Make alleles optional in command-line predictors (default None)

[iskandr]

Motivation

A predictor built only to score explicit (peptide, allele) pairs via predict_pairs() still had to be handed a placeholder allele at construction — and for predictors that validate against the tool's -listMHC list, a valid one, because the constructor resolves alleles eagerly and raises UnsupportedAllele. In practice that forced awkward calls like NetMHCstabpan(alleles=[]).

Change

Default alleles=None (normalized to []) across the command-line predictors and their factories (NetMHCpan, NetMHC, NetMHCIIpan), so:

NetMHCstabpan().predict_pairs_dataframe(peptides, alleles, skip_unsupported=True)  # no placeholder
NetMHCpan().predict_pairs(pairs)

both work with no constructor alleles. Passing alleles still works unchanged.

The classic predict() / predict_peptides() path scores against self.alleles, and _allele_groups returns [] on empty — so with no alleles it would silently return nothing. A fail-loud _require_alleles() guard now raises there instead, preserving the loud failure the old required-argument gave for the forgot-the-alleles mistake:

ValueError: NetMHCpan42 was constructed without alleles; pass alleles=[...] to the
constructor, or use predict_pairs(peptides, alleles) to score explicit (peptide, allele) pairs.

NetMHCIIpanBase and BaseCommandlinePredictor stay required — they are internal bases that always receive alleles by keyword from their subclasses/factories, and each has a required positional (parse_output_fn) after alleles.

Scope is the command-line (BaseCommandlinePredictor) predictors, where the eager -listMHC validation makes the placeholder especially painful. BasePredictor gains the None → [] normalization (shared) but BigMHC / PRIME / MixMHC* signatures are left required for now — they take a different predict path without the shared guard.

Tests

New CI-safe cases (stub predictor, no binary): BasePredictor normalizes None/omitted/str alleles; predict_pairs works with empty constructor alleles; predict/predict_peptides raise ValueError (not silent-empty) with none. Full suite: 672 passed, 51 skipped, 2 xfailed. Also verified against the real binaries here (NetMHCstabpan() / NetMHCpan() construct, pair-score, and guard).

https://claude.ai/code/session_01SZZUQruWNznzAJSN4FpDGs