Machine Learning and Proteomics Reveal Why Kidney Injury Turns Chronic

When a person survives an episode of acute kidney injury, the danger is often not over. The kidney may appear to recover, yet months or years later a substantial fraction of these patients slide silently into chronic kidney disease, a progressive, irreversible condition that eventually requires dialysis or transplantation. Clinicians have long lacked reliable tools to predict which patients will make this transition, and the molecular events that drive a healing kidney toward scarring have remained frustratingly opaque. Now a team of researchers led by investigators at Xiangya Hospital of Central South University has built a multi-layered discovery pipeline that combines population-scale proteomics, machine learning, genetic causal inference, single-cell sequencing and structural biology to identify the proteins that shepherd this transition — and, remarkably, to pinpoint atomic-level features on one of those proteins that drug developers could target.

The scale of the discovery effort is one of its defining strengths. Drawing on the UK Biobank, the team screened 53,014 participants who had blood protein measurements generated with the Olink platform, a technology that quantifies thousands of circulating proteins simultaneously using proximity extension assays. Within this cohort, they identified 2,532 patients who had experienced acute kidney injury, of whom 410 subsequently progressed, providing a rich set of outcome events for predictive modeling. Starting from 2,923 measured proteins, the researchers applied what they call a triple-orthogonal screening strategy, a design intended to ensure that every surviving candidate is supported by independent lines of evidence rather than by a single statistical signal that might reflect confounding or chance.

The first layer of this strategy is an ensemble machine learning framework built from six algorithms, including least absolute shrinkage and selection operator regression, known as LASSO, and random forest recursive feature elimination. LASSO works by shrinking the coefficients of uninformative features toward zero, effectively discarding proteins that add little predictive value, while random forest recursive feature elimination iteratively removes the least important variables across thousands of decision trees. By requiring proteins to be retained across multiple algorithms rather than just one, the ensemble approach guards against the overfitting that plagues high-dimensional biomedical data, where the number of measured features vastly exceeds the number of patients.

The second layer applies Cox proportional hazards models, the standard statistical tool for time-to-event analysis, to ask which of the machine-learned candidates actually carry prognostic weight for progression to chronic kidney disease over follow-up. The third and perhaps most distinctive layer is two-sample Mendelian randomization, a technique that exploits naturally occurring genetic variants as proxies to test whether altered protein levels causally contribute to disease rather than merely correlating with it. Because genetic variants are randomly assorted at conception in a manner analogous to a randomized trial, this approach can help distinguish drivers of disease from innocent bystanders whose levels rise simply because the kidneys are already failing. Only proteins that survived all three orthogonal filters — machine learning, survival modeling and genetic causal inference — advanced to the final list.

That list contained 25 proteins, each possessing both statistical prognostic value and genetic support for a causal role in the acute-to-chronic transition. When the researchers trained an ensemble prediction model on these 25 features, it achieved an area under the receiver operating characteristic curve of 0.939, a level of discrimination that, if validated prospectively, would represent a major advance over existing clinical risk tools for this patient population. In practical terms, a score near 0.94 means the model can separate patients who will progress from those who will not with far greater accuracy than a coin flip or most conventional biomarker panels, potentially allowing clinicians to intensify monitoring and nephroprotective treatment precisely for those at highest risk.

To move beyond association and into mechanism, the team turned to experimental validation. They performed single-cell RNA sequencing to map gene expression across individual kidney cell types, and they studied a mouse model of ischemia-reperfusion injury — the interruption and restoration of blood flow that mimics the most common forms of human acute kidney injury — using paired bulk RNA sequencing and tandem mass tag-based quantitative proteomics. This dual approach revealed something unexpected about one of the top candidates, galectin-3, encoded by the LGALS3 gene. During the maladaptive repair phase, around day 14 after injury, the researchers documented what they describe as transcriptional-translational discordance: the gene’s messenger RNA fell silent, yet the protein itself continued to accumulate in the tissue. This decoupling suggests that galectin-3 is regulated after transcription during fibrotic remodeling, and it carries a broader lesson for biomarker research — measuring RNA alone can badly misjudge which proteins are actually driving pathology.

Galectin-3 is a beta-galactoside-binding lectin with well-documented roles in inflammation, macrophage activation and fibrosis across multiple organs, making its prominence in this analysis biologically coherent rather than a statistical curiosity. But the study went further than naming the culprit. Using structural profiling of the protein, the researchers identified a hotspot spanning residues Glu185 and Gln187, a region where the binding pocket that accommodates the inhibitor olitigaltin and anhydrous lactose spatially converges with high-scoring B-cell epitopes — the molecular surface patches that antibodies recognize. This convergence is significant for drug design: a pocket that binds a known small-molecule inhibitor and sits within an immunologically accessible surface region offers a concrete, structurally defined starting point for developing molecules that could neutralize galectin-3’s fibrogenic activity in the kidney.

The translational logic of the study is what sets it apart from most biomarker papers. Rather than stopping at a ranked list of statistically significant proteins, the authors built a workflow that moves from population-scale discovery through causal triangulation to experimental confirmation and finally to structure-guided therapeutic positioning, delivering what they describe as atomic-level coordinates for precision intervention against renal fibrosis. The animal work was conducted under ARRIVE guidelines with blinded outcome assessment, and the human data derive from the UK Biobank, which operates under ethics approval from the North West Multi-center Research Ethics Committee with written informed consent from all participants — details that strengthen confidence in the rigor of the underlying evidence.

The clinical implications extend in two directions. Diagnostically, a compact 25-protein panel measurable in blood could, in principle, be deployed shortly after an episode of acute kidney injury to stratify patients by their risk of chronic progression, enabling earlier referral, tighter blood pressure and metabolic control, and enrollment into trials of antifibrotic agents. Therapeutically, the galectin-3 findings suggest that intercepting this protein during the vulnerable window of maladaptive repair — when protein accumulation persists even as transcription shuts down — might blunt the fibrotic cascade before it becomes self-sustaining. The convergence of a druggable binding pocket with defined epitope features provides medicinal chemists with a map rather than a guess.

Caveats remain, as they do in any study of this ambition. The prediction model was developed and evaluated within UK Biobank data, and prospective validation in independent, ethnically diverse cohorts with serial kidney function measurements will be essential before clinical deployment. Mendelian randomization rests on assumptions about genetic instrument validity that can never be fully verified, and the mouse model of ischemia-reperfusion injury, while informative, does not capture every route to human kidney injury, from sepsis to nephrotoxic drugs. The article itself was published as an early, peer-reviewed accepted version subject to further edits. Even so, the framework demonstrated here — ensemble machine learning filtered through survival statistics and genetic causality, then stress-tested across single-cell, transcriptomic and proteomic layers, and finally anchored to protein structure — offers a template that could be applied to many other disease transitions where a reversible insult hardens into a chronic, fibrotic fate. For the millions of acute kidney injury survivors worldwide who quietly progress toward dialysis, that template may prove to be the most important output of all.

Subject of Research: Multi-omics identification of protein drivers of the transition from acute kidney injury to chronic kidney disease

Article Title: Integrated multi-omics decodes the AKI-to-CKD transition: from ensemble discovery to structure-guided translational targeting

Article References: Cao, X., Xiao, Y., Wang, Y., Huang, H., Kong, W., Chen, Z., Zheng, Y., Li, J., Li, E., Gong, Y., Yuan, Q., Ge, H., & Xiao, X. (2026). Integrated multi-omics decodes the AKI-to-CKD transition: from ensemble discovery to structure-guided translational targeting. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08950-0

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08950-0

Keywords: acute kidney injury, chronic kidney disease, proteomics, UK Biobank, machine learning, Mendelian randomization, galectin-3, LGALS3, renal fibrosis, single-cell RNA sequencing, structural biology, biomarkers