ProteoNexus: an integrative database to characterize genetic architecture, estimate mediation effects, and construct and evaluate prediction models of the plasma proteome
Kaixin Shao, Zixin Luo, Peng Huang, Sheng Yang · 2025 · 8 citationsRead the paper
Proteins are biological effectors that mediate the effects of exposures on diseases and serve as predictors for constructing high-performance disease prediction models. However, an integrative, sex-specific proteomic resource using a biobank-scale dataset remains unavailable. Here, we introduce ProteoNexus, a database featuring a standardized best-practice pipeline integrating protein pQTLs mapping, mediation analysis, and risk prediction. Following stringent quality control, ProteoNexus comprises three categories of exposures: 129 measurement-based variables, 54 environmental variables, 1 251 123 single-nucleotide polymorphisms (SNPs), and 57 incident diseases among 33 325 European participants. ProteoNexus identifies 16 998 putative causal pQTLs, of which 5 979 are cis-pQTLs and 11 019 are trans-pQTLs in the combined-sex dataset, while 9 464 and 7 832 pQTLs were identified in the female and male datasets, respectively. Using a two-step screening strategy, ProteoNexus identifies 308 325, 144 975, and 1 336 significant pathways caused by measurement-based variables, environmental variables, and SNPs, respectively, followed by enrichment analysis of proteins associated with these exposures. With 21 optimized parameters for four machine learning algorithms, ProteoNexus provides an online analysis module that enables users to analyze their own proteomic data. Users can search for results by protein, reported disease, ICD-10 code, or exposure, with accompanying summary statistics for each query. ProteoNexus is freely accessible at https://www.proteonexus.com/.
What the full text shows
Working prototype- Identified putative causal pQTLs in the combined-sex dataset. — 16,998 total (5,979 cis-pQTLs and 11,019 trans-pQTLs) (33,325 European-ancestry participants)
- Identified significant mediation pathways from measurement-based, environmental, and genetic variables to diseases. — 308,325 M-P-D, 144,975 E-P-D, and 1,336 G-P-D pathways (Two-step screening strategy using FDR < 0.05 or FWER < 0.05)
- Developed disease prediction models using baseline proteomic profiles. — AUC > 0.75 for 26 of 57 diseases (Four ML algorithms (PLR, XGBoost, LightGBM, MLP) optimized via TPE)
- XGBoost was the top-performing algorithm for the most outcomes. — Best for 26 diseases (combined), 27 (female), and 24 (male) (Across sex-specific and combined datasets)
- Demonstrated high predictive accuracy for Type II Diabetes (T2D). — AUC = 0.884 (Using XGBoost model)
- Setup
- 33,325 European-ancestry participants from UK Biobank (17,988 female, 15,337 male) with 2,919 proteins and 1,251,123 SNPs.
- How
- Integrated a pipeline of linear regression for pQTL mapping, SuSiE for fine-mapping, and the medflex R package for mediation analysis. Four machine learning classifiers were trained and optimized using five-fold cross-validation and the Optuna framework.
- Limitations
- Restricted to European-ancestry participants · Mediation analyses are not formal causal inferences · Used linear regression instead of linear mixed models for pQTL mapping · Incident outcomes analyzed with logistic regression rather than Cox proportional hazards models
Summarized by Gemma 4 from the open-access full text · 8 Oct 2026. Read it