Bioinformatics strategies and biomarker refinement using high-throughput transcriptome data in transplantation

Renal transplantation is the treatment of choice for kidney failure, but most transplants fail prematurely and barely half of recipients survive with a functioning graft for more than a decade. Strategies to induce operational tolerance are therefore at the cutting edge of transplant research, exploiting the dynamic plasticity of the immune system to recapitulate neonatal ontogeny and permit gradual withdrawal of immune suppression. We have shown that whole blood gene expression is profoundly altered in uremia and following graft implantation, and that changes in the blood transcriptome are characteristic of rejection injury. But deriving simple, robust and parsimonious classifiers presents challenges, and pre-filtering methods of varying stringency have been proposed to enhance predictive accuracy.

We re-analyzed our previous data documenting transcriptome changes in rejection using a case-control design to compare analytical strategies in subjects with or without biopsy-proven rejection. Five pre-filtering methods and eight multivariate classification methods were evaluated using multiple partition nested cross-validation to obtain unbiased estimation of classifier performance.

The most permissive strategy identified 800 unique genes and the most restrictive identified 71 nested genes differentially expressed in rejectors of which 31%-45% were downregulated and 55%-69% were upregulated, reflecting neutrophil degranulation, regulated necrosis, programmed cell death, pyroptosis, interleukin signaling and other functional pathways. Of the ten most common genes or probe-sets over all panels, nine were increased in BCAR.

No individual combination of methods presented superior performance among all those considered although the PAM and XGBoost classifiers were more resistant to over-fitting. It is therefore advisable to apply multiple analytical combinations and compare performances in transcriptome analysis. In limited resource situations, evaluation of at least two complementary classifiers with fixed pre-filter and ranking methods is advisable. For small panel size constraints, feature-selecting methods like PAM or EN could be considered.