Bio Immunoinformatics Immunogenicity Scoring

Version Compatibility

Reference examples tested with: NeoFox 1.0+, pVACtools 4.1+, pandas 2.2+, numpy 1.26+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Notes specific to this skill: NeoFox annotates ~16 published features (it does not rank candidates automatically); PRIME 2.x requires MixMHCpred v3.0+ on PATH; BigMHC has separate -m el and -m im heads. ImmunoBERT is a PRESENTATION model, not an immunogenicity predictor — do not use it here. PRIME2.0 is the Cell Systems 2023 paper (the Cell Reports Medicine 2021 paper is PRIME v1). Re-verify tool versions and the supported-allele lists before scoring.

Immunogenicity Scoring

"Rank my neoantigen candidates by how likely a T cell responds" -> Annotate presentation + recognition features and order candidates within a patient; never assign an absolute immunogenicity verdict.

• Python: NeoFox to compute the published feature panel; PRIME / BigMHC -m im for recognition scores
• CLI: pVACtools aggregate-report tiering as the auditable, rule-based default ranking

The Single Most Important Modern Insight -- this is the least-solved layer; rank within a patient, never threshold

Binding/presentation is genuinely good (AUROC high-0.9s); immunogenicity is not close. Predicting whether a displayed peptide provokes a T-cell response requires knowing whether a cognate TCR exists in this patient's repertoire, whether that clone survived thymic negative selection (escaped tolerance), and whether it activates in a suppressive tumor microenvironment — none observable from sequence. Dedicated immunogenicity tools land around AUROC 0.6-0.7 on their own test sets and worse on independent data; in TESLA the dedicated in-silico immunogenicity scores correlated poorly with validated immunogenicity, while presentation strength, binding stability, abundance/expression, agretopicity, and foreignness carried the signal. Two operational rules follow. First, immunogenicity scores are calibrated within a context (a tool, an allele, often a patient's HLA), so they are legitimate for ordering one patient's candidate list and illegitimate for absolute go/no-go or cross-patient/cross-allele comparison. Second, a confident single composite number is a red flag: stacking weak, correlated, IEDB-bias-trained scores into one value launders the bias at higher apparent precision. The honest deliverable is an ordered, feature-annotated shortlist with its uncertainty stated out loud.

Why "Best Binder" Lost to "Best Quality"

The best-binder heuristic fails on a tolerance argument: a peptide that binds MHC superbly but closely resembles a self-peptide the thymus presented has had its cognate T cells deleted, so display does not help. A moderate binder that looks strikingly un-self may have a full, un-tolerized repertoire. The modern requirement is conjunctive — a useful neoantigen must be both PRESENTED (binding) AND FOREIGN enough (different from self) to have escaped tolerance. The Łuksza/Balachandran fitness model formalizes this: quality = amplitude (how much better the mutant is presented than its WT, a DAI-like term) x recognition potential R (resemblance to known immunogenic foreign epitopes). This is why agretopicity and foreignness, not raw affinity, recur in every validated analysis.

Tool Taxonomy

Decision Tree by Scenario

Annotate Features, Then Rank Within Patient

Goal: Order one patient's candidates without collapsing fragile features into a single over-trusted number.

Approach: Compute the feature panel (NeoFox), apply the non-negotiable expression/clonality filters first, then sort by presentation + abundance + quality features, keeping the features visible side by side for human curation. Cross-patient comparison is invalid.

import pandas as pd

def rank_within_patient(df, expr_col='gene_expression', vaf_col='rna_vaf'):
    '''Filter (not score) on expression/clonality first, then order by presentation,
    abundance, and quality. Returns a feature-annotated table for human curation, not
    a verdict. Scores are within-patient only - never compare across patients/alleles.'''
    keep = df[(df[expr_col] >= 1.0) & (df[vaf_col] >= 0.25)].copy()
    sort_cols = ['presentation_rank', 'gene_expression', 'agretopicity', 'foreignness']
    ascending = [True, False, False, False]
    cols = [c for c in sort_cols if c in keep.columns]
    asc = [a for c, a in zip(sort_cols, ascending) if c in keep.columns]
    return keep.sort_values(cols, ascending=asc)

Compute Agretopicity (DAI) Defensively

Goal: Use the mutant-vs-WT binding gain without falling into its two traps.

Approach: Agretopicity (ratio, IC50_WT / IC50_MT; the DAI family — Duan 2014 uses the difference form) rewards a mutant that binds while WT does not. Trap 1: an anchor-position mutation inflates it without changing the TCR-facing surface (quarantine via the Anchor tier). Trap 2: when WT binds very poorly, the denominator explodes and the ratio is dominated by prediction noise — a value of 200 on a barely-estimable WT is not 100x more meaningful than a value of 2.

def defensive_dai(df, wt='wt_ic50', mt='mt_ic50', anchor='mutation_at_anchor', wt_cap=5000):
    '''Flag anchor-inflated and denominator-unstable DAI rather than trusting the number.'''
    out = df.copy()
    out['dai'] = out[wt] / out[mt]
    out['dai_anchor_artifact'] = out[anchor]                 # surface unchanged -> DAI is artifact
    out['dai_unstable'] = out[wt] > wt_cap                   # WT barely presented -> ratio is noise
    out['dai_trustworthy'] = ~out['dai_anchor_artifact'] & ~out['dai_unstable']
    return out

Per-Method Failure Modes

Treating a score as a verdict

Trigger: "score > X means immunogenic" or comparing scores across patients. Mechanism: scores are calibrated within tool/allele/patient. Symptom: false confidence; cross-patient mis-ranking. Fix: rank within a patient; state uncertainty; never threshold absolutely.

The composite-score illusion

Trigger: summing/modeling DAI + foreignness + dissimilarity + hydrophobicity + PRIME into one number. Mechanism: components are weak, correlated (several measure "un-selfness"), and trained on ill-defined negatives. Symptom: an authoritative-looking 3-decimal number hiding fragile assumptions. Fix: keep features side by side; use auditable tiers; let a human weigh axes.