Proteomics Analysis of Plasma for Risk of Sepsis: Findings from the Atherosclerosis Risk in Communities Study

Abstract

Background

Sepsis is a life-threatening complication of infection with high mortality. A high-throughput analysis of circulating blood proteins may provide mechanistic insight and potent therapeutic targets for the prevention of sepsis.

Methods

We used multivariable Cox regression analysis to examine the association of 4955 plasma proteins, measured by SomaScan, with the risk of incident sepsis among 11 065 participants of the Atherosclerosis Risk in Communities (ARIC) Study (visit 3 in 1993 to 1995; mean age, 60.1 years, 54.4% female, 21.0% Black). Proteins (false discovery rate [FDR] of P < 0.05) discovered at visit 3 were replicated using data at visit 5 (n = 4869 in 2011 to 2013: mean age, 75.5 years) and in the Cardiovascular Health Study (CHS) (n = 3512 in 1992 to 1993; mean age, 74.5 years). Canonical pathways were identified by enrichment analyses.

Results

At ARIC visit three, 669 proteins were associated with the risk of sepsis; 175 were replicated at visit 5. Of these, 90 were validated in the CHS. The top 20 proteins ranked by P value were relevant to acute inflammatory signaling in innate immunity. Pathway analyses implicated activation of pro-inflammatory pathways (e.g., cytokine storm signaling) as well as inhibition of anti-inflammatory pathways (e.g., liver X receptor/retinoid X receptor [LXR/RXR] activation), which also play relevant roles in lipid metabolism.

Conclusions

In this analysis, levels of acute inflammatory proteins measured during routine visits were associated with the subsequent incidence of sepsis. An increased risk of sepsis associated with the inhibition of anti-inflammatory pathways, such as LXR/RXR warrants further mechanistic investigation.

IMPACT STATEMENT.

Sepsis is a life-threatening complication of infection with high mortality, yet preventive strategies remain limited. In a large proteomics analysis of more than 11 000 participants from the ARIC study, we identified circulating proteins associated with future risk of sepsis, with replication in independent cohorts. Findings highlight 90 reproducible proteins, particularly those involved in innate immunity and dysregulated inflammatory signaling. Importantly, inhibition of protective anti-inflammatory pathways such as LXR/RXR was linked to increased sepsis risk. These results provide novel insight into sepsis pathogenesis and reveal promising therapeutic targets for prevention, advancing precision strategies to reduce sepsis burden.

INTRODUCTION

Sepsis is a major public health concern linked to significant morbidity and mortality, affecting over 1.7 million adults annually in the United States (1) and representing the leading cause of in-hospital deaths (2). Sepsis disproportionally affects immunocompromised hosts, such as older adults and individuals with chronic conditions (e.g., diabetes and chronic kidney disease) (3, 4). Metabolic alterations in these conditions, such as the accumulation of glycated and uremic proteins, may increase sepsis susceptibility. Prior studies have used targeted approaches to associate specific biomarkers, such as interleukin-6 (IL-6) and tumor necrosis factor-α (TNFα), with sepsis incidence (5). However, these offer limited insight into the complex, multi-pathway pathophysiology of sepsis.

Advancements in -omics technology now allow high-throughput analysis of thousands of proteins to better understand disease mechanisms. Using data from the Atherosclerosis Risk in Communities (ARIC) Study, we applied a proteomics approach to investigate sepsis risk. We identified plasma proteins associated with sepsis in ARIC visit 3 (1993 to 1995) and replicated these in ARIC visit 5 (2011 to 2013) and the Cardiovascular Health Study (CHS) (1992 to 1993). We conducted pathway analysis to uncover key biological pathways, used Mendelian randomization to explore causal relationships, and assessed the predictive value of proteomic biomarkers for sepsis risk.

MATERIALS AND METHODS

Study Population

The ARIC Study is a community-based cohort of US adults (6, 7). Study visits occurred in 1987 to 1989 (visit 1), 1990 to 1992 (visit 2), 1993 to 1995 (visit 3), 1996 to 1998 (visit 4), 2011 to 2013 (visit 5), and 2016 to 2017 (visit 6). After visit 6, participants were invited for annual or biennial site visits. Blood protein levels were measured at visits 3 and 5. The primary analysis used visit 3 as baseline due to larger sample size and longer follow-up; visit 5 served as an internal replication. The primary visit 3 cohort included 11 065 participants (online Supplemental Fig. 1). The internal replication cohort at visit 5 included 4869 participants.

For external replication, we used data from the CHS, a population-based cohort of adults ≥65 years of age. A total of 5888 participants were enrolled at baseline (1989 to 1990 and 1992 to 1993). Proteomic measurements were obtained from 3678 participants during the 1992 to 1993 exam (8). For both the ARIC Study and the CHS, written informed consent was obtained from all participants, and the institutional review board at each study site approved the study.

Proteomics

In the ARIC Study, plasma samples from visits 3 and 5 were stored at −80°C and analyzed in 2018 to 2019 using the SomaScan v4 platform (SomaLogic). Proteins were captured by SOMAmer reagents and quantified via fluorescent arrays, and the SomaScan assay demonstrated overall stability across measurements and time points. In a study that examined the analytic, short- and long-term variation of the SomaScan assay in ARIC Study samples (9), the median coefficients of variation (CV) and intraclass correlations (ICC) were 5.0% and 0.96, respectively, across 3693 analytes from 40 participants. Regarding measurement stability over the short term (9 weeks) and long term (approximately 20 years), statistically significant differences were observed in one analyte (<0.1%) and 866 analytes (23.4%), respectively. After two-step quality control, 4955 of 5284 aptamers passed and were analyzed. In the CHS, EDTA-plasma samples stored at −70°C were analyzed using SomaScan v4 (n = 3188) or v4.1 (n = 490). Scaling factors aligned v4.1 aptamers with v4.0. Deprecated, nonhuman, or low-quality aptamers were excluded.

Outcome

Our primary outcome was incident hospitalization with sepsis (hereafter “incident sepsis”), which was captured through the International Classification of Diseases, Ninth or Tenth Revision, Clinical Modification (ICD-9/10-CM) on hospital discharge records (online Supplemental Table 1). In both the ARIC Study and the CHS, research staff asked participants or their proxies through annual phone calls or during study site visits whether they had been hospitalized since the last contact (10). Active surveillance ascertained all available hospitalization records including the list of discharge diagnoses and corresponding ICD codes.

For the present study, we defined incident sepsis as recording of sepsis at the primary diagnostic position. This approach assumes that sepsis was the primary reason for hospitalization or primary clinical problem during the hospitalization. We captured concomitant ICD9/10 codes for infections (e.g., pneumonia) (online Supplemental Table 2) and causative pathogens of sepsis when there were ICD-9/10 codes for pathogen-specific infections (e.g., Streptococcus pneumoniae) (online Supplemental Table 3). To avoid double-counting events between the analyses of visit 3 and visit 5, we applied specific criteria. If a participant attended both visit 3 and visit 5, we censored the data at the date of visit 5 for the primary analysis that used visit 3 as the baseline. For participants who attended visit 3 but did not attend visit 5, they were censored at the last date of visit 5. Other censoring criteria included participants who had an event of interest, died, were lost to follow-up, or were administratively censored on December 31, 2019, whichever occurred first.

Covariates

Age, sex, race, center, education, and smoking status were self-reported. Education was categorized as basic (<high school), intermediate (high school/vocational), or advanced (≥some college). Body mass index (BMI) was calculated from weight and height. Diabetes and hypertension were defined using standard criteria. Estimated glomerular filtration rate (eGFR) was calculated using the CKD-epidemiology creatinine equation (11). History of coronary heart disease, heart failure, and stroke were based on self-report, hospital discharge records, and physicians’ adjudication, as appropriate (6).

Statistical Analysis

Protein Levels and Risk of Incident Sepsis

Multivariable Cox models estimated hazard ratios (HRs) and 95% CIs. Protein levels were log2-transformed and treated as continuous variables. The models were adjusted for age, sex, race, center, education attainment, BMI, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure. We used FDR-adjusted P values (Benjamini–Hochberg, <5%) to address multiple testing. Significant proteins from visit 3 were tested at visit 5, then externally replicated in the CHS.

Pathway Analysis

We used Ingenuity Pathway Analysis (IPA, QIAGEN, Inc.) to examine the biological mechanisms involved in sepsis. IPA is a web-based application that contains a large, curated database of molecular interactions and gene-to-phenotype association knowledge. For this analysis, we used the proteins that were significantly associated with sepsis both at visit 3 and 5 of the ARIC Study. Pathway enrichment was assessed by P value and z-score. P value is based on the right-tail Fisher exact test. The threshold for statistical significance was a P value of <0.05 after Benjamini–Hochberg FDR adjustment. The threshold for Z-value was absolute z-scores of ≥2.

Causal Inference Analysis

Protein-wide association studies (PWAS) were conducted using Mendelian randomization with cis-protein quantitative trait loci (pQTL) (12). In the ARIC Study, a previous study related protein levels to cis-pQTL and developed models to genetic variants in the encoding region (pQTL) and models to infer putatively causal effects of these proteins on other traits by Mendelian randomization (13). We utilized the FUSION work flow (14), incorporating elastic net modeling, with weights derived from the European ancestry subpopulation of the ARIC Study. These weights were then combined with the corresponding European ancestry in-sample linkage disequilibrium reference (13). By combining PWAS model weights with summary statistics from large genome-wide association studies (GWAS) of select phenotypes, evidence of potential causal relationships was evaluated. In the current study, summary statistics from a GWAS of the ICD-9 code of 038 (Sepsis) in UK Biobank were combined with protein models from the ARIC Study (15). Significance was set at FDR <0.05.

Prediction Model for Sepsis Risk

To examine the ability of proteins to predict the risk of sepsis, we developed models using ARIC Study visit 3 data to predict the 10-year risk of incident sepsis. For this analysis, all participants were censored at 10 years if they remained in the risk set at that time. The first (“base”) model included variables of age, sex, race, education attainment, BNI, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure. The second (“LASSO”) model included proteins that were selected based on LASSO regression analysis, in addition to covariates for the base model.

For each of the base and LASSO models, we ran Cox proportional analyses and calculated the Harrel c-statistics. Models were validated by recalibrating β-coefficients with baseline hazard from validation cohorts (16). Prediction improvement was assessed by refitting models, in which multivariable Cox models were re-run using the same set of proteins as in the primary model but allowing for different beta coefficients. Fine and Gray competing risk models estimated sepsis incidence accounting for death. Calibration plots compared predicted and observed risks. All statistical analyses were performed using Stata 18.0 (StataCorp) and R statistical software (v4.1.1; R Core Team 2021).

RESULTS

Proteins Associated with Sepsis

In the ARIC Study visit 3 cohort (n = 11 065, mean age, 60.1 [SD, 5.7], 54.4% female, 21.0% Black) (online Supplemental Table 4), there were 463 cases of incident sepsis prior to visit 5 during a median follow-up of 17.3 (IQI, 14.5 and 18.5) years. Crude incidence rate per 1000 person-years was 2.7 (95% CI, 2.5–2.9). Causative organisms were documented in 29% of these cases, including 17% caused by gram-negative rods and 11% caused by gram-positive cocci (online Supplemental Fig. 2). Along with sepsis, concomitant infections were documented in 69% of the cases, with the most common infection being pneumonia (34%), followed by urinary tract infections (25%), cellulitis and osteomyelitis (6%), and gastrointestinal tract infections (4%) (online Supplemental Fig. 3).

In multivariable Cox analysis, 669 of 4955 proteins were significantly associated with incident sepsis (FDR <0.05) (“significant proteins”) (Fig. 1 and online Supplemental Tables 5 and 6). We then assessed the associations of these 669 proteins with sepsis in the ARIC Study visit 5 cohort, approximately 18 years after visit 3 (n = 4869, mean age, 75.5 years, 56.8% female, 18.8% Black). During a median follow-up of 7.2 (IQI [interquartile interval], 5.6 and 7.8) years after visit 5, there were 357 cases of sepsis; 175 of the 669 proteins remained significant (online Supplemental Tables 5 and 7). These 175 proteins were further evaluated in the CHS (n = 3512, mean age 74.5, 60.8% female, 18.2% Black). In the CHS, there were 321 cases of sepsis during a median follow-up of 12.5 (IQI, 7.7 and 18.2) years. Of the 175 proteins associated with sepsis in both ARIC Study visit 3 and visit 5, 90 proteins remained significant (Fig. 2 and Supplemental Table 7).

Fig. 1.

Volcano plots for risk of sepsis in ARIC Study visit 3 cohort. The models were adjusted for age, sex, race, center, education attainment, body mass index, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure. Red dots indicate statistical significance by FDR criteria, green dots indicate non-significance by FDR criteria. Top 20 proteins ranked by P values are labeled in the volcano plots.

Fig. 2.

Scatter plots comparing adjusted hazard ratios for sepsis between primary cohort (ARIC Study visit 3) and replication cohort (CHS). (A), ARIC visit 3 and visit 5; (B), ARIC visit 3 and CHS; (C), ARIC visit 5 and CHS. The models were adjusted for age, sex, race, education attainment, body mass index, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure.

Of these 90 proteins, the 20 with the lowest P values at visit 3 included 18 unique proteins, as shown in Table 1. These proteins were involved in innate immunity (i.e., initial response to pathogens), such as growth/proliferation factors (growth differentiation factor 15 [GDF15], epidermal growth factor receptor [EGFR], contactin 1 [CNTN1], hepatoma-derived growth factor [HDGF], neuroblastoma suppressor of tumorigenicity 1 [NBL1]); cytokines or their receptors (tumor necrosis factor receptor superfamily 1A/1B [TNFRSF1A/1B], interleukin-15 receptor subunit alpha [IL15RA], signaling lymphocytic activation molecule family 1 [SLAMF1]); or immune modulators (four-disulfide core domain protein 2 [WFDC2], beta-2 microglobulin [B2M], matrix metalloproteinase 7 [MMP7], triggering receptor expressed on myeloid cells 1 [TREM1], hepatitis A virus cellular receptor 2 [HAVCR2], thymopoietin [TMPO]). Positive coactivator 4 (PC4) and serine/arginine-rich splicing factor 1 (SRSF1) interacting protein 1 (PSIP1) is an essential protein for human immunodeficiency virus (HIV) integration (17). Epidermal growth factor (EGF)-containing fibulin-like extracellular matrix protein 1 (EFEMP1) (18) and sushi, von Willebrand factor type A, EGF and pentraxin domain containing 1 (SVEP1) (19) are extracellular matrix glycoproteins that interact with integrins to mediate immune cell communications. For the directionality of the association, EGFR and CNTN1 were associated with a reduced risk of sepsis, while the rest of the 16 proteins were associated with an increased risk of sepsis.

Table 1.

Top 20 proteins associated with sepsis risk in both ARIC Study and CHS cohorts.^a

Protein name	Full protein name	ARIC visit 3			ARIC visit 5			CHS
		HR	95% CI	P value	HR	95% CI	P value	HR	95% CI	P value
GDF15	Growth/differentiation factor 15	2.32	(1.94, 2.78)	4.21E−20	1.98	(1.61, 2.43)	1.33E−10	1.97	(1.48, 2.63)	3.61E−06
EGFR	Epidermal growth factor receptor	0.17	(0.10, 0.27)	4.36E−13	0.15	(0.09, 0.25)	1.97E−13	0.24	(0.12, 0.47)	2.62E−05
WFDC2	WAP four-disulfide core domain protein 2	2.04	(1.66, 2.50)	1.23E−11	2.80	(2.08, 3.78)	1.07E−11	2.04	(1.43, 2.92)	8.13E−05
TNFRSF1B	Tumor necrosis factor receptor superfamily member 1B	2.50	(1.91, 3.29)	4.29E−11	2.00	(1.68, 2.38)	1.41E−14	2.59	(1.80, 3.73)	2.57E−07
EFEMP1	EGF-containing fibulin-like extracellular matrix protein 1	2.88	(2.07, 4.02)	3.37E−10	1.94	(1.32, 2.85)	7.49E−04	2.04	(1.27, 3.27)	3.01E−03
B2M	Beta-2-microglobulin	2.36	(1.80, 3.09)	3.89E−10	2.21	(1.59, 3.09)	3.14E−06	2.67	(1.75, 4.07)	5.10E−06
TNFRSF1A	Tumor necrosis factor receptor superfamily member 1A	2.24	(1.74, 2.88)	4.28E−10	2.48	(1.76, 3.48)	1.72E−07	2.25	(1.51, 3.36)	6.45E−05
SVEP1	Sushi, von Willebrand factor type A, EGF and pentraxin domain-containing protein 1	2.04	(1.62, 2.56)	1.30E−09	1.82	(1.51, 2.20)	4.94E−10	1.61	(1.20, 2.15)	1.33E−03
MMP7	Matrilysin	1.68	(1.42, 2.00)	2.34E−09	1.52	(1.23, 1.89)	1.38E−04	1.24	(1.05, 1.48)	1.25E−02
TREM1	Triggering receptor expressed on myeloid cells 1	1.87	(1.52, 2.29)	2.36E−09	1.63	(1.27, 2.09)	1.33E−04	1.57	(1.14, 2.16)	6.07E−03
SVEP1	Sushi, von Willebrand factor type A, EGF and pentraxin domain-containing protein 1	1.93	(1.54, 2.41)	9.27E−09	1.94	(1.58, 2.39)	2.14E−10	1.49	(1.12, 1.98)	5.76E−03
IL15RA	Interleukin-15 receptor subunit alpha	2.03	(1.59, 2.60)	1.78E−08	1.99	(1.50, 2.64)	2.21E−06	1.60	(1.15, 2.22)	5.43E−03
HAVCR2	Hepatitis A virus cellular receptor 2	1.91	(1.52, 2.39)	1.88E−08	1.72	(1.31, 2.25)	9.55E−05	1.65	(1.18, 2.32)	3.61E−03
TNFRSF1B	Tumor necrosis factor receptor superfamily member 1B	1.88	(1.50, 2.36)	5.06E−08	1.98	(1.64, 2.40)	1.06E−12	2.69	(1.87, 3.87)	9.47E−08
CNTN1	Contactin-1	0.35	(0.24, 0.51)	7.30E−08	0.52	(0.34, 0.81)	3.27E−03	0.49	(0.30, 0.78)	2.68E−03
HDGF	Hepatoma-derived growth factor	1.70	(1.40, 2.07)	8.52E−08	1.48	(1.12, 1.96)	5.42E−03	1.61	(1.19, 2.17)	1.95E−03
PSIP1	PC4 and SFRS1-interacting protein	1.74	(1.42, 2.13)	1.10E−07	1.61	(1.21, 2.14)	1.12E−03	1.54	(1.16, 2.04)	3.00E−03
NBL1	Neuroblastoma suppressor of tumorigenicity 1	1.79	(1.44, 2.22)	1.34E−07	1.69	(1.36, 2.11)	2.21E−06	1.76	(1.34, 2.30)	4.43E−05
TMPO	Lamina-associated polypeptide 2, isoforms beta/gamma	1.52	(1.29, 1.78)	2.95E−07	1.59	(1.34, 1.90)	1.43E−07	1.66	(1.25, 2.19)	3.92E−04
SLAMF1	Signaling lymphocytic activation molecule	1.57	(1.32, 1.87)	2.99E−07	1.56	(1.27, 1.92)	3.16E−05	1.61	(1.26, 2.06)	1.36E−04

^aDue to duplicate proteins (SVEP1 and TNFRSF1B), there were 18 unique proteins. The models were adjusted for age, sex, race, education attainment, body mass index, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure. Statistical significance was based on false discovery rate (FDR) P < 0.05.

Abbreviation: HR, hazard ratio.

Pathway Analysis

The pathway analysis revealed 18 significantly enriched pathways, including the liver X receptor/retinoid X receptor (LXR/RXR) activation, pathogen-induced cytokine storm signaling pathway, wound-healing signaling pathway, acute phase response signaling, immunogenic cell death signaling pathway, IL-6 signaling, peroxisome proliferator-activated receptor (PPAR) signaling, and crosstalk between dendric cells and natural killer cells (Fig. 3 and online Supplemental Table 8A). For most pathways, activation was associated with a higher risk of sepsis (positive z score in Fig. 3). On the other hand, LXR/RXR activation and PPAR signaling are pathways relevant to anti-inflammatory signaling, and their inhibition was associated with a higher risk of sepsis (negative z score in Fig. 3). When we used the 90 proteins that were replicated in the CHS cohort for pathway analysis, 5 pathways were enriched with |z score| of ≥2, and 4 of them overlapped with the above 18 pathways. However, none of the 5 pathways remained significant after multiple testing correction (online Supplemental Table 8B).

Fig. 3.

Enriched canonical pathways associated with risk of sepsis determined by the Ingenuity Pathway Analysis (IPA). Input parameters comprised the 175 proteins that were significant at ARIC Study visits 3 and 5. Among the 175 proteins, 2 were not mapped and 3 were duplicate identifiers, resulting in 170 proteins that mapped to the IPA database. A small P value indicates that there are more proteins included in the pathway than expected by chance. The z score quantifies the consistency in the directionality of an association (e.g., inhibition vs activation) between observed and reference datasets. For example, a high positive z score indicates that known activating proteins in a pathway consistently showed positive associations in the observed dataset. The threshold for statistical significance was a P value of <0.05 after Benjamini–Hochberg FDR adjustment. The threshold for Z-value was absolute z scores of ≥2.

Causal Inference Analysis

Of the 175 proteins significant both at visit 3 and visit 5, PWAS models were available for 102 proteins (i.e., 102 proteins had valid instruments and were linked to the protein levels) (13) (online Supplemental Table 9). In these 102 proteins, the heritability of the gene (i.e., proportion of variation in protein levels explained by pQTL) was generally low (median [IQI], 4.9% [2.0% and 11.0%]). PWAS did not discover proteins that were causally associated with the risk of sepsis based on an FDR-corrected P value of <0.05 but did highlight 8 proteins at a nominal significance threshold of <0.05: macrophage-capping protein (CAPG), neural cell adhesion molecule 1 (NCAM1), ribonuclease T2 (RNASET2), cathepsin Z (CTSZ), TNFRSF1B, alpha-2-HS-glycoprotein (AHSG), fibulin-5 (FBLN5), and insulin-like growth factor-binding protein 2 (IGFBP2) (Table 2).

Table 2.

Protein-wide association study for risk of sepsis.^a

Gene symbol	Chromosome	SNPs (n)	Heritability of the gene, %	R 2 for prediction model	PWAS z score	PWAS P value
CAPG	2	55	7.88	0.16	−2.9323	0.00337
NCAM1	11	115	4.84	0.13	2.89506	0.00379
TNFRSF1B	1	50	3.47	0.049	−2.46657	0.0136
AHSG	3	78	11.00	0.19	−2.4245	0.0153
RNASET2	6	66	9.82	0.16	−2.349926	0.018777
FBLN5	14	75	0.94	0.022	2.09694	0.036
IGFBP2	2	34	0.44	0.011	−2.0968	0.03601
CTSZ	20	27	1.99	0.031	2.054	0.03997

^aOf the 102 proteins assessed in this analysis, only 8 highlighted proteins at a nominal significance threshold of <0.05 are presented in this table. P values for the rest of the 94 proteins are shown in online Supplemental Table 9.

Predictive Performance of Protein Models

Finally, we developed prediction models for risk of sepsis. The base model had a c-statistic of 0.710 (Table 3 and online Supplemental Table 10). Using LASSO regression, we identified 105 predictive proteins from the 669 proteins that were significant in the ARIC Study visit 3 cohort. Adding these proteins significantly improved the c-statistic (Δc-statistic, 0.085 [95% CI, 0.068–0.101]).

Table 3.

c-Statistics for predicting risk of sepsis in the discovery cohort (ARIC visit 3) and internal/temporal (ARIC visit 5) and external (CHS) replication cohorts.^a

Approach/Models	ARIC visit 3 (discovery)		ARIC visit 5 (internal/temporal validation)		CHS (external validation)
	c-Statistic	Δc (95% CI)	c-Statistic	Δc (95% CI)	c-Statistic	Δc (95% CI)
Approach 1
Base model	0.710	Reference	0.649	Reference	0.633	Reference
LASSO model	0.795	0.084 (0.068, 0.101)	0.691	0.041 (0.025, 0.058)	0.653	0.020 (−0.006, 0.046)
Approach 2
Base model	—	—	0.661	Reference	0.652	Reference
LASSO model	—	—	0.754	0.093 (0.065, 0.121)	0.732	0.080 (0.052, 0.108)

^aLASSO regression models selected the list of predictive proteins and their corresponding β-coefficients, where the number of predictive proteins was determined by a tuning parameter λ. To balance between the predictive performance and model overfitting, we chose a λ that minimized the cross-validated partial likelihood deviance based on 10-fold cross-validation. Approach 1 used the same set of predictors and regression coefficients from the discovery cohort for the validation cohorts. Approach 2 used the same set of predictors but allowed refitting regression coefficients within each validation cohort. The models were adjusted for age, sex, race, education attainment, body mass index, smoking status, eGFR, diabetes, hypertension, prevalent coronary heart disease, prevalent stroke, and prevalent heart failure.

This LASSO model also significantly improved prediction in the c-statistics in the ARIC Study visit 5 cohort (Δc-statistic, 0.041 [95% CI, 0.025–0.058]), while the improvement was not statistically significant in the CHS (0.020 [95% CI, −0.006 to 0.046]). However, the same set of 105 proteins considerably improved the c-statistic when we refitted the models (Δc-statistic, 0.093 [95% CI, 0.065–0.121] for the ARIC Study visit 5 data; and 0.080 [95% CI, 0.052–0.108] for the CHS data) (Table 3).

For the ARIC Study visit 3, the LASSO model showed excellent calibration of 10-year sepsis risk, with both observed and predicted risks exceeding 7% in the top decile and remaining below 0.5% in the first through fifth deciles (online Supplemental Fig. 4). These findings were also consistent in Cox analyses: the HRs comparing the bottom vs top decile were 51.4 (95% CI, 24.1–109.7) for the LASSO model (online Supplemental Fig. 5A and B). These findings were consistent when using ARIC Study visit 5 data although the risk gradients were less evident, reflecting a higher baseline risk in this population (online Supplemental Fig. 6A and B). In the CHS, the LASSO model showed good calibrations in the low predicted risk groups (i.e., first to sixth deciles) but not in the high predicted risk groups (i.e., seventh to 10^th deciles). (online Supplemental Fig. 6C and D).

DISCUSSION

In this study, we identified 669 proteins significantly associated with sepsis risk in the midlife (visit 3) ARIC Study cohort. Of these, 175 proteins were replicated in the late-life (visit 5) ARIC Study cohort, and 90 were further replicated in the external CHS cohort. The top hits included 18 proteins involved in acute inflammatory processes and innate immunity. Pathway analysis confirmed enrichment of both pro-inflammatory pathways, such as cytokine storm signaling, and anti-inflammatory pathways, such as LXR-RXR activation. Additionally, we showed that a proteomics-informed model can effectively identify individuals at high risk of sepsis.

Of the top 20 proteins (including 18 unique proteins), 16 proteins were positively associated with the risk of sepsis, including TNFRSF1A and TNFRSF1B. Although tumor necrosis factor receptor (TNFR) is a membrane-bound protein, the extramembrane portion of TNFRSF1A and TNFRSF1B is cleaved by the enzyme to form soluble TNFRSF1A and TNFRSF1B. In an event of sepsis, the levels of soluble TNFRSF1A and TNFRSF1B are markedly elevated in response to TNFα production (20). Soluble TNFR has capacity to bind TNFα and inhibit the excessive activation of acute TNFα signaling (21). Meanwhile, persistent TNF signaling activation can disrupt the translocation of nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) to the nucleus (22, 23), and cause hypo-responsiveness to toll-like receptor signaling (24). Similar mechanisms of persistent stimulation leading to receptor desensitization have also been reported for chemokines (25), interleukins (26), and growth factors (27). Our findings support a hypothesis where chronic “irritation” of inflammatory pathway while free of the disease, as manifested by higher levels of inflammatory markers in the blood, may impair a healthy immune response against infection, thereby increasing susceptibility to sepsis.

On the other hand, some proteins such as EGFR and CNTN1 showed an inverse association with the risk of sepsis (i.e., higher levels were associated with a lower risk of sepsis). EGFR has been extensively researched as an oncogenic protein highly expressed in cancer cells, and it also plays a significant role in innate immunity. In healthy tissues, EGFR exhibits minimal expression, but its levels rapidly escalate through toll-like receptor signaling activation, triggering various responses such as interleukin activation, neutrophil recruitment, and epithelial repair (28). CNTN1 has been initially identified as a neuronal protein that regulates signaling between myelin and axon and recently garnered attention as a potent oncogenic protein (29). Regarding its role in immunity, CNTN1 suppresses retinoic acid-inducible gene I (RIG-I) and mitochondrial antiviral-signaling protein (MAVS) signaling in innate immunity (30), and the loss of Cntn1 function resulted in global immune deficiency (31). However, both EGFR and CNTN1 are membrane-bound, and the mechanisms governing their release into the bloodstream, as well as the biological functions of their soluble forms, remain poorly understood.

Pathway analyses provide further insight into a pathophysiological landscape beyond the individual roles of proteins. Enriched pathways were consistently related to acute inflammatory pathways particularly in the innate immune response, such as LXR-RXR activation, cytokine storm signaling pathway, wound-healing signaling pathway, and IL-6 signaling pathway. All of these pathways included at least one of interleukin-1 (IL-1), TNFα, or IL-6, which are known to activate classic pro-inflammatory transcription factors such as the Janus kinase/signal transducer and activator of transcription (JAK/STAT) and phosphatidylinositol 3-kinase and protein kinase B (PI3K/Akt) and the mammalian target of rapamycin (mTOR).

Interestingly, some enriched pathways, such as LXR-RXR activation and PPAR signaling, showed their inhibition to be associated with the risk of sepsis, rather than activation. Recent evidence has revealed that these pathways activate anti-inflammatory signaling by suppressing pro-inflammatory nuclear receptors, such as NFkB and other pro-inflammatory cytokines (32). Further, both LXR and PPAR are nuclear receptors (i.e., they regulate gene transcriptions) that play central roles in lipid and glucose metabolism (33). For example, LXR is activated by cholesterol overload in macrophages, and induces the efflux of cholesterol and the expression/secretion of cholesterol transporters (e.g., apolipoprotein E, high-density lipoprotein [HDL] cholesterol) (33). Epidemiological studies have demonstrated associations of dyslipidemia, such as low HDL, with the risk of pneumonia (34). Further, several medications that modulate lipid metabolism, such as statin (35) and metformin (36), have been explored for the prevention or treatment of infections, although clinical trial data are limited.

PWAS did not discover significant causal proteins for sepsis once accounting for multiple comparison. However, low heritability (<10%) of assessed proteins may limit statistical power. Furthermore, the development of sepsis is influenced by non-host factors (e.g., pathogen factors). Nonetheless, highlighted proteins (i.e., unadjusted P values < 0.05) including TNFRSF1B, are reasonably linked to immune response. RNASET2 (37) and CTSZ (38) are enzymes that degrade microbial proteins or RNAs to be recognized by toll-like receptors. CAPG, FBLN5, and NCAM1 are primarily expressed in macrophages and natural killer cells, regulating cell motility, adhesion, and migration during initial immune response against pathogens. On the other hand, AHSG and IGFBP2 are inhibitors of the insulin receptor or components of the insulin-like growth receptor, respectively. These proteins have pleiotropic effects on cell growth signaling and lipid and glucose metabolism but also modulate the immune response (39).

Finally, we found that a prediction model based on proteins showed an ability to discriminate individuals at risk of sepsis, particularly those with the highest risk of sepsis. The model was not necessarily well replicated in an external older-age cohort. Further, the exploratory nature of this analysis should be noted, and application of the prediction model may be limited since proteomics data are yet to be utilized in routine clinical care or health screening settings. Nonetheless, the same set of 105 proteins considerably improve c-statistics, including the external CHS cohort, when we refit the models, which is noteworthy given that the risk for communicable diseases should be determined by both host and pathogen factors.

Several limitations should be acknowledged. First, observational studies are subject to residual confounding. Second, the generalizability of our findings may be restricted to White and Black individuals in the US community. Third, our outcome ascertainment relied on ICD codes recorded in the primary diagnostic position on discharge records. This approach may offer high specificity but low sensitivity. Fourth, internal validation in the ARIC Study late-life (visit 5) cohort may be confounded because some participants contributed to the risk set for both the midlife and late-life cohorts, although events were not double-counted between these analyses. Finally, caution is warranted when interpreting our findings in the context of other proteomics assay platforms. Although newer platforms (e.g., SomaScan 11 K and Olink Explore HT) are capable of capturing a greater number of proteins, their assay characteristics (e.g., precision and detectability) and corresponding implications may differ from those of the current platform (40).

In conclusion, in this large-scale proteomics analysis, levels of acute inflammatory proteins measured during routine visits were associated with the subsequent incidence of sepsis. An increased risk of sepsis associated with the inhibition of anti-inflammatory pathways, such as LXR/RXR activation, warrants further mechanistic investigation.

Author Declaration

A version of this paper was previously posted as a preprint on medRxiv as https://doi.org/10.1101/2025.03.07.25323594.

Data Availability

Pre-existing data access policies for each of the parent cohort studies specify that research data requests can be submitted to each steering committee; these will be promptly reviewed for confidentiality or intellectual property restrictions and will not unreasonably be refused. Please refer to the data sharing policies of these studies. Individual-level patient or protein data may further be restricted by consent, confidentiality or privacy laws/considerations. These policies apply to both clinical and proteomic data.

Supplemental Material

Supplemental material is available at The Journal of Applied Laboratory Medicine online.

Nonstandard Abbreviations: ARIC, Atherosclerosis Risk in Communities; FDR, false discovery rate; CHS, Cardiovascular Health Study; LXR/RXR, liver X receptor/retinoid X receptor; TNFα, tumor necrosis factor-α; ICD, International Classification of Diseases; PWAS, protein-wide association studies; IQI, interquartile interval; EGFR, epidermal growth factor receptor; CNTN1, contactin 1; TNFRSF1A/1B, tumor necrosis factor receptor superfamily 1A/1B; PPAR, peroxisome proliferator-activated receptor.

Human Genes:  CAPG: macrophage-capping protein; NCAM1: neural cell adhesion molecule 1; RNASET2: ribonuclease T2, CTSZ: cathepsin Z; TNFRSF1B: tumor necrosis factor receptor superfamily member 1B; AHSG: alpha-2-HS-glycoprotein; FBLN5: fibulin-5; IGFBP2: insulin-like growth factor-binding protein 2.

Author Contributions:  The corresponding author takes full responsibility that all authors on this publication have met the following required criteria of eligibility for authorship: (a) significant contributions to the conception and design, acquisition of data, or analysis and interpretation of data; (b) drafting or revising the article for intellectual content; (c) final approval of the published article; and (d) agreement to be accountable for all aspects of the article thus ensuring that questions related to the accuracy or integrity of any part of the article are appropriately investigated and resolved. Nobody who qualifies for authorship has been omitted from the list.

Authors’ Disclosures or Potential Conflicts of Interest:  Upon manuscript submission, all authors completed the author disclosure form.

Research Funding: The Atherosclerosis Risk in Communities Study has been funded in whole or in part with federal funds from the National Heart, Lung, and Blood Institute, National Institutes of Health, Department of Health and Human Services, under Contract Nos. (75N92022D00001, 75N92022D00002, 75N92022D00003, 75N92022D00004, 75N92022D00005). SomaLogic Inc. conducted the SomaScan assays in exchange for use of ARIC data. This work was supported in part by NIH/NHLBI grant R01 HL134320. The Cardiovascular Health Study research was supported by NHLBI contracts HHSN268201200036C, HHSN268200800007C, HHSN268201800001C, N01HC55222, N01HC85079, N01HC85080, N01HC85081, N01HC85082, N01HC85083, N01HC85086, and 75N92021D00006; and NHLBI grants U01HL080295, R01HL087652, R01HL105756, R01HL103612, R01HL120393, U01HL130114, and HL144483 with additional contribution from the National Institute of Neurological Disorders and Stroke (NINDS). Additional support was provided through R01AG023629 from the National Institute on Aging (NIA).

Disclosures: J. Coresh was a scientific advisor to Soma Logic until November 2023.

Role of Sponsor: The funding organizations played no role in the design of study, choice of enrolled patients, review and interpretation of data, preparation of manuscript, or final approval of manuscript.

Acknowledgments: The authors thank the staff and participants of the ARIC Study for their important contributions. A full list of principal Cardiovascular Health Study investigators and institutions can be found at CHS-NHLBI.org.