Abstract: SA-PO1064
Application of Machine Learning for Repertoire Analysis in Antibody-Mediated Rejection (ABMR) of Kidney Transplantation
Session Information
- Transplantation: Clinical - II
November 04, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 2102 Transplantation: Clinical
Author
- Yang, John J., Korea University Guro Hospital, Seoul, Korea (the Republic of)
Background
The adaptive immune receptor repertoire (AIRR)-sequencing analyzes the individual repertoire at sequence level. Here we compare AIRRs of kidney transplant recipients using repertoire analyses and machine learning, regarding ABMR.
Methods
Total of 30 repertoires (9 ABMR and 21 No rejection) were analyzed. Bulk gDNA (2μg) was sequenced at 500,000 target read-depth (equivalent of 10,000 B lymphocytes). The MiXCR tool (v4.3.2) was used for preprocessing and normalization. Repertoire analysis using the immunarch R package (v1.0.0) included clonality, diversity and V(D)J gene usage analyses. Machine learning by ImmuneML (v2.2.4) used four algorithms of KNN, SVM, RF and logistic regression. Using 3-mer aminoacids encoding and ABMR status as classifiers, the performances of ML algorithms were assessed by AUC.
Results
From repertoire analyes, decreased number of clonotypes (P=0.008) and decreased diversity (P=0.02) were seen in ABMR. There were differences in V gene utilization, from which IGHV1-2 and IGHV1-8 were decreased in ABMR group, while IGHV3-30 usage was higher (P=0.026, 0.014 and 0.018). The AUC of ML algorithms KNN, LR, RF and SVM were 0.77±0.11, 0.8±0.09, 0.81±0.11 and 0.75±0.06. Analysis of 3mer correlated V gene usages, from which NEI, TEC, WAK and WAR were statistically significant 3mers after Bonferroni correction (P<0.00000625).
Conclusion
Reproducible and clinically applicable AIRR-seq workflow was implemented using IVD product. Repertoire analysis was capable of demonstrating differences from kidney transplant recipients. ML models using different algorithms generated overall favorable performances in classifying ABMR status, indicating that direct interpretation of AIRR-seq data is possible.
Differences in number of clonotypes (A), diversity (B) between ABMR and No rejection groups. The performances of ML algorithms (C).