Xilin Jiang

Predicting disease risk from prior diagnoses is fundamental to clinical decision-making, particularly during health emergencies such as the COVID-19 pandemic, when individuals with long-term conditions may be disproportionately vulnerable to adverse outcomes. Despite intense interest in developing models to predict disease risk from prior diagnoses ^1–3 , most prediction models do not estimate effects of each prior diagnosis on disease risk conditional on other diagnoses, limiting interpretability and clinical utility. We developed the Comorbidity Risk Score (CRS), trained on 13 million individuals (age 40-69) from linked electronic health record (EHR) datasets of the entire population of England, to predict COVID-19 hospitalisation and 87 other disease outcomes. CRS was trained at close to saturated sample size and precisely estimated the effects of 212 prior diagnoses on the 88 disease outcomes, conditional on all other prior diagnoses. Correlations of CRS effect sizes across outcomes (e.g. 0.76 for myocardial infarction vs. hyperlipidaemia) matched the corresponding genetic correlations (e.g. 0.79 for myocardial infarction vs. hyperlipidaemia), confirming that comorbidity architectures capture disease aetiology. On average, CRS identified 5% of the population with 3.4-fold higher disease risk, including myocardial infarction (4.4-fold), lung cancer (6.5-fold), and COVID-19 hospitalisation (6.3-fold). Using prior diagnoses alone, CRS outperformed state-of-the-art clinical COVID-19 models ⁴ . Furthermore, CRS ( N =13 million) substantially outperformed state-of-the-art AI ¹ ( N =0.5 million) and linear ³ ( N =0.5 million) models in predicting disease risk, suggesting that training sample size outweighs model complexity. CRS attained near-perfect transferability across self-reported ethnicities (e.g., Black vs. White: AUROC ratio = 97.3%). Finally, CRS distinguished independently predictive comorbidities from indirect associations, e.g., lipid metabolism disorder was a strong predictor of myocardial infarction risk but not ischaemic stroke, after conditioning on other prior diagnoses. In conclusion, CRS provides a comprehensive resource for understanding the impact of comorbidities on COVID-19 and other future diseases, revealing disease aetiology while enabling powerful prediction of disease risk.

Proteomics holds great promise for identifying potentially druggable effectors of common diseases, yet its application at population-scale across diverse ancestries, remains challenging. Here, we developed genetic imputation models for 2,594 plasma proteins using proteomic and genetic data from 54,219 UK Biobank participants, validating their performance across multiple ancestry groups and in an independent cohort. Plasma proteomes were then imputed for over 640,000 participants in the UK Biobank and the All of Us Research Program. To assess its aetiological value at population-scale, a further proteome-wide association study of cardiovascular diseases was performed across six genetic ancestries. We identified ∼9000 protein-disease associations across 89 cardiovascular conditions (PheCodes), the majority of which show consistent effects across ancestries and biobanks, with many comprising known targets of drugs either approved or under development. The associations reveal both shared and distinct proteomic signatures across cardiovascular conditions and defined clusters of distinct pathophysiology with shared underlying molecular pathways. Integration of data on tissue specificity and single-cell transcriptomics prioritised liver-derived proteins in circulation as candidate effectors of coronary artery disease, highlighting inter-alpha-trypsin inhibitor heavy chain H4 (ITIH4) as a putative effector. Using a liver-targeted CRISPR gene-editing platform, we show that in vivo disruption of ITIH4 reduces plasma cholesterol and pro-atherogenic lipid species in a preclinical model, consistent with a causal role in cardiovascular disease. Our study enables study of large-scale proteomics in diverse populations, provides a systematic map of protein associations of cardiovascular diseases, and demonstrates the utility of genetically imputed proteomes for target discovery and experimental validation. To facilitate proteomic analyses for the research community, the resultant models and association results have been made freely available through the OmicsPred platform.

Electronic health records (EHRs) provide rich longitudinal disease histories, but existing methods for analysing these data typically treat diseases in isolation¹ and rarely integrate germline genetics. Here we present ALADYNOULLI, a Bayesian generative framework that jointly models longitudinal EHR diagnoses, age and polygenic risk to recover latent time-varying disease signatures and patient-specific signature loadings; the model is formulated as a mixture of probabilities rather than a probability of a mixture², correctly accommodating simultaneous and chronic conditions. Applied to three independent biobanks (UK Biobank³, Mass General Brigham⁴ and All of Us; total n > 683,000) spanning up to 52 years of follow-up and 348 diseases, the model recovers 21 replicable signatures with high cross-cohort composition preservation (median of 80%) and reveals biological subtypes within diagnostic categories (Cohen's d up to 4.25; P ≤ 1 × 10^-8 for 95% of comparisons). Signatures are concordant with established disease biology: carriers of familial hypercholesterolaemia⁵ enrich in the cardiovascular signature; carriers of clonal haematopoiesis of indeterminate potential⁶ in the inflammation signature; and a rare variant burden in LDLR, TTN and BRCA2 (refs. ^7,8) aligns with disease specificities. A signature-based genome-wide association study identifies 151 genome-wide significant loci including cardiovascular associations missed by single-trait analyses. An explicit likelihood enables inverse probability weighting for selection bias⁹ while preserving biological signal. For disease prediction, ALADYNOULLI outperforms Pooled Cohort Equation (PCE), PREVENT and Gail at 1-year and 10-year horizons; disease-level (PheCode) predictions complement code-level foundation models such as Delphi-2M (ref. ¹⁰).

Proteome-wide association studies (PWAS) typically link genetically predicted protein levels to disease using cis-pQTLs, which can be limited by low cis-heritability for disease-critical genes under negative selection and by tagging due to co-regulation among nearby genes. Trans-pQTLs provide complementary information when large sample sizes are available to detect weak polygenic effects, enabling associations between trans-predicted protein levels and disease. We developed PolyPWAS, a functionally informed, summary statistics-based framework for associating both cis- and trans-predicted protein levels to disease. PolyPWAS integrates 96 functional annotations with proteome-wide pleiotropy to improve protein prediction, while correcting for PCs of predicted protein levels to limit tagging effects. We applied PolyPWAS to 2.8K plasma proteins measured in 34K UKB-PPP participants, analyzing GWAS summary statistics for 88 diseases and complex traits (average N=336K). Trans-predicted protein levels explained 21% of disease heritability (vs. 9.6% for cis-predicted protein levels), leveraging a 24% relative improvement in trans-prediction accuracy from functional priors. Trans-PWAS identified more significant protein-disease associations (and more conditionally significant associations) than cis-PWAS. Cis and trans associations showed only modest excess overlap (1.18, 95% CI: 1.11-1.26). Accordingly, combining evidence from cis and trans associations improved disease gene prioritization evaluated using gene sets from rare variant association studies (+11% relative improvement) and PoPS (+7.0% relative improvement) relative to cis-only approaches. PWAS associations to disease replicated across protein level cohorts, with strong UKB-PPP/deCODE concordance after adjusting for cohort-specific prediction accuracy. We provide examples where trans-regulatory effects link multiple disease-critical genes, underscoring the importance of integrating cis- and trans-regulatory effects to map protein-mediated disease biology.

Proteome-wide association studies (PWAS) typically link genetically predicted protein levels to disease using cis-pQTLs, which can be limited by low cis-heritability for disease-critical genes under negative selection and by tagging due to co-regulation among nearby genes. Trans-pQTLs provide complementary information when large sample sizes are available to detect weak polygenic effects, enabling associations between trans-predicted protein levels and disease. We developed PolyPWAS, a functionally informed, summary statistics-based framework for associating both cis- and trans-predicted protein levels to disease. PolyPWAS integrates 96 functional annotations with proteome-wide pleiotropy to improve protein prediction, while correcting for PCs of predicted protein levels to limit tagging effects. We applied PolyPWAS to 2.8K plasma proteins measured in 34K UKB-PPP participants, analyzing GWAS summary statistics for 88 diseases and complex traits (average N=336K). Trans-predicted protein levels explained 21% of disease heritability (vs. 9.6% for cis-predicted protein levels), leveraging a 24% relative improvement in trans-prediction accuracy from functional priors. Trans-PWAS identified more significant protein-disease associations (and more conditionally significant associations) than cis-PWAS. Cis and trans associations showed only modest excess overlap (1.18, 95% CI: 1.11-1.26). Accordingly, combining evidence from cis and trans associations improved disease gene prioritization evaluated using gene sets from rare variant association studies (+11% relative improvement) and PoPS (+7.0% relative improvement) relative to cis-only approaches. PWAS associations to disease replicated across protein level cohorts, with strong UKB-PPP/deCODE concordance after adjusting for cohort-specific prediction accuracy. We provide examples where trans-regulatory effects link multiple disease-critical genes, underscoring the importance of integrating cis- and trans-regulatory effects to map protein-mediated disease biology.

Background and Aims: Clinical biomarkers, nuclear magnetic resonance (NMR) metabolomics biomarker scores, and polygenic risk scores (PRS) have shown promise for improving cardiovascular disease (CVD) prediction but have not yet been evaluated in the context of current prediction models (SCORE2) and ESC recommendations for 10-year prediction of fatal and non-fatal CVD. Methods: NMR metabolomic biomarker scores were constructed and compared to clinical biomarkers, PRS and SCORE2 in 297 463 UK Biobank participants (8919 incident CVD cases) aged 40–69 without previous CVD, diabetes, or lipid-lowering treatment. Improvement in risk discrimination when added to SCORE2 was assessed using Harrel’s C-index. Improvement in risk stratification following ESC guideline risk thresholds was assessed using categorical net reclassification. Population modelling was subsequently applied to estimate the impact on CVD prevention if applied at scale. Results: Risk discrimination provided by SCORE2 (C-index: 0.719) improved when 11 clinical biomarkers (ΔC-index: 0.014 [0.012–0.015]), NMR metabolomic biomarker scores (ΔC-index: 0.010 [0.009–0.012]) and PRSs (ΔC-index 0.009; [0.008–0.011]) were added individually. The combination of 11 clinical biomarkers, NMR metabolomic biomarker scores, and PRSs yielded the largest improvement risk discrimination, with ΔC-index 0.024 (0.022–0.027). Concomitant improvements in risk stratification were observed in categorical net reclassification index, with net case reclassification of 16.66% (15.50%–17.81%). Modelling suggested that addition of these biomarkers to SCORE2 for targeted risk reclassification would increase the number of CVD events prevented per 100 000 screened from 229 to 413 (ΔCVDprevented: 184 [174–194]) while essentially maintaining the number of statins prescribed per CVD event prevented. Conclusions: Combining NMR metabolomic, polygenic, and clinical biomarkers with SCORE2 enhanced prediction of first-onset CVD and could have substantial population health benefit if applied at scale.

Our understanding of the dependence of the genetic and environmental architecture of common diseases on age is incomplete. Here, we use longitudinal data to separate age-dependent genetic and environmental variance across complex traits and diseases. When applying our approach to 16 UK Biobank quantitative traits (average N=180K), we found environmental variance accumulates with age, leading to decreasing heritability with age, which follows an exposure accumulation model that is distinct from gene-age interaction model. Heritability decreases with age for 5 traits including systolic blood pressure and lung functon(FEV1/FVC), with an average change of −17.8% with age per 10-years. We demonstrated that majority of the decreasing heritability comes from exposure accumulation, instead of gene-age interaction, using longitudinal phenotypic correlations. For diseases, environmental variance in disease liability also accumulates with age for 5 of 9 diseases, with an average decrease in heritability of −18.1% per 10 years. A liability-threshold model with exposure accumulation explains 86% of decreasing PRS prediction R ² with age in 9 UK Biobank complex diseases. Finally, we show that both genetic and non-genetic predictors have decreasing prediction accuracy with age in 61 predictor-disease pairs. Taken together, our results demonstrate that the age-dependent liability- threshold model with exposure accumulation is a general disease model, suggesting ascertaining younger cohorts is more powerful for training both genetic and non-genetic risk prediction models.

Variance quantitative trait loci (vQTLs), which capture genetic contributions to phenotypic variability, remain underexplored in proteomic studies, particularly across diverse ancestries. We systematically mapped cis -vQTLs for 2,923 plasma proteins in 52,706 UK Biobank participants of European (EUR, N = 45,486), African (AFR, N = 1,336), and Central/South Asian (CSA, N = 934) ancestries, identifying 2,162 vQTLs (P _VE < 5 x 10 ^-8 ) for 781 proteins. We identified ancestry-specific and shared cis -vQTLs, including those for 30 proteins which were shared across all ancestries, with a few proteins, exhibiting stronger associations in non-EUR ancestry groups despite smaller sample sizes. Across ancestries, 7% (EUR), 25% (AFR), and 14% (CSA) of associations had variance effects only (vQTL _only ), lacking corresponding mean effects (P _ME > 0.05), with chromosome X enriched for vQTL _only associations. Finally, multivariable Mendelian randomization revealed that, independent of genetically predicted mean protein levels, genetically predicted variance of three proteins influenced disease risk of coronary artery disease (Lp(a) and VAMP5) or type 2 diabetes (ANGPTL4). The MR effects for protein levels and variance were independent yet directionally consistent and significant (FDR < 0.05). Taken together, this study identifies novel protein vQTLs, highlights their transferability and demonstrates the potential therapeutic relevance of protein variance.

We propose an age-dependent topic modelling (ATM) model, providing a low-rank representation of longitudinal records of hundreds of distinct diseases in large electronic health record data sets. The model assigns to each individual topic weights for several disease topics; each disease topic reflects a set of diseases that tend to co-occur as a function of age, quantified by age-dependent topic loadings for each disease. The model assumes that for each disease diagnosis, a topic is sampled based on the individual’s topic weights (which sum to 1 across topics, for a given individual), and a disease is sampled based on the individual’s age and the age-dependent topic loadings (which sum to 1 across diseases, for a given topic at a given age). The model generalises the Latent Dirichlet Allocation (LDA) model by allowing topic loadings for each topic to vary with age. References: Jiang (2023) .