Prognostic implications of unbiased molecular categorization in kidney disease

In this commentary, a novel approach to the reclassification of chronic kidney disease is reviewed. In the revisited study, the investigators identify 4 distinct subtypes of kidney disease derived from an unbiased self-organizing map of transcriptomic data from kidney biopsy samples. These molecular subtypes then are characterized by biologic cell processes, clinical and histopathologic features, urinary proteomics, and disease progression. The strengths and limitations of the self-organizing map approach are assessed; the prognostic, diagnostic, and therapeutic implications are considered briefly. In this commentary, a novel approach to the reclassification of chronic kidney disease is reviewed. In the revisited study, the investigators identify 4 distinct subtypes of kidney disease derived from an unbiased self-organizing map of transcriptomic data from kidney biopsy samples. These molecular subtypes then are characterized by biologic cell processes, clinical and histopathologic features, urinary proteomics, and disease progression. The strengths and limitations of the self-organizing map approach are assessed; the prognostic, diagnostic, and therapeutic implications are considered briefly. For decades, oncologists have used molecular expression derived from tumor biopsy samples to establish diagnoses, predict prognosis, and select optimal therapy.1Early Breast Cancer Trialists' Collaborative group (EBCTCG)Trastuzumab for early-stage, HER2-positive breast cancer: a meta-analysis of 13 864 women in seven randomised trials.Lancet Oncol. 2021; 22: 1139-1150Abstract Full Text Full Text PDF PubMed Scopus (146) Google Scholar The renal scientific community finally is emulating these approaches, generating molecular information from kidney biopsy specimens.2de Boer I.H. Alpers C.E. Azeloglu E.U. et al.Rationale and design of the Kidney Precision Medicine Project.Kidney Int. 2021; 99: 498-510Abstract Full Text Full Text PDF PubMed Scopus (84) Google Scholar, 3Hogan J.J. Owen J.G. Blady S.J. et al.The feasibility and safety of obtaining research kidney biopsy cores in patients with diabetes: an interim analysis of the TRIDENT study.Clin J Am Soc Nephrol. 2020; 15: 1024-1026Crossref Scopus (10) Google Scholar, 4Lake B.B. Menon R. Winfree S. et al.An atlas of healthy and injured cell states and niches in the human kidney.Nature. 2023; 619: 585-594Crossref PubMed Scopus (65) Google Scholar The integration of clinical and histopathologic data into the molecular sphere remains a challenge. Clinical studies in kidney disease principally enroll subjects and assess outcomes based on information from the clinical domain. For example, the inclusion criteria and outcomes of the Supportive Versus Immunosuppressive Therapy of Progressive IgA nephropathy (STOP-IgAN) study depend on clinical criteria such as glomerular filtration rate (GFR) and proteinuria, without regard to the Oxford-MEST (mesangial hypercellularity, endocapillary hypercellularity, segmental glomerulosclerosis, and tubular atrophy/interstitial fibrosis) histologic categorizations.5Rauen T. Eitner F. Fitzner C. et al.Intensive supportive care plus immunosuppression in IgA nephropathy.N Engl J Med. 2015; 373: 2225-2236Crossref PubMed Scopus (482) Google Scholar The integration of histopathologic and clinical information subsequently came to enhance risk prediction.6Barbour S.J. Coppo R. Zhang H. et al.Evaluating a new international risk-prediction tool in IgA nephropathy.JAMA Intern Med. 2019; 179: 942-952Crossref PubMed Scopus (243) Google Scholar Only recently have molecular data further refined our understanding of seemingly benign IgA nephropathy.7Rivedal M. Mikkelsen H. Marti H.-P. et al.Glomerular transcriptomics predicts long term outcome and identifies therapeutic strategies for patients with assumed benign IgA nephropathy.Kidney Int. 2024; 105: 717-730Abstract Full Text Full Text PDF PubMed Scopus (2) Google Scholar As molecular data accumulate, dimensionality reduction will prove essential for integration into clinical studies. After all, the human genome contains over 20,000 protein-coding genes, and correcting for each as a covariate within a clinical study is nigh impossible. In this issue of Kidney International, Reznichenko, Kretzler, and their colleagues take an important step forward in recategorizing chronic kidney disease (CKD) according to molecular subtypes.8Reznichenko A. Nair V. Eddy S. et al.Unbiased kidney-centric molecular categorization of chronic kidney disease as a step towards precision medicine.Kidney Int. 2024; 105: 1263-1278Google Scholar Their vision is to move beyond functional classifications such as CKD stage 3 or 4, or etiologic classifications, such as minimal change disease or diabetic kidney disease. Instead, they chart a future for nephrology wherein molecular features drive disease classification without a priori knowledge or bias. The molecular subtypes of CKD uncovered in this study independently augment prediction of CKD progression, superseding clinical and histopathologic features such as proteinuria and interstitial fibrosis. To achieve their vision, tubulointerstitial datasets from individuals with kidney disease and from healthy living kidney donors were culled and merged from multiple publicly available cohorts: the European Renal cDNA Bank, the Nephrotic Syndrome Study Network (NEPTUNE), and the Clinical Phenotyping Resource and Biobank Core (C-PROBE). Subjects were categorized based on a self-organizing map, which helped surface the unifying cell-based disease processes in each unbiased group. What is a self-organizing map? Although this artificial neural-network tool was conceived in the 1980s by Kohonen,9Kohonen T. Essentials of the self-organizing map.Neural Netw. 2013; 37: 52-65Crossref PubMed Scopus (1054) Google Scholar its application to kidney disease is novel. For every subject added to an analysis (or analytical space), that subject is assigned a weight vector that summarizes all available information. In the case of Reznichenko et al.,8Reznichenko A. Nair V. Eddy S. et al.Unbiased kidney-centric molecular categorization of chronic kidney disease as a step towards precision medicine.Kidney Int. 2024; 105: 1263-1278Google Scholar this vector summarizes the expression of thousands of transcripts in each patient's kidney biopsy specimen. During training of the map, a wide variety of subjects' weight vectors are included, hopefully representing the diversity of the patient pool. Subsequently, additional subjects are selected at random and mapped in the analytical space. The process is competitive in that each subject is assessed for similarity and either will be mapped near other subjects, reinforcing the unique nature of an existing group, or will establish a new direction on the map, potentially merging 2 groups or forming a new group (Figure 1a). After multiple iterations, the map grows and facilitates self-selection into different groups without existing biases. Reznichenko et al. used the self-organizing map to identify 4 groups, as follows: CKD-blue, characterized by diffuse downregulation of canonical cellular processes; CKD-gold, with upregulation of immune cell and matrix formation processes; CKD-olive, with upregulation of mitochondrial and peroxisome processes; and CKD-plum, characterized by downregulation of ion transport (Figure 1b). Strengths of the study are its ability to do the following: (i) demonstrate independent predictive value of molecular subtypes above histopathologic and clinical features; (ii) link the molecular signature of a biopsy specimen with noninvasive urinary biomarkers; and (iii) use a relevant control. Despite these strengths, the study should be praised more for its vision than its final output. Certain limitations portend that translating these 4 molecular CKD subtypes into clinical practice is an unlikely prospect. First, few nephrologists would find surprising the fact that patients with lower GFR and higher interstitial fibrosis and tubular atrophy will have progressive CKD. Such is the plight of the CKD-gold subjects, a group whose signature is enriched for immune cell activation and extracellular matrix formation, but who also had the lowest level of kidney function and the highest level of scarring at the time of biopsy. The CKD-gold status held its independence within a fully adjusted model to predict faster CKD progression, but these patients were already at high risk. One interpretation is that the CKD-gold signature enables us to recategorize the minority of subjects who otherwise might have been misclassified by GFR, proteinuria, and fibrosis alone (Figure 1c). The scientific community will refine these categories over time with other data types, but this need for refinement does not undermine the important advancements in approach and framework outlined in this study. A second consideration is the bulk nature of the molecularly derived signatures. The use of the microdissected tubulointerstium is a reasonable choice to demonstrate proof of principle. However, gene signatures will vary based on the underlying tissue composition. We cannot rule out the possibility that a portion of the ion transport effects observed in CKD-plum subjects are related to the underlying proportion of cortex or medulla within the samples. Further, processes involving microdissection are unlikely to transition to a clinical laboratory improvement amendments (CLIA) output, given the technical expertise required. As multiplexed spatial and single-cell technologies mature, these will afford the categorization of signatures for distinct cell types or functional tissue units such as glomeruli, vessels, or the medullary interstitium, complementing the cortical tubulointerstitial groups defined in this study. Early signatures in some forms of CKD are likely to be sensitive to regional changes in the kidney, such as arterionephrosclerosis in vessels or sickle cell nephropathy in the medulla. One decision point for the community is when to group or split cell types and disease states based on a priori knowledge, and then when to subsequently apply unbiased machine learning. Splitting feels anathemic to the unsupervised philosophy espoused in this study, yet a goal that remains incumbent upon us is to confidently distinguish biological signals from end-organ phenomena, such as interstitial fibrosis and tubular atrophy, and the noise of technical artifacts. In this study, subjects with heterogenous disease etiologies were grouped. The result was to identify a progressive CKD-gold group defined by broad cross-cutting features such as a lower baseline GFR, a higher level of inflammation, and fibrosis. The CKD-blue, CKD-olive, and CKD-plum groups were molecularly distinct, but therapeutic implications were not identified. Although samples were aggregated from multiple consortia, a larger sample size is required to tease out more-subtle molecular features with clinical relevance in these 3 groups. As the sample size grows, an exciting process will be disentangling the molecular features predictive of progression within groups such as CKD-blue, CKD-olive, and CKD-plum. To offset study limitations, the investigators link the transcriptomic signatures to urinary proteomics, identifying a set of markers associated with molecular subtype and progression. The liquid biopsy may prove to be the future of molecular subtyping. Going forward, much work is needed to differentiate acute and chronic signals, as well as those which vary under normal physiological conditions. For example, the expression of aquaporin 2 (AQP2) or uromodulin (UMOD) changes based on the fasting or feeding state. Until these parameters are fully understood using repeated measures, the field is apt to focus on maximizing the value of a biopsy with molecular information, as is demonstrated by Reznichenko et al., rather than replacing the biopsy outright. Modeling the success of our oncologist counterparts, perhaps the tissue sample provides context to understand baseline urinary biomarker expression, and their change over time facilitates the measurement of therapy response, in a manner safer than serial biopsy procedures. An important strength of this study is the use of living kidney donor biopsy samples because they are processed identically to diagnostic cores. Nephrectomy samples may have tumor effects or prolonged hypoxia, which alter their signature. One unaddressed issue is the heterogeneity of "normal" healthy samples. Aging, senescence, genetic background, sex, medications, comorbidities, and diet all contribute to the standard deviation of reference samples. This study predominantly pulls subjects from North American and European cohorts with few demographic details. Although the momentum within the practice of medicine is to deemphasize race, a diverse patient pool matters for establishing the control comparison, and race is not referenced in this study. The caveats to this study do not diminish its importance in establishing a pathway forward for the field. We should periodically question what we know as technological advances allow us to view kidney disease through new prisms. New therapeutics for CKD and glomerular disease finally have emerged after a long drought. The gauntlet laid for the kidney community is to navigate beyond prognostication and generate evidence to support molecular-guided therapy selection. Diagnostic subtypes of disease should be more than molecularly distinct; they should hold prognostic significance or drive clinical trials. For example, the prominence of human epidermal growth factor receptor 2 (HER2)-positive molecular subtyping rose, once the therapeutic utility of trastuzumab was established in breast cancer.1Early Breast Cancer Trialists' Collaborative group (EBCTCG)Trastuzumab for early-stage, HER2-positive breast cancer: a meta-analysis of 13 864 women in seven randomised trials.Lancet Oncol. 2021; 22: 1139-1150Abstract Full Text Full Text PDF PubMed Scopus (146) Google Scholar Just as the human genome project heralded advancements in pharmacogenomics, hopes are high that the Kidney Precision Medicine Project,2de Boer I.H. Alpers C.E. Azeloglu E.U. et al.Rationale and design of the Kidney Precision Medicine Project.Kidney Int. 2021; 99: 498-510Abstract Full Text Full Text PDF PubMed Scopus (84) Google Scholar the Human BioMolecular Atlas Program,4Lake B.B. Menon R. Winfree S. et al.An atlas of healthy and injured cell states and niches in the human kidney.Nature. 2023; 619: 585-594Crossref PubMed Scopus (65) Google Scholar the Transformative Research in Diabetic Nephropathy (TRIDENT) study,3Hogan J.J. Owen J.G. Blady S.J. et al.The feasibility and safety of obtaining research kidney biopsy cores in patients with diabetes: an interim analysis of the TRIDENT study.Clin J Am Soc Nephrol. 2020; 15: 1024-1026Crossref Scopus (10) Google Scholar and others will identify new targets and advance existing therapy through pharmacotranscriptomics and pharmacoproteomics. The armamentarium of kidney disease treatments is growing; approaches such as those of Reznichenko et al. will allow us to distill molecular signatures and advance therapeutic algorithms beyond GFR and proteinuria. The author declared no competing interests. MTE is supported by National Institutes of Health/National Center for Complementary and Integrative Health R01AT011463. Unbiased kidney-centric molecular categorization of chronic kidney disease as a step towards precision medicineKidney InternationalVol. 105Issue 6PreviewCurrent classification of chronic kidney disease (CKD) into stages using indirect systemic measures (estimated glomerular filtration rate (eGFR) and albuminuria) is agnostic to the heterogeneity of underlying molecular processes in the kidney thereby limiting precision medicine approaches. To generate a novel CKD categorization that directly reflects within kidney disease drivers we analyzed publicly available transcriptomic data from kidney biopsy tissue. A Self-Organizing Maps unsupervised artificial neural network machine-learning algorithm was used to stratify a total of 369 patients with CKD and 46 living kidney donors as healthy controls.