Add XSignal zero-shot baseline

[Billwanttobetop]

Summary

This PR adds XSignal, a zero-shot evidence-integration baseline for ProteinGym substitution assays. XSignal derives 14 mutation-scoring channels from ProteinGym MSAs, MSA weights, precomputed AlphaFold2 structures, and fixed amino-acid physicochemical descriptors. Each channel is transformed to assay-wise percentile ranks and the final score is an equal-weight average.

The final scorer reads only variant-key columns from DMS assay CSVs (mutant, mutated_sequence). It does not read DMS labels, binary labels, supervised folds, public baseline score packages, or official merged-score files during final scoring. XSignal uses no pretrained protein language model weights; it does use precomputed AlphaFold2 structures and ten self-generated intermediate evidence score directories.

Score Package

Audited release: https://github.com/Billwanttobetop/ProteinGym/releases/tag/xsignal-proteingym-scores-20260723-audited

Score asset: XSignal_scores_substitutions_20260723.zip

• 217 ProteinGym substitution assays
• 2,465,767 variants
• columns: mutant, mutated_sequence, XSignal_score
• SHA256: 0e3c9fa348d83b60e23ee3f57a7c1d517ea1109cc277556bdc8598762538b367

Evaluation

Official ProteinGym wrapper metrics on the full substitution benchmark:

Metric	Score
Spearman	0.476
AUC	0.761
MCC	0.369
NDCG	0.778
Top recall	0.213

A full final-fusion reproduction regenerated 217 assay score files / 2,465,767 variants from the archived upstream XSignal sensor outputs plus ProteinGym MSA weights and AF2 structures, matching the frozen internal candidate PureGraphGEContactPair_v1 with max_abs_diff = 0.0. The release also includes a compact audit/provenance bundle.

Scope Boundary

This submission is for the frozen XSignal substitution score package only. Public RefSeq MSA expansion experiments and private/non-public datasets are not included.