BGPT: Paper Review: Proteomics of CKD progression in the chronic renal insufficiency cohort

Explore by Goal

Quick Explanation Copied

Paper-at-a-glance (science-focused)

Large-scale plasma proteomics (SomaScan V4.0; 4638 proteins) in CRIC identified hundreds of proteins associated with 10-year CKD progression (≥50% eGFR decline or kidney failure), built a 65-protein prognostic model (C-statistic ~0.862 in the CRIC test set), validated external performance in ARIC (C-statistic ~0.840), and added a genetics-informed “potential causality” layer via Mendelian randomization for a subset of proteins. Key biological themes emphasized: ephrin signaling, BMP antagonists/growth factor modulation, and prothrombin activation. Source: 10.1038/s41467-023-41642-7

Long Explanation

Proteomics of CKD progression in the chronic renal insufficiency cohort

Nature Communications (2023). DOI: 10.1038/s41467-023-41642-7

Known vs inferred vs uncertain (epistemic hygiene)

Known from this paper: proteomic associations, pathway over-representation results, model discrimination/calibration summaries, MR screening results, and which proteins were considered druggable.
Inferred: biological plausibility links (ephrin/BMP/prothrombin themes) are interpretive; MR is used to argue “potential causality,” but MR assumptions may not hold perfectly (e.g., pleiotropy/invalid instruments).
Uncertain / not directly demonstrated: whether any specific protein is a true causal mediator in human CKD progression (requires stronger causal triangulation: tissue mechanistic work, longitudinal perturbation, and replication in independent proteomics platforms/cohorts).

Figure 1. Study design, cohorts, and outcomes (as reported)

Visual map of cohort sizes, event counts, and endpoint definitions. Values are taken directly from the paper text provided.

Figure 2. Multiplicity & filtering: from 4638 proteins to “top hits” and final model

Figure 3. Prognostic discrimination: protein vs clinical vs hybrid summaries

The paper provides C-statistic and confidence intervals for the protein models and hybrid models at key time horizons. Where the expanded clinical model’s C-statistics are not numerically provided in the excerpt, this chart focuses on the protein/hybrid points explicitly stated.

Figure 4. Pathways enriched among CKD-progression-associated proteins

Top canonical pathways reported from IPA over-representation analysis among proteins associated after eGFR adjustment (FDR<0.05) are summarized as “ratio = significant proteins / proteins in pathway measured.”

Figure 5. Reported Mendelian randomization “potentially causal” proteins (count summary)

This is a count-level summary of the MR results stated in the paper excerpt.

Core scientific claims (grounded in reported results)

1) Proteome-wide associations with CKD progression

The paper reports that among 4638 measured plasma proteins in CRIC, 330 proteins were associated with the 10-year primary outcome under a fully adjusted model at FDR q<0.05, and 100 remained significant under a Bonferroni-corrected threshold.

2) Biological pathway convergence (ephrin, BMP antagonists, prothrombin activation)

IPA pathway analysis identified canonical pathways enriched among proteins associated with the primary outcome (e.g., ephrin A signaling and intrinsic/extrinsic prothrombin activation; plus LXR/RXR activation and matrix metalloprotease inhibition). The discussion interprets these themes as mechanistically relevant to CKD progression.

3) A multi-protein prognostic model with reported discrimination and calibration

Using elastic-net Cox modeling in an 80/20 train/test partition of CRIC, the authors report a 65-protein model for the 10-year primary endpoint with C-statistic ~0.862 (95% CI 0.835–0.889) in the CRIC testing set, plus external validation in ARIC with C-statistic ~0.840 (95% CI 0.785–0.896). For the secondary 4-year outcome (eGFR slope), they report a 20-protein model with C-statistic ~0.728 (95% CI 0.708–0.748).

4) Causality-adjacent triangulation via Mendelian randomization and druggability mapping

The authors performed MR on a selected subset (reported as 76 aptamers/75 proteins), and report 8 proteins with MR significance after multiple-testing correction in at least one GWAS. They also report druggability counts among the proteins in their risk models (e.g., 14/65 druggable targets in the primary model; 3/20 in the secondary model).

Skeptical critique: key limitations, bias risks, and what could change the conclusion

Residual confounding in observational associations: Even with substantial covariate adjustment, unmeasured factors (disease severity proxies, medications not captured, inflammation/vascular comorbidity not fully represented) could generate associations that later look “causal-like.” The paper acknowledges the limits of epidemiological association and residual confounding risk.
Multiplicity and selection pipeline: Proteins are screened and then filtered (FDR, effect-size ranking, and replication selection). This can still lead to winner’s curse—models that look strong in one cohort can be unstable across platforms or populations. The external validation helps, but the validation is limited to ARIC and the proteomics platform is specific to SomaScan.
Platform biology mismatch (plasma vs kidney tissue): Plasma proteins may reflect systemic processes (inflammation, coagulation, liver production, clearance effects) rather than kidney-local mechanisms. The paper explicitly measures circulating proteins and discusses future tissue correlation work as needed.
MR assumptions not fully resolved: MR treats genetic instruments as proxy for lifelong differences in protein levels, but instruments can violate exclusion restriction via pleiotropy or linkage disequilibrium. The paper’s MR strategy supports “potentially causal” signals, but it still requires replication and deeper sensitivity analyses to confirm robustness (e.g., pleiotropy-robust methods, colocalization checks). The paper also notes that MR analyses could be augmented using more comprehensive renal-function GWAS.
Generalizability and ancestry coverage: MR pQTL selection used European ancestry restriction for a CRIC subset (as stated in methods), and GWAS resources used are typically European-heavy. That can reduce portability of causal claims across ancestries and may affect effect sizes for protein associations.
Outcome definition anchored to kidney function estimation: The endpoint relies on eGFR-based decline and kidney failure. eGFR estimation models can have systematic biases (e.g., creatinine-related confounding), though the paper includes sensitivity analyses with alternative eGFR equations and reports similar discrimination for the primary protein model.

What would disprove or materially weaken the paper’s practical implications?

(i) Failure to reproduce the 65-protein model’s discrimination in additional independent CKD cohorts (especially different labs/platforms), (ii) calibration failure in external settings, (iii) evidence that MR-significant proteins are explained by pleiotropic pathways rather than protein-mediated effects, and/or (iv) lack of mechanistic coherence between plasma protein signals and kidney tissue biology.

Druggability: interpretation caution

The paper reports that a subset of proteins in the risk model are “druggable” based on a therapeutic target database search. This mapping supports target feasibility, not therapeutic validity. A key skeptical point: plasma protein association + druggability does not guarantee that modulating the target will slow CKD progression, nor that it will be safe/efficacious in the relevant patient strata.

Author reviews (open BGPT deep-dives)

Feedback:

Updated: April 22, 2026