Transplant Biopsy Assessment in 21st Century

The Achilles heel of solid organ transplantation is the alloimmune reaction to foreign tissue, or rejection. Two tools, serum creatinine and biopsy histology, are traditionally used to assess for allograft function, viability, and rejection. Our u nderstanding of rejection is based on histologic features of immune cell infiltration in various regions of the allograft: capillaries, tubules, interstitium, and subendothelium. More recently, assessment of capillary complement deposition and circulating anti-donor antibodies have been added to diagnostic algorithms. On the basis of these findings, defined by the Banff criteria, we make diagnoses of T cell–mediated rejection, antibody-mediated rejection, borderline rejection, or no rejection.¹ However, it is acknowledged that these categories do not fully accommodate for clinical or immunologic heterogeneity between patients. There is also likely a spectrum of immunologic processes across these broad categories of rejection. The histologic approach is also limited by its subjective nature and interobserver variability.

The study of gene expression from biopsy tissue may offer a more objective way to assess the state of an allograft. In recent decades, the transcriptomes of hundreds of biopsy specimens have been studied, which has helped to improve diagnostics and inform our understanding of rejection.^2–5 The ability to study thousands of genes expressed in a sample allows us to identify cellular processes and pathways turned on or off relative to other biopsy samples studied. Furthermore, patterns of gene expression can be defined and trained on traditional categories of rejection or specific histologic features. By using such an approach, probabilities can be assigned to each biopsy specimen for each traditional category of rejection or histologic feature. This removes the problem of interobserver variability. However, as pointed out by Buscher et al. in this edition of JASN,⁶ this approach is still dependent on histology to train the genetic data and categorize each biopsy specimen. Therefore, the known limitations of histology will affect all downstream analyses.

Buscher et al. describe a histology-independent approach to harnessing large transcriptomic datasets to define allograft phenotypes that can better predict outcomes and help uncover candidate biomarkers of disease. They used a number of publicly available, microarray-based, transplant biopsy specimen datasets to test and validate their bioinformatics approach. First, the authors created a gene signature, KID9plus3. This was composed of differentially expressed gene sets that define nine immune cell types (on the basis of isolated cells in vitro) previously compiled by others (“KID9”). To this, they added a gene signature from kidney tubules, glomeruli, and human fibroblasts, also publicly available (“plus3”). They used this KID9plus3 gene signature to deconvolute microarray data from each biopsy specimen studied using the Cibersort tool.⁷ In other words, they estimated the proportion of each cell type present in each biopsy sample. Next, the authors used PRESTO (a predictive t-stochastic neighbor embedding tool for -omics) to perform a gene network analysis of 2296 variable genes from a dataset of 411 biopsy specimens. This tool uses data variation to identify coregulated networks by unsupervised dimensionality reduction across a multiomic dataset.⁸(preprint) The selected genes clustered into gene networks using k-means and were plotted using t-distributed stochastic neighbor embedding. This defined 13 different networks of gene expression. Finally, the authors generated a matrix containing samples (411 biopsy and eight nephrectomy samples) defined by the PRESTO network and KID9plus3 composition (expression parameter[gene network activity and KID9plus3 cell-type estimations]×sample). They then used another computational tool, Phenograph,⁹ to cluster biopsy samples on the basis of phenotypic similarity using this matrix as input. Phenograph creates a nearest neighbor graph in high dimensional space, and then clusters phenotypically similar subpopulations using the graph partition algorithm Louvain. This generated seven kidney transplant phenotypes (A–G) on the basis of gene networks and cell-type proportions. The authors posit these seven phenotypes represent the full spectrum of diagnoses present in the datasets studied. Of note, the test set included all categories of rejection and nonrejecting samples.

The authors then sought to determine whether a phenotype score assigned to each biopsy sample could predict allograft survival. They performed this part of the analysis using two independent datasets of biopsy sample transcriptomes. Using Cibersort, they calculated a phenotype enrichment score for each biopsy sample. They found phenotypes B, C, and E were associated with poor allograft outcome; and phenotypes D, F, and G were associated with good allograft outcome. Interestingly, the molecular phenotypes performed better at predicting poor outcomes than histology alone. The authors characterized each phenotype based on gene network and cell-type composition. For example, phenotype B (poor allograft outcome phenotype) is associated with fibroblasts/M2 macrophages and extracellular matrix/collagen generation gene network (network 4). Lastly, the authors sought to determine if their approach can be used to uncover candidate biomarkers of disease. They filtered genes differentially expressed in both poor outcome phenotypes and network 4. They then ranked these genes by M2 macrophage abundance and identified lysyl oxidase like 2 (LOXL2). LOXL2 + macrophage number (assessed by immunofluorescence) positively correlated with biopsy specimen interstitial fibrosis and tubular atrophy. Furthermore, the presence of LOXL2 + macrophages in non-rejection biopsy samples predicted worsening interstitial fibrosis and tubular atrophy and increases in creatinine.

This study by Buscher et al. nicely demonstrates how high-dimensional genomics data can be harnessed to better predict transplant outcome and provide mechanistic insights. However, there remain challenges when interpreting computational analyses of large genomic datasets. Results can vary depending on the computational algorithms used and variable definitions. For example, the number of clusters generated by the k-means algorithm is dependent on the value chosen for “k.” Therefore, it is critical that noncomputational experiments are performed to validate findings. The importance of data sharing is also highlighted in this study because it includes multiple datasets generated by other investigators. The kidney biopsy remains a useful tool for both the clinician and researcher. Transplant rejection is complex and continues to prove difficult to navigate for the transplant clinician. A better understanding of rejection will require the continued study of biopsy tissues. The development of new tools, such as the “liquid biopsy,” will only be possible once a complete understanding of rejection is achieved. We are now in a very exciting era where machine learning intersects with genomics and is driving new discoveries at a rapid pace. The application of such methods to transplantation research will undoubtedly benefit future patients.

Disclosures

A. Malone reports having ownership interest in AstraZeneca, and serving as a scientific advisor for, or member of, CareDx.

Funding

None.