Introduction: This study aims to assess the ability of machine learning (ML) classifiers which based on peripheral blood transcriptome data, a timely and non-invasive approach, for the classification of patients with kidney allograft acute rejection (AR) and with stable function (STA) posttransplant.
Materials and Methods: 90 AR and 72 STA patient transcriptome and pathological data updated on Gene Expression Omnibus were included. Batch effect was removed by Surrogate Variable analysis[1]. Then 3 kinds of supervised machine learning (ML) model were performed. Support Vector Machine (SVM) classifiers were constructed using the space of component loadings matrix of whole transcriptome principal component analysis (PCA), and cross validated by leave-one out method. Logistic regression classifier was based on 8 differentially expression RNAs (DEGs) identified by Empirical Bayes method. Step regression was used to assay key molecular signatures in the 8 DEGs. We evaluated the accuracy, sensitivity, and specificity of classifiers which based on whole transcriptome data, logistic regression, and step regression.
Results: We found that a classifier combined with SVM and Principal Component analysis (PCA-SVM) could better distinguish patients with AR and STA condition, with the accuracy, sensitivity and specificity of 100%. Furthermore, 8 RNA molecular signatures (XIST, ITGA4, RPS4Y1, FAM3C, PRKACB, SLC8A1, SGPP and HLA-DRB4) identified from peripheral blood could also be used to classify AR and STA patients (accuracy 87.9%, sensitivity 85.0%, specificity 92.3%, and the area under the curve (AUC) is 0.912. Step regression revealed that using 3 mRNA signatures (SLC8A1, SGPP1 and RPS4Y1) was able to detect AR patients with similar performance (accuracy 84.8%, sensitivity 80.0%, specificity 92.3%). The AUC of 3 mRNA signature classifier is 0.904.
Discussion: Small amount changes of regulatory genes may have a greater impact on the biological effect. PCA have advantage in the preservation of original information, benefiting to preserve the effect to which regulatory genes contribute to AR, which partly contributed to the high accuracy of PCA-SVM classifier. HLA-DRB4, expressed in antigen presenting cells, presenting peptides of extracellular proteins, plays a crucial role in immune system, and associated with immune-related disease, such as type 1 diabetes and Anti-LGI1 encephalitis[2,3]. RPS4Y1, one of minor histocompatibility antigen genes, have shown strongly correlated with the kidney acute rejection in gender-mismatched transplantation[4]. The relationship between HLA-DRB4, SLC8A1 and kidney acute rejection has not been reported, and further research is needed
Conclusions: Supervised ML classifiers base on peripheral blood transcriptome data may be potential measures to monitor kidney allograft status early, accurately and dynamically.