Multiple omics-based machine learning reveals peripheral blood immune cell landscape during acute rejection of kidney transplantation and constructs a precise non-invasive diagnostic strategy.

Abstract

Kidney transplantation is the optimal treatment for end-stage renal disease (ESRD), but acute rejection (AR) remains a major factor affecting graft survival and patient prognosis. Currently, renal biopsy is the gold standard for diagnosing AR, but its invasiveness limits the application of dynamic monitoring. This study aims to analyze changes of immune cell and gene expression in the peripheral blood of AR recipients and construct a non-invasive AR diagnosis strategy. All datasets were downloaded from the GEO database. Single cells were annotated based on the expression profiles of surface proteins and changes of immune cell in the peripheral blood of AR and stable transplant (STA) recipients were compared. The high-dimensional weighted gene co-expression network analysis (hdWGCNA) algorithm was used to analyze gene modules related to AR and to screen out hub genes by integrating bulk RNA-Seq. Based on hub genes, consensus clustering stratified recipients into two sub-clusters and a non-invasive AR diagnostic model was constructed using Convolutional Neural Networks (CNNs). Additionally, we also constructed a predictive model for long-term graft survival through combinations of 111 machine learning algorithms and validated the expression of hub genes in the rat AR model. AR recipients had higher abundance of memory B cells, effector memory T cells, terminally differentiated effector memory T cells (TEMRA), and NK T cells but lower Tregs in the peripheral blood compared to STA recipients. Through hdWGCNA analysis, we identified gene modules associated with these immune cells and screened out four hub immune-related genes (TBX21, CX3CR1, STAT1, and NKG7) after integrating bulk RNA-Seq. Based on these hub genes, recipients can be stratified into two sub-clusters with distinct clinical outcomes and biological characteristics. We also innovatively constructed a non-invasive AR diagnostic model using CNNs, which can effectively address the issues caused by batch effects and demonstrate a high diagnostic accuracy. Besides, the predictive model for long-term graft survival constructed using the RSF algorithm can divided recipients into high- and low-risk groups, with significantly higher rates of AR and long-term graft failed in the high-risk group. This study successfully identified immune cell subsets and hub genes related to AR. Based on hub genes, we successfully identified two distinct molecular sub-clusters of kidney transplant recipients, and constructed a non-invasive diagnostic model for AR and a predictive model for long-term graft survival. These models offer new tools for precise diagnosis and prognosis in kidney transplantation and may advance precision medicine.