Predicting kidney graft survival with a machine learning model based on for-cause biopsy transcriptomics

Kidney transplantation significantly improves outcomes in end-stage renal disease, however, long-term graft survival remains suboptimal. Traditional prognostic models often yield limited predictive power. Given the routine use of for-cause biopsies in clinical practice, we aimed to develop and compare ML models to predict graft survival using gene expression profiles from indication biopsies. Data from for-cause renal biopsies were collected from six cohorts in the Gene Expression Omnibus: GSE21374 (n = 282), GSE48581 (n = 306), GSE36059 (n = 411), GSE50058 (n = 101), GSE72925 (n = 168) and GSE25902 (n = 24). Differential expression and Cox regression analysis were applied to identify prognostic genes. These were used to train and validate 117 machine learning models. A total of 11 genes associated with graft loss were identified and used to train ML models. The Gradient Boosting Machine (GBM) achieved the highest C-index (> 0.85) and accurately stratified patients by graft survival. External validation in diagnostic tasks across four independent cohorts confirmed the model's value of overall rejection (AUCs 0.760-0.826). Patients classified as High-risk showed reduced graft survival, enrichment of pro-inflammatory immune pathways and marked differences in immune cell composition. We developed and validated a transcriptome-based GBM model that accurately predicts graft survival using for-cause biopsy samples. The model demonstrated strong prognostic and diagnostic performance across multiple independent cohorts and reflects key immunological mechanisms of rejection. Its compact gene set offers practicality, supporting its potential integration into routine transplant care.