Absorption and Co-expression Modules Show Where Polygenic and Proteomic Risk Scores Diverge in Neurodegenerative Diseases

Polygenic and proteomic risk scores are both proposed for pre-symptomatic stratification, yet the extent to which they provide overlapping or complementary information has not been measured across neurodegenerative disease. To quantify this overlap, we define absorption as the fraction of a polygenic score's predictive contribution accounted for by an out-of-fold proteomic score and estimate it among 9,434 to 9,820 UK Biobank participants. Absorption did not track h2, as Alzheimer's with APOE, Alzheimer's without APOE and Parkinson's carried matched SNP h2 of 0.068, 0.061 and 0.069 yet absorbed 0.73, 0.39 and 0.19, with amyotrophic lateral sclerosis at 0.22. Residual genetic signal remained in all four, indicating that proteomic risk scores did not fully capture the predictive information contained in polygenic risk scores. The proteins associated with a polygenic score and the proteins a proteomic score selects overlap no more often than chance, converging only where APOE dominates. Across 33 plasma co-expression modules built in 41,358 disease-free participants, germline signal concentrates in modules rather than spreading, with the summary component of seven modules associated with the score for Alzheimer's with APOE and none for Parkinson's, even though Parkinson's carries the lysosomal genetic architecture that the lysosomal module M32 encodes. This module has 89% of its members associated with AD germline risk yet none was used by the proteomic score. Rebuilding the co-expression modules in All of Us gave an adjusted Rand index of 0.582 against the 0.686 attainable within that cohort, and 29 of 33 modules stayed together above a permutation null. Cross-cohort transferability was predictable from module coherence in the discovery cohort, supporting the reuse of this module partition as a fixed reference dictionary.