Benchmarking partner-specific protein representations for phenotype discrimination in human allelic heterogeneity

Abstract Background Pathogenic variants within the same human gene can cause distinct Mendelian phenotypes. Partner-specific protein representations provide molecular context for studying this allelic heterogeneity. Their technical availability, phenotype-balanced coverage and incremental phenotype information require separate evaluation. We benchmarked structural/interface representations, MutPred-PPI disruption scores and sequence-based SPPIDER-seq and X-PAIR representations in a ClinVar-based cohort. The labeled-versus-sorted control removed named partner assignments by sorting each residue's raw score multiset before independent standardization. Cohort eligibility and statistical procedures were specified before phenotype testing. The combined support criterion required a positive median ΔR² with its 95% confidence interval (CI) strictly above zero. Positive ΔR² was also required in at least 70% of phenotype-ready genes. Results On a common denominator of 132 genes, MutPred-PPI scored 10,917/11,150 submitted structural rows (97.91%). Nine genes reached the representation-coverage threshold, and three met phenotype-balance criteria. Sequence-based representations expanded usable coverage. SPPIDER-seq reached the threshold in 127 genes, with 23 phenotype-balanced genes; X-PAIR reached it in 107 genes, with 15 phenotype-balanced genes. The primary cohort comprised 10 genes and 547 unique wild-type (WT) residue analysis units. X-PAIR median ΔR² was 0.00179553 (95% gene-bootstrap CI − 0.00410526 to 0.00495717), with positive ΔR² in 8/10 genes. The prespecified combined criterion was not met, and cross-model findings limited generalization. Distinct substitutions at the same WT residue received the same residue × partner vector. The contribution of this design property to the global result was not directly tested. Conclusions Sequence-based partner representations substantially expanded phenotype-balanced coverage. Named partner identity provided no robust global incremental phenotype information within the evaluated cohort under the prespecified criterion. The framework distinguishes molecular feature availability from information relevant to phenotype discrimination. Phenotype-resolved interpretation may require combining substitution-specific molecular consequences with partner context and disease-relevant biological context.