Abstract: TH-PO0940
Unsupervised Learning of Five-Biomarker Data for Patient-Specific Kidney Rejection Monitoring
Session Information
- Transplantation: Clinical - Glomerular Diseases, Infections, and Rejection
November 06, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 2102 Transplantation: Clinical
Authors
- Singla, Akhil, Northwestern University, Evanston, Illinois, United States
- Chen, Kenny, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Robelly, Blade H, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Rebello, Christabel, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Lantz, Connor, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Zhao, Lihui, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Lacombe, Ronnie, Eurofins Transplant Genomics, Lenexa, Kansas, United States
- Sinha, Rohita, Eurofins Transplant Genomics, Lenexa, Kansas, United States
- Hanna, Genevieve S, Eurofins Transplant Genomics, Lenexa, Kansas, United States
- Kleiboeker, Steven, Eurofins Transplant Genomics, Lenexa, Kansas, United States
- Park, Sookhyeon, Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
- Mehrotra, Sanjay, Northwestern University, Evanston, Illinois, United States
- Friedewald, John J., Northwestern University Feinberg School of Medicine, Chicago, Illinois, United States
Background
Population-level supervised models assume a single “average” rejection signature and ignore recipient heterogeneity. We propose unsupervised clustering using five biomarkers—blood gene-expression profiling (GEP), torque teno virus (TTV), donor-derived cell-free DNA (dd-cfDNA), and urinary CXCL9/CXCL10—to derive patient subgroups and train subgroup-tailored prediction models for subclinical acute rejection (subAR) versus stable function (TX).
Methods
We clustered the five biomarker data from the CTOT-08 study (226 recipients, 462 biopsy-paired samples; 121 subAR, 341 TX). Clusters dominated (>85%) by a single phenotype were classified as that phenotype; the remaining clusters employed stepwise logistic models incorporating biomarkers plus recipient factors (race, age, ethnicity, sex, transplant type, induction therapy, and days post-transplant). Performance was validated using 54 subAR and 92 TX samples from 136 Northwestern University Mini-Kidney Biorepository subjects.
Results
Five clusters emerged: Cluster 1 (n=106; 405 average days post-transplant) had 11% subAR (7.5% AMR+1.9% cellular+1.9% mixed); Cluster 2 (n=100; 466 days) had 51% subAR (20% AMR+20% cellular+11% mixed); Cluster 3 (n=63; 157 days) had 4.8% subAR (3.2% AMR+1.6% mixed); Cluster 4 (n=106; 471 days) had 32% subAR (13% AMR+12% cellular+7% mixed); Cluster 5 (n=87; 233 days) had 24% subAR (4.5% AMR+15% cellular+4.5% mixed) cases. Clusters 1 and 3 were TX-dominant (89% and 95% TX). For the other clusters, stepwise models retained CXCL10/Cr+dd-cfDNA+induction therapy (Cluster 2), dd-cfDNA+GEP (Cluster 4), and GEP+dd-cfDNA+CXCL9/Cr (Cluster 5) as predictors. This cluster-specific workflow achieved accuracy 0.77, sensitivity 0.68, specificity 0.81, NPV 0.88, PPV 0.55, and AUC 0.8. The validation cohort showed similar results.
Conclusion
Clustering uncovers biologically meaningful subgroups and highlights the importance of patient heterogeneity for rejection monitoring, compared to a single global model.
Decision-making workflow.
Funding
- NIDDK Support – Eurofins-Viracor