Machine Learning-Based Urine Peptidome Analysis to Predict and Understand Mechanisms of Progression to Kidney Failure
Ziad A. Massy, Oriane Lambert, Marie Metzger, Mohammed Sedki, Adeline Chaubet, Benjamin Breuil, Acil Jaafar, Ivan Tack, Thao Nguyen‐Khoa, Mélinda Alvès, Justyna Siwy, Harald Mischak, Francis Verbeke, Griet Glorieux, Yves-Édouard Herpe, Joost P. Schanstra, Bénédicte Stengel, Julie Klein, Natália Alencar de Pinho, Carole Ayav, Dorothée Cannet, Christian Combe, Jean‐François Deleuze, Denis Fouque, Luc Frimat, Yves-Édouard Herpe, Christian Jacquelinet, Maurice LAVILLE, Sophie Liabeuf, Ziad A. Massy, Christophe Pascal, Bruce Robinson, Roberto PECOITS-FILHO, Joost P. Schanstra, Bénédicte Stengel, Céline Lange, Marie Metzger, Élodie Speyer
Abstract
Introduction: The identification of patients with chronic kidney disease (CKD) at risk of progressing to kidney failure (KF) is important for clinical decision-making. In this study we assesed whether urinary peptidome (UP) analysis may help classify patients with CKD and improve KF risk prediction. Methods: The UP was analyzed using capillary electrophoresis coupled to mass spectrometry in a case-cohort sample of 1000 patients with CKD stage G3 to G5 from the French CKD-Renal Epidemiology and Information Network (REIN) cohort. We used unsupervised and supervised machine learning to classify patients into homogenous UP clusters and to predict 3-year KF risk with UP, respectively. The predictive performance of UP was compared with the KF risk equation (KFRE), and evaluated in an external cohort of 326 patients. Results: More than 1000 peptides classified patients into 3 clusters with different CKD severities and etiologies at baseline. Peptides with the highest discriminative power for clustering were fragments of proteins involved in inflammation and fibrosis, highlighting those derived from α-1-antitrypsin, a major acute phase protein with anti-inflammatory and antiapoptotic properties, as the most significant. We then identified a set of 90 urinary peptides that predicted KF with a c-index of 0.83 (95% confidence interval [CI]: 0.81-0.85) in the case-cohort and 0.89 (0.83-0.94) in the external cohort, which were close to that estimated with the KFRE (0.85 [0.83-0.87]). Combination of UP with KFRE variables did not further improve prediction. Conclusion: This study shows the potential of UP analysis to uncover new pathophysiological CKD progression pathways and to predict KF risk with a performance equal to that of the KFRE.