phagescout-ctsg-peptide-relaxed-unaligned

ELANE model	CTSG model	Native calculation
phagescout-elane-pwm-deseq2	phagescout-ctsg-pwm-deseq2	Mean of complete five-mer sums of DESeq2-derived positional weights
phagescout-elane-pwm-relaxed-unaligned	phagescout-ctsg-pwm-relaxed-unaligned	Mean of complete five-mer sums of significant-peptide log2 enrichment weights
phagescout-elane-pwm-relaxed-aligned	phagescout-ctsg-pwm-relaxed-aligned	Sum of P5–P4′ positional log2 enrichment weights
phagescout-elane-peptide-relaxed-unaligned	phagescout-ctsg-peptide-relaxed-unaligned	Mean log2 fold change of matching released five-mers spanning a bond
phagescout-elane-peptide-relaxed-aligned	phagescout-ctsg-peptide-relaxed-aligned	Log2 fold change of the first matching released aligned wildcard profile
phagescout-elane-peptide-deseq2	phagescout-ctsg-peptide-deseq2	Mean DESeq2 log2 fold change of matching released five-mers; optional full tables

Optional full DESeq2 peptide profiles

mhctools fetch phagescout
mhctools ls phagescout --json
mhctools cleavage --sequence AAACD --model phagescout-elane-peptide-deseq2

Each enzyme's full table contains 652,509 distinct five-mers, including negative DESeq2 log2 fold changes. These are peptide lookups, separate from the DESeq2 PWM summaries. The two unmodified files total about 32.6 MB and are excluded from the package. Fetching streams the fixed Zenodo record 21387981, checks file sizes and SHA-256, and atomically installs both tables with source provenance, attribution and the CC BY 4.0 license. A failed download leaves no partial installation. An invalid existing destination must be moved aside before fetching again.

The managed data directory can be selected with MHCTOOLS_DATA_DIR or mhctools fetch phagescout --data-dir PATH. For existing copies of the exact deposited files, set PHAGESCOUT_HOME to their directory, or pass PhageScout(enzyme="CTSG", profile="peptide-deseq2", profile_dir=PATH). Explicit --data-dir overrides PHAGESCOUT_HOME for fetch/inventory; inference uses PHAGESCOUT_HOME or MHCTOOLS_DATA_DIR. The selected file is verified before inference. Invalid explicit paths raise an actionable error. mhctools cleavage --list-models discovers both models without reading their tables; only explicit selection loads the lookup.

Bond b splits sequence[:b] | sequence[b:]. The aligned ninth-position window puts residue b at column 5, the inferred P1 anchor; the whole window is P5–P4′. DECIPHER alignment of enriched phage five-mers infers this anchor. It does not experimentally locate a scissile bond. An unaligned five-mer can contribute to four candidate bonds; the adapter retains that ambiguity.

Five-mer modes average up to four complete windows spanning each bond. Aligned PWM scoring omits missing terminal flanks, following the native calculation, and explains that omission in the site's reason. Aligned peptide-profile matching requires all nine positions. Dashes in the released profiles are wildcards, including internal gaps; when multiple rows match, the first source-ordered row supplies the score. Row order is preserved.

Only canonical uppercase L-peptide sequences with assumed free termini are supported. Modified or unknown terminal chemistry abstains. The computational context requirements are not absolute enzyme length cutoffs. Enzyme exposure, subsequent fragments and repeated cleavage are not simulated.

Missing scores and interpretation

Every available native score is retained, including negative and small PWM scores. No threshold, min-max transformation, cross-model average or peptide loss percentage is added. Significant-peptide profiles are sparse: no match means unassessed, not zero or resistance. The result's conditions["unassessed_bonds"] JSON object gives a reason for each omitted bond; an entirely unassessed input also has unsupported_reason. API, CLI and batch epitope overlays use the same canonical bond coordinates.

The assay screened immobilized randomized five-mer phages against active elastase (5 nM, 1.5 hours) or cathepsin G (10 nM, 6 hours), with FLAG-eluted controls. These times describe the source assay. They do not predict a new peptide's half-life. Transfer from phage display to free vaccine peptides, cellular processing and whole-serum survival remains unvalidated.

Source reproduction and remaining models

Offline fixtures retain the authors' deposited scores on mature alpha-1 antitrypsin (P01009, 393 bonds) and 14-3-3 theta (P27348, 244 bonds), with matching UniProt sequence identities. All ten models reproduce their source values and missing scores: 3,822 numeric PWM scores and 36 numeric significant-peptide-profile scores, plus every unassessed source position. The optional full tables additionally reproduce both enzymes' deposited pep_deseq2 values and missing masks on these proteins. CI fetches the real files and runs that comparison without skips. These are implementation-conformance examples, not independent accuracy validation or experimentally measured non-cleavages.

The optional distance-weighted PWM variant is excluded: the source notebook multiplies the accumulated sum at successive positions, rather than weighting each residue independently. Substituting an ordinary distance-weighted sum would change its released scores.

All four deposited phage-only XGBoost boosters load in the official R XGBoost 3.2.1.1 runtime. They require 18 features: raw, within-mature-protein min-max, and sample-standard-deviation Z scores for six sequence features. Their saved attributes do not contain the training median vectors used to impute sparse profile matches. New-sequence classifier inference needs that preprocessing state and the exact mature-protein normalization context. The adapter exposes the reproduced native sequence scores while that work remains tracked in known gaps. Structural classifiers additionally require actual applicable structure features. Balanced-classifier output and normalized landscape scores are not calibrated serum-loss probabilities.