The authors introduce absorption as a metric quantifying how much of a polygenic risk score's (PRS) predictive contribution is accounted for by an out-of-fold proteomic risk score (ProRS). Absorption is estimated in UK Biobank participants, with sample sizes reported between 9,434 and 9,820 depending on the analysis. Proteomic co-expression modules were constructed from plasma proteomes measured in 41,358 disease-free participants.
This framework was used to directly compare predictive information contained in PRS and ProRS across multiple neurodegenerative diseases, and to evaluate how germline signal maps onto plasma protein networks.
A central finding is that absorption does not track SNP-based heritability (h2). The authors report matched SNP h2 values for three diseases: Alzheimer’s with APOE (h2 = 0.068), Alzheimer’s without APOE (h2 = 0.061), and Parkinson’s disease (h2 = 0.069). Despite similar SNP h2, absorption differed markedly: Alzheimer’s with APOE absorbed 0.73, Alzheimer’s without APOE absorbed 0.39, and Parkinson’s absorbed 0.19. Amyotrophic lateral sclerosis (ALS) showed absorption of 0.22.
Importantly, the analyses identified a residual genetic signal in all four conditions, indicating that proteomic risk scores did not fully capture the predictive information contained in polygenic risk scores. Thus, the ProRS and PRS provide overlapping but non-redundant information for pre-symptomatic stratification.
The proteins associated with a PRS and the proteins selected by a ProRS overlapped no more than expected by chance in most cases. Convergence between genetic and proteomic selections occurred primarily in contexts dominated by APOE. This suggests that, outside dominant locus effects, mapping from germline genetic architecture to circulating protein predictors is incomplete and often disparate.
The authors partitioned the plasma proteome into 33 co-expression modules. Germline signal tended to concentrate in particular modules rather than being diffusely distributed across the proteome. The summary (principal) components of seven modules were associated with the Alzheimer’s score that included APOE, whereas no module summary components were associated with the Parkinson’s score.
This modular organization implies that genetic risk projects onto specific protein subnetworks. Identifying which modules carry germline signal can help explain where a ProRS might succeed or fail to recapitulate PRS-derived risk.
The lysosomal module M32 is notable because it encodes a lysosomal genetic architecture associated with Parkinson’s disease. Despite this, the module shows an unexpected pattern: 89% of its members were associated with Alzheimer’s germline risk, yet none of M32’s proteins were used by the proteomic score. This highlights a disconnect between modules that harbor germline associations and the proteins selected into predictive proteomic models.
The M32 example exemplifies how disease-relevant genetic architecture can be concentrated in modules that are not exploited by ProRS approaches, and underscores the complex mapping from germline variation to circulating protein predictors.
To evaluate module generalizability, the authors rebuilt the co-expression modules in the All of Us cohort. The cross-cohort comparison produced an adjusted Rand index of 0.582 against a within-cohort attainable value of 0.686. Additionally, 29 of 33 modules remained coherent above a permutation null distribution. These results indicate predictable cross-cohort transferability driven by module coherence in the discovery cohort and support reuse of the module partition as a fixed reference dictionary for plasma proteome analyses.
Predictability of transferability suggests that module-based approaches may be robust across large diverse cohorts, though absolute concordance is imperfect.
The study analyzed de-identified secondary data from the UK Biobank and the All of Us Research Program; both sources require authorized access applications to retrieve underlying data. The authors provide a code and resource link to their repository. Ethical approvals and participant informed consent were reported for both resources; the authors declared no competing interests.