Proteomic analysis of cardiorespiratory fitness for prediction of mortality and multisystem disease risks

Abstract

Despite the wide effects of cardiorespiratory fitness (CRF) on metabolic, cardiovascular, pulmonary and neurological health, challenges in the feasibility and reproducibility of CRF measurements have impeded its use for clinical decision-making. Here we link proteomic profiles to CRF in 14,145 individuals across four international cohorts with diverse CRF ascertainment methods to establish, validate and characterize a proteomic CRF score. In a cohort of around 22,000 individuals in the UK Biobank, a proteomic CRF score was associated with a reduced risk of all-cause mortality (unadjusted hazard ratio 0.50 (95% confidence interval 0.48-0.52) per 1 s.d. increase). The proteomic CRF score was also associated with multisystem disease risk and provided risk reclassification and discrimination beyond clinical risk factors, as well as modulating high polygenic risk of certain diseases. Finally, we observed dynamicity of the proteomic CRF score in individuals who undertook a 20-week exercise training program and an association of the score with the degree of the effect of training on CRF, suggesting potential use of the score for personalization of exercise recommendations. These results indicate that population-based proteomics provides biologically relevant molecular readouts of CRF that are additive to genetic risk, potentially modifiable and clinically translatable.

Fig. 1. Study design.

We developed and validated a circulating proteomic signature of CRF across four cohorts and various exercise modalities. In the UKB, we examined the relationship a proteomic CRF signature with a broad range of clinical endpoints and examined its interaction with polygenic risk. In HERITAGE, we examined the association of the proteomic CRF signature with response to exercise training and correlated changes in signature with changes in CRF. NAFLD, nonalcoholic fatty liver disease.

Fig. 2. Development of the proteomic CRF score in CARDIA.

a, Correlations between the proteomic CRF score and CRF (defined by ETT time) in CARDIA across derivation (left) and validation (right) samples. b–d, Correlations of the proteomic CRF score with age (b), sex and race (c) and BMI (d). Colors on scatter plots represent density of overlapping observations, with red being the most dense and blue the least dense. P values in a, b and d are from Spearman rank correlation tests. P values in c are from linear regression modeling of the proteomic CRF score as a function of sex and race. All P values are from two-sided tests.

Fig. 3. Proteomic CRF score, polygenic risk and multisystem clinical outcomes.

a, Forest plot of Cox model results with proteomic score as the main predictor, grouped by outcome category. The ‘full’ adjustment model includes adjustment for age, sex, race, BMI, systolic blood pressure, diabetes, Townsend deprivation index, smoking, alcohol and LDL. Error bars, 95% CI. The adjoining table reports the C-index for Cox models without proteomic score (Base) and with the score (Score). Base models include age, sex, race, BMI, systolic blood pressure, diabetes, Townsend deprivation index, smoking, alcohol and LDL. Reported P value is from comparison testing of C-indices by z distribution (two-sided) without correct for multiple comparison. b, Cox beta coefficients from models including an interaction between the protein score of CRF and PRSs of the indicated conditions or diseases. Error bars, 95% CI. c, Contour map of the model predicted HR across the range of protein score of fitness and PRSs. The referent hazard was set at the median of the protein score and median of the PRS. Values reported and visualized are from point estimates and 95% CI. d, Comparison of Cox model coefficients from a parsimonious 21-protein panel and the full 307-protein panel. The halo represents the 95% CI around the model coefficient. P value is from two-sided Spearman rank correlation test. For visualization, we reversed the sign of the beta coefficients. Full data on sample sizes, model estimates and results of statistical testing may be found in Supplementary Tables 7 and 13.

Fig. 4. Proteins related to CRF whose levels are dynamic with exercise training are related to cardiometabolic risk factors and diseases.

Heatmap of Pearson correlations between individual proteins and cardiometabolic risk factors and disease in CARDIA using the CARDIA validation sample (N = 589–669). Proteins visualized are included in the proteomic CRF score and change after a 20-week exercise intervention in HERITAGE (false discovery rate < 5%). Proteins marked with an asterisk are included in the abbreviated 21-protein score. Cells marked with an asterisk indicate Pearson correlations with false discovery rate < 5%. AAC, abdominal aorta calcification; AHA LS7, American Heart Association Life Simple 7; CAC, coronary artery calcification; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; FC, fold change; GLS, global longitudinal strain; HbA1c, hemoglobin A1c; HDL, high density lipoprotein; LV, left ventricular; PA, physical activity; SAT, subcutaneous adipose tissue; SBP, systolic blood pressure; VAT, visceral adipose tissue.

Extended Data Fig. 1. Relationship of a protein score of fitness with VO₂ max, age, sex, race and BMI in 3 validation cohorts.

The proteomic CRF score was scaled (mean 0, variance 1) in BLSA and HERITAGE cohorts. Colors on scatter plots represent density of overlapping observations with red being the most dense and blue the least dense. P values on panels showing the relationship of the proteomic CRF score with sex and race are from linear regression models of the proteomic CRF score as a function of sex and race. All other panels report P values from Spearman rank correlation tests. P values below 2.2 × 10^–16 are reported as p < 2.2e-16.

Extended Data Fig. 2. Relations of a protein score of fitness with age, sex, race and BMI in UK Biobank.

Colors on scatter plots represent density of overlapping observations with red being the most dense and blue the least dense. P values on panels showing the relationship of the proteomic CRF score with sex and race are from linear regression models of the proteomic CRF score as a function of sex and race. All other panels report P values from Spearman rank correlation tests. P values below 2.2 × 10^–16 are reported as p < 2.2e-16.

Extended Data Fig. 3. Correlation of change in proteomic CRF score with change in peak VO₂ with exercise training in HERITAGE.

After a 20-week exercise training program in HERITAGE, we observed correlation between changes in the proteomic CRF score with changes in peak VO₂, which were replicated in regression models. P value is from two sided Spearman rank correlation test.