DSL extensions: logistic, Column, WT, peptide properties

[iskandr]

Intent

Extend the ranking/filtering DSL to support the signals needed for Vaxrank-style composite scoring and general immunogenicity/manufacturability analysis. After this PR, users can express:

score = (
    0.4 * Affinity["netmhcpan"].logistic(350, 150)
    + 0.3 * Presentation["mhcflurry"].score
    - 0.1 * Column("cysteine_count")
    - 0.1 * abs(Column("charge"))
    + 0.1 * (Affinity.score - WT(Affinity).score)  # differential binding
)

Changes

1. .logistic(midpoint, width) on Expr

Logistic sigmoid transform: 1 / (1 + exp((x - midpoint) / width)). Vaxrank uses midpoint=350, width=150 for IC50→score conversion. Same pattern as existing .norm().

2. Column("name") — arbitrary DataFrame column access

New Expr subclass that reads any column from the group DataFrame. This is the key unlock: external signals (read counts, expression, peptide properties, custom annotations) become first-class ranking signals without special-casing each one.

Column("hydrophobicity") >= -0.5
Column("n_alt_reads").sqrt()

CLI: column(name) in filter/ranking strings. Errors with available columns when column is missing.

3. WT(accessor) — wildtype comparison wrapper

Wraps a KindAccessor to read wildtype prediction columns (wt_value, wt_score, wt_percentile_rank) instead of the mutant columns. Avoids duplicating every field on KindAccessor.

WT(Affinity).value                        # WT IC50
WT(Affinity["netmhcpan"]).score           # qualified WT
Affinity.score - WT(Affinity).score       # differential

Works with method qualification — same prediction_method_name filtering as the mutant side. When WT columns don't exist (non-variant inputs), evaluates to NaN.

4. topiary.properties — peptide property columns

New module computing amino acid properties directly on the peptide column using vectorized pandas string operations (fast enough for proteome-scale):

from topiary.properties import add_peptide_properties

df = add_peptide_properties(df, groups=["manufacturability"])

Named groups:

• "core": charge, hydrophobicity, aromaticity, molecular_weight
• "manufacturability": core + cysteine_count, instability_index, max_7mer_hydrophobicity, cterm_7mer_hydrophobicity, difficult_nterm, difficult_cterm, asp_pro_bonds
• "immunogenicity": core + tcr_charge, tcr_aromaticity, tcr_hydrophobicity

Computation uses pandas str.count(), str[n], apply() with lookup dicts — no external dependencies.

Properties accessible in DSL via Column("charge"), Column("cysteine_count"), etc.

Not in this PR

• add_wildtype_predictions() function (the step that actually runs WT predictions and populates the wt_* columns) — separate PR, depends on predictor refactoring
• Evaluator performance refactor (vectorized filter/sort) — separate PR

[coveralls]

coverage: 85.885% (+0.4%) from 85.441% — dsl-extensions into master