Plasma Proteomic Signature Predicts Myeloid Neoplasm Risk

Abstract

Purpose: Clonal hematopoiesis (CH) is thought to be the origin of myeloid neoplasms (MN). Yet, our understanding of the mechanisms driving CH progression to MN and clinical risk prediction of MN remains limited. The human proteome reflects complex interactions between genetic and epigenetic regulation of biological systems. We hypothesized that the plasma proteome might predict MN risk and inform our understanding of the mechanisms promoting MN development.

Experimental design: We jointly characterized CH and plasma proteomic profiles of 46,237 individuals in the UK Biobank at baseline study entry. During 500,036 person-years of follow-up, 115 individuals developed MN. Cox proportional hazard regression was used to test for an association between plasma protein levels and MN risk.

Results: We identified 115 proteins associated with MN risk, of which 30% (N = 34) were also associated with CH. These were enriched for known regulators of the innate and adaptive immune system. Plasma proteomics improved the prediction of MN risk (AUC = 0.85; P = 5×10-9) beyond clinical factors and CH (AUC = 0.80). In an independent group (N = 381,485), we used inherited polygenic risk scores (PRS) for plasma protein levels to validate the relevance of these proteins toMNdevelopment. PRS analyses suggest that most MN-associated proteins we identified are not directly causally linked toMN risk, but rather represent downstream markers of pathways regulating the progression of CH to MN.

Conclusions: These data highlight the role of immune cell regulation in the progression of CH to MN and the promise of leveraging multi-omic characterization of CH to improveMN risk stratification. See related commentary by Bhalgat and Taylor, p. 3095.

Figure 1.

Association between plasma protein levels, incident myeloid neoplasm (MN), and clonal hematopoiesis (CH). A, Multivariable Cox regression adjusted for age, sex, smoking status, year of sample collection, blood count measurements, and the top 10 principal components of genetic ancestry was used to test for an association between plasma protein expression and MN (left). Multivariable logistic regression adjusted for the same covariates was used to test for an association with CH (right). The y-axis displays the z-score transformed beta coefficients from the regressions. Each circle represents one protein colored by statistical significance (FDR-corrected P value < 0.05) for MN (blue), CH (green), MN and CH (red), or none (gray). The top 20 results in each plot are labeled by protein name. B, Multivariable Cox for the top 25 proteins associated with MN risk with (teal) or without (red) correction for CH (by maximum VAF) regression for MN risk. Shown are the hazard ratios and their 95% confidence intervals. C, Discrimination of MN risk incorporating basic clinical features alone or alongside CH and/or proteomics. The clinical model incorporates age, sex, ancestry, smoking status, blood counts, and sample age. Receiver operating (ROC) curves include the mean AUC and 95% confidence interval (lighter shading) generated from 100-fold cross-validation.

Figure 2.

Association between polygenic risk score (PRS) for plasma protein levels and MN in an independent subset of the UKBB (A). Association between global (i.e., genome-wide) PRS (PRS^global-pQTL) for plasma protein levels and risk of MN. For each protein, PRSs for each individual were constructed by summing the product of the beta coefficients for the association between and the number of alternate alleles for each protein quantitative loci detected through a genome-wide interrogation. The beta coefficients were obtained from SNP-protein association analysis. Displayed are the hazard ratio and 95% confidence intervals for proteins in which genetically predicted levels were significantly associated with MN risk. B, Two-sample Mendelian randomization (MR) analyses of the causal impact of plasma protein levels on MN risk. Each dot represents the estimated MR effect for the corresponding protein with the corresponding 95% confidence interval.