While short-term kidney transplant outcomes are generally excellent, maintaining long-term graft function remains a challenge.1 Standard clinical surveillance still relies heavily upon estimated glomerular filtration rate and proteinuria, signaling damage only after significant physiological injury has occurred.2 Consequently, there is a clear need to identify molecular signals that allow for earlier interventions.3 In this context, Shehab et al4 present a study integrating untargeted plasma proteomics with machine learning to predict death-censored graft survival. Using pre-event biobank plasma samples from kidney transplant recipients who later experienced graft failure, and event-free matched controls, the authors apply Boruta-based feature selection within a Random Survival Forest framework. Boruta is an established method to identify variables that consistently exceed noise signals. This approach distills a high-dimensional proteomic dataset into a biomarker signature, which is given biological context through network and pathway analyses. The authors report that a protein-only model outperforms a clinical-only model. This finding suggests that the plasma proteome encodes a substantial latent representation of graft health5,6 with proteins such as apolipoprotein L1 and apolipoprotein C-I emerging as prominent contributors. Although the gain over standard clinical variables is modest, the predictive performance of the proteome-only model is encouraging. Several factors may explain why proteomic data are only marginally superior to clinical features alone. First, despite the high quality of the cohort, the analysis faces an inherent dimensionality challenge: approximately 800 proteins are evaluated across 78 samples, with only 29 graft failure events. Second, the study is restricted to a single time-point per patient: Longitudinal clinical and proteomic measurements would likely capture dynamic trajectories that are more informative than static snapshots. Third, biological and clinical heterogeneity within the cohort can obscure prognostic information when markers are evaluated uniformly across all samples. To explore this third aspect, exploratory post hoc computational analyses were performed using data shared by the authors. Applying the new method of dataset restriction led to small but consistent improvements in predictive performance (approximately a 1% increase in C-index) for both proteome-only and combined proteomic-clinical models.7,8 Dataset restriction identifies biomarker ranges where reliable predictions are possible. Restriction also sharpened univariate signals for several Boruta-selected proteins, including inter-alpha-trypsin inhibitor heavy chain H2 and paraxonase 3 , suggesting that part of the observed attenuation may reflect signal dilution rather than absence of biological relevance. In conclusion, Shehab et al4 provide a rigorous, transparent study demonstrating that plasma proteomics reveals prognostic information relevant to late kidney graft failure. At the same time, the findings highlight the challenges of extracting substantial incremental value from single-time-point molecular measurements in complex, heterogeneous clinical settings. Future progress in long-term graft surveillance will likely benefit from longitudinal clinical data complemented by proteomic measurements. Advanced computational approaches may be particularly important for capturing the dynamic processes underlying graft deterioration and for translating molecular signals into clinically meaningful predictions.
Gunther Glehr (Tue,) studied this question.