Molecular Profiling Improves Diagnoses of Rejection and Infection in Transplanted Organs
The monitoring of transplanted hearts is currently based on histological evaluation of endomyocardial biopsies, a method that is fairly insensitive and that does not always accurately discriminate between rejection and infection in the heart. Accurate diagnosis of rejection and infection is absolutely crucial, however, as the respective treatments are completely different. Using microarrays, we analyzed gene expression in 76 cardiac biopsies from 40 heart recipients undergoing rejection, no rejection, or Trypanosoma cruzi infection. We found a set of genes whose expression patterns were typical of acute rejection, and another set of genes that discriminated between rejection and T cruzi infection. These sets revealed acute rejection episodes up to 2 weeks earlier, and trypanosome infection up to 2 months earlier than did histological evaluation. When applied to raw data from other institutions, the 2 sets of predictive genes were also able to accurately pinpoint acute rejection of lung and kidney transplants, as well as bacterial infections in kidneys. In addition to their usefulness as diagnostic tools, the data suggest that there are similarities in the biology of the processes involved in rejection of different grafts and also in the tissue responses to pathogens as diverse as bacteria and protozoa.
The health of transplanted hearts is most often assessed by histological analysis of small biopsies that are routinely taken at least once a month for the first year after transplant. Although this is still considered the gold standard for diagnosis of acute rejection, a comparison between histological diagnosis of biopsies and of the whole organ at autopsy1,2 showed that the overall sensitivity of histological evaluation of endomyocardial biopsy is only ≈70% and that a number of important rejections can be missed.
In addition to rejection, infection is also a serious complication in the posttransplant period,3 and histology does not discriminate very well between a rejection response and the type of immune response involved in fighting an infection in the heart.4 Nevertheless, discrimination is essential because the treatment for the 2 conditions is quite different. A patient undergoing acute rejection needs an increase in the immunosuppressive regimen to dampen the damaging immune response, whereas a patient experiencing infection needs specific treatment for the infectious agent, along with a decrease in immunosuppression so as to allow the immune system to mount a strong response against the pathogen. Each year, in spite of heroic efforts by transplant physicians, a number of transplant patients die who might have been saved by a better diagnostic technique.
In this study, we tested whether molecular profiling by microarray analysis of gene expression patterns might offer a more accurate picture. As a model transplant, we studied cardiac allografts, and, as a model infection, we chose Chagas disease (infection with Trypanosoma cruzi). In Latin America, a large proportion of heart recipients receive transplants because of chronic heart failure resulting from this infection.5 Relapse of Chagas disease is consequently one of the most frequent posttransplant complications. We analyzed mRNA amplified from biopsy samples that had been taken from heart recipients with cases of rejection, no rejection, or infection.
We found that there are indeed specific molecular profiles that discriminate between patients undergoing rejection, no rejection, or infection. When tested on the raw data from other institutions, these profiles also turned out to be useful for analysis of kidney and lung transplants. In addition, although Chagas infection is rare in North America,6 the approach turned out to be useful for other infections that are more universally encountered.
Materials and Methods
Patients and Samples
All except 1 cardiac allograft recipient were transplanted at São Paulo Hospital, Federal University of São Paulo, Sao Paulo, Brazil. The protocol was approved by the Ethics Committee of the Federal University of São Paulo, and informed consent was obtained from all subjects under study. All recipients were adults maintained on standard triple therapy immunosuppression. Treatment for rejection consisted of pulse therapy with methylprednisolone (1 g daily for 3 days) and/or augmentation of the oral doses of prednisone and cyclosporine.
Endomyocardial biopsies for rejection monitoring were routinely performed during the first 6 months after transplantation according to a standard schedule (weekly during the first month, every 15 days during the second month, and monthly thereafter). Additional biopsies were taken when rejection was suspected in any posttransplant period. Seventy-five percent of the biopsies comprising this study were performed within the first 3 months after transplantation and 90% within the first 6 months. Rejection was graded according to the scoring scheme of the International Society for Heart and Lung Transplantation (ISHLT), after the examination of 3 or 4 fragments of each biopsy. In the group with rejection, there were 1 biopsy with grade 1B (1R according to the new classification7), 1 biopsy with grade 2 (1R), 20 biopsies with grade 3A (2R), and 2 biopsies with grade 3B (3R).
Chagas myocarditis after transplantation was diagnosed by histological observation of T cruzi organisms surrounded by infiltrates in endomyocardial biopsies as described elsewhere.8 Toxoplasma and leprosy myocarditis were similarly diagnosed.
Figure I in the online data supplement (available at http://circres.ahajournals.org) shows a flowchart summarizing all training and test sets of samples. For microarray experiments, we used 76 biopsy samples from 40 patients. Twenty-one patients had more than 1 biopsy analyzed (mostly 2 to 3), and some were consecutive biopsies.
One or 2 fragments of the biopsy specimens were snap frozen and stored in liquid nitrogen until the time for RNA isolation. The first microarray reference RNA sample consisted of a pool of amplified RNAs from 92 nonrejection and 35 rejection biopsies. The second reference RNA sample consisted of a pool of amplified RNAs from 22 nonrejection and 10 rejection biopsies.
Preparation of RNA
Seven hundred fifty nanograms of total RNA, isolated from each biopsy using TRIzol Reagent (Invitrogen), were amplified in the first round, purified, and 750 ng of obtained antisense RNA (aRNA) were converted into double-stranded cDNA, all steps following a previously described protocol.9 To prepare RNA for the microarrays, the second round of in vitro transcription reaction consisted of 16 μL of purified cDNA, 3 μL of 50 mmol/L 5-(3-aminoallyl)-UTP, 2 μL of 75 mmol/L UTP, 4 μL of each 75 mmol/L ATP, GTP, and CTP, 4 μL of 10× reaction buffer, and 4 μL of transcription enzyme mix (all from Ambion, Austin, Tex) and was incubated at 37°C for 13 hours. After an additional 15-minute incubation with 4 μL of DNase I, aRNA was cleaned-up using RNAeasy mini kit (Qiagen, Valencia, Calif). Quality of total RNA as well as of aRNA was checked by electropherograms using Agilent 2100 BioAnalyzer (Palo Alto, Calif).
Microarrays contained 70-mer oligonucleotides corresponding to almost 14 000 human genes designed and synthesized by Qiagen/Operon (Alameda, Calif) and printed onto the slides at NIAID Microarray Facility. One hour before hybridization, the slides were treated at 42°C for 1 hour with a blocking solution containing 1% BSA, 25% 20× standard saline citrate (SSC), and 0.1% sodium dodecyl sulfate (SDS) in nuclease-free water followed by 2 washes in nuclease free water and 1 wash in isopropanol.
Five micrograms of aRNA from the biopsy samples were vacuum dried and resuspended in 4.5 μL of 0.1 mol/L sodium bicarbonate buffer (pH 8.5 to 9.0). After addition of 2.5 μL of nuclease-free water, the sample aRNA was coupled for 1 hour at room temperature with 3 μL of monoreactive Cy3 (Amersham Biosciences, Piscataway, NJ) or with 4 μL of monoreactive Cy5, each previously dissolved in 45 μL of DMSO. All test RNAs were labeled with Cy3, and reference aRNA with Cy5. The reaction was quenched by adding 6 μL of 4 mol/L hydroxylamine and incubated for additional 15 minutes. The coupled aRNA was purified using the RNAeasy minikit clean-up protocol (Qiagen) and was concentrated to 10 μL in Microcon YM-30 columns (Millipore, Billerica, Mass). The Cy3 and Cy5 labeled aRNAs were mixed, and 1 μg of poly(dA), 10 μg of human Cot I DNA, and 4 μg of yeast tRNA were added. The mixture was heated to 98°C for 3 minutes followed by addition of an equal volume of hybridization solution consisting of 50% formamide, 50% 20× SSC, and 0.2% SDS. The resulting solution was applied to the slides previously treated with the blocking solution and incubated at 60°C overnight. After hybridization, before scanning, slides were washed twice in 1× SSC with 0.05% SDS and 3 times in 0.1× SSC at 55°C. After spin-drying, slides were scanned at 10 μm using a GenePix 4000B microarray scanner (Axon Instruments, Foster City, Calif).
Expression values were obtained by GenePix Pro Software (Axon Instruments). On average, ≈70% of the spots presented no signal. Analyses of the remaining data were performed using BRB ArrayTools (version 3.0) developed by Dr Richard Simon and Amy Peng Lam (http://linus.nci.nih.gov/BRB-ArrayTools.html). We used the following filtering criteria: spots with diameter between 10 and 300 μm; intensity >100 after background subtraction; nonflagged spots; genes found in at least 90% of arrays. Lowess intensity–dependent normalization was used to adjust for differences in labeling intensities of the Cy3 and Cy5 dyes. The adjusting factor varied over intensity levels. We performed 14 technical replicates of the samples, which showed good correlation (mean±SD Pearson correlation coefficient 0.80±0.15 for replicates versus 0.24±0.1 for independent samples in the group) and were averaged.
We identified genes that were differentially expressed among the 2 classes, ie, clinical states, using a random-variance t test.10 Genes were considered statistically significant if their probability value was <0.01 (for biological data mining) and 0.001 or 0.005 (for class prediction). A global permutation test with 1000 random permutations was computed to evaluate the probability of finding a number of differentially expressed genes by chance. The proportion of the permutations that gave at least as many significant genes as with the actual data were the significance level of the global test. We accepted a set of genes when the probability of getting that given number of genes, if there are no real differences between the classes, was lower than 0.01. We also used the multivariate permutation test to provide 90% confidence that the false discovery rate was less than 10%. Hierarchical clustering of genes was performed using centered correlation and average linkage.
To classify samples based on expression profiles, we used class prediction tools. These included 6 cross-validated multivariate classification models (compound covariate predictor, diagonal linear discriminant analysis, 1- and 3-nearest neighbor predictor, nearest centroid predictor, support vector machine). The models incorporated genes that were differentially expressed at the 0.001 or 0.005 significance level, as assessed by the random-variance t test. To prevent a predictor from overfitting, we estimated the prediction error of each model using leave-one-out cross-validation (LOOCV) as described.11 For each LOOCV training set, the entire model building process was repeated, including the gene selection process. We also evaluated whether the cross-validated error-rate estimate for a model was significantly less than one would expect from random prediction. The class labels were randomly permuted, and the entire LOOCV process was repeated. The significance level is the proportion of the random permutations that gave a cross-validated error rate no greater than the cross-validated error rate obtained with the real data. One thousand random permutations were used. Among the 6 models, we considered the results of the ones that yielded a permutation probability value <0.05. Most samples were classified equally by all models. For those that had 50% of discordance between the methods, we applied the term “unclassified.” In addition to the LOOCV process, the predictor was also tested on independent test samples (supplemental Figure I).
For exploring biological pathways in the gene lists, we used a Database for Annotation, Visualization and Integrated Discovery (DAVID 2.1; http://david.abcc.ncifcrf.gov). We applied a Functional Annotation Tool to discover functional categories overrepresented in a gene list relative to the representation within the human genome. We used terms provided under Biological Process and Molecular Function in Gene Ontologies and in KEGG Pathways. For the redundant terms, only 1 main term was chosen.
The data regarding 52 heart biopsies that constituted the first training and test sets have been deposited in the Gene Expression Omnibus (GEO) of the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov/geo) under GEO Series accession no. GSE2596. Results of the remaining 24 samples collected and analyzed later have been deposited under series GSE4470.
To technically validate the specificity of microarray detection for expression of individual genes, we performed real-time RT-PCR for 2 selected12 reference genes and for 10 differentially expressed genes.
Analysis of Kidney and Lung Transplant Data From the Literature
The raw data on kidney and lung transplant samples were extracted from NCBI GEO Datasets, record numbers GDS365, GDS724, and GDS999. The original articles,13–15 as well as the GEO descriptions of the samples, were the source of all clinical details about the samples. After data extraction, the analyses were performed in BRB ArrayTools using the same procedures as described above but incorporating into the predictor set only genes discovered during the analysis of heart data. Sequences between different platforms were matched using gene symbols.
PCR for T cruzi
For T cruzi detection, we performed a PCR for short interspersed repetitive elements (SIRE) of T cruzi. SIRE are present in ≈1500 to 3000 copies per parasite genome and are included in the 3′ end of several mRNAs, contributing the polyadenylation site in 63% of parasite cDNAs.16 We amplified 20 ng of total RNA isolated from the biopsies and transformed into cDNA in 25 μL of PCR reaction with Eppendorf MasterMix (Eppendorf) and S2 and S3 primers published elsewhere.16 PCR was performed for 40 cycles with an annealing temperature of 56°C. Products of amplification were separated by electrophoresis on ethidium bromide–stained 2% agarose gels.
Rejection Versus Nonrejection
We first asked whether we could find an expression profile that discriminates rejection from nonrejection as accurately as does histology. We used 23 biopsies that fell into 3 groups. In 2 groups (8 samples that had been classified as “rejection” and 12 as “nonrejection”), the histology and the clinical outcomes were in agreement. The third group consisted of three samples for which the histology and clinical outcomes were not in concordance. Two samples had been histologically classified as nonrejection samples, but a rejection episode had occurred shortly thereafter (14 and 20 days). In a previous study, we had found by RT-PCR analysis that samples collected a short period (≈2 weeks) before a rejection episode display increased expression of some immune activation molecules, and we called these “prerejection” samples.17–19 The third “nonconcordant” sample had been classified as rejection but was retrospectively determined not to be a rejection episode, but, instead, a sign of leprosy, as the patient soon presented with cutaneous and cardiac manifestations of this disease.
There is currently no consensus about the best mathematical algorithm for class prediction, and the use of multiple algorithms increases the confidence and the validity of the results.11 We therefore used 6 different class prediction algorithms, as well as leave-one-out cross-validation (see Materials and Methods), to analyze the data sets from the 23 samples. With rare exceptions (see below), we found that the assignments were similar between them, imparting a certain robustness to the results.
Using the 23 samples, we found a set of 98 genes (Figure 1a; for the list of genes, see supplemental Table I) that correctly discriminated all 8 rejection and all 12 nonrejection samples. In addition, the 2 prerejection samples that had been labeled as nonrejection by histology were correctly classified as rejections by the microarrays, and the leprosy sample was classified as a nonrejection (supplemental Figure I). Therefore, microarray assignments correlated with the actual clinical situations and flagged cases missed by histology.
To be useful as a clinical test, microarray analysis should be applicable across the world, in different labs, with different sets of microarray slides and different sets of reference RNAs. To test the robustness and general applicability of our 98-gene rejection/nonrejection (R/NR) predictor set, we processed 2 new independent sets of biopsies: test set 1 consisted of 20 samples done 3 months later, on a different continent, and with a different batch of chips, whereas test set 2 consisted of 24 samples gathered prospectively after the original analysis. The results from these 2 sets were similar, and we have gathered them together here (see individual results in supplemental Table II).
Altogether, the 44 test samples consisted of 15 rejections, 26 nonrejections, and 3 samples that had been labeled as nonrejections by histology but for which the clinical data showed otherwise. Two were prerejection samples, and the third was a sample from a patient who presented with cardiac dysfunction (flutter). Although the histology showed no sign of rejection, the flutter was clinically taken to be a sign of a rejection episode and was successfully resolved with antirejection treatment.
The microarray analysis using the 98-gene R/NR predictor correctly classified 14 of 15 rejection samples and also correctly classified the prerejection and flutter samples as rejections. Of the remaining 26 histologically classified nonrejection samples, 20 were also labeled as nonrejections by the microarrays, three were not classifiable, as the six prediction algorithms gave only 50% concordance, and the other three were “false positives” labeled as rejections by microarrays.
Up to this point, taking all 67 samples together, the microarrays provided a diagnosis (ie, assigned a sample to either the rejection or nonrejection group) for 94% (63/67) of samples. Considering all available clinical information about the individual samples, 95% (60/63) of the samples were correctly classified (Table 1), and 6 cases (4 prerejections, 1 clinically defined rejection, and 1 case of leprosy) that had been missed by histology were correctly classified by microarrays.
To overcome any bias that might occur because of nonrandom representation of samples in the training set, we repeated the analysis described above twice more, using test sets 1 and 2 as training sets (and the remaining 2 groups as test sets), and obtaining 130- and 188-gene predictors, respectively (see lists in supplemental Tables III and IV). The classifications were remarkably stable, with 95% and 95.5% of correctly assigned samples. To see whether we could find a predictor set with a smaller number of genes, we looked for genes common to all 3 predictors (generated from the three training sets). We found that the 14 common genes (supplemental Table V) gave the same performance for classification (95.5%).
Kidney and Lung Transplants
To test whether our approach could be applied to the diagnosis of other types of transplants, we analyzed the data from 2 earlier studies on transplanted kidneys13,14 and lungs.15 Although these studies were done in different laboratories, using different procedures and microarray platforms, the analytic comparison was remarkably revealing. Using our first R/NR predictor set, we obtained almost perfect agreement with the histology from all 3 studies,13–15 as we were able to discriminate between acute rejection and nonrejection in 51 of 52 samples from the first kidney study,13 29 of 31 in the second,14 and 31 of 34 for the lung study15 (Figure 2a). Thus the overall correctness of microarray diagnosis for kidney and lung acute rejection was ≈94%. The second and third R/NR predictor sets, and the small 14-gene predictor, gave very similar results (supplemental Table VI). Thus, the predictor sets discovered in hearts are also diagnostically useful for kidneys and lungs.
Rejection Versus Infection
The finding that leprosy was misclassified as rejection by histology and not by microarray led us to test whether the R/NR profile can discriminate between rejection and infections. We chose Chagas reactivation as a model infection because heart biopsies from patients with reactivation of Chagas disease are often histologically misdiagnosed as rejections (unless actual parasites are seen).4 We used R/NR gene predictor to analyze 8 samples from patients with documented Chagas reactivation, ie, samples that were visibly positive for T cruzi organisms on histology. Seven samples were classified as rejections. Thus the expression profile of rejection and trypanosome infections is similar enough that the R/NR predictor set could not distinguish them (Figure 1a).
Nevertheless, intrigued by the leprosy result, we went back to the basics to determine whether there might exist a different gene profile able to discriminate trypanosome infection from rejection. We used 11 rejection and 8 Chagas disease samples as a training group and found 87 rejection/Chagas (R/Ch) predictor genes (Figure 1b and supplemental Table VII) that correctly classified all of the 8 Chagas samples and 10 of 11 rejection samples (supplemental Figure I).
Interestingly, for 1 Chagas sample, the microarray seemed to give more accurate information than histology. This patient had no evidence of parasites on histology but presented subcutaneous reactivation of Chagas and responded well to anti-Chagas treatment without any anti-rejection therapy (Table 2, patient D). Retrospectively, therefore, the infiltrate in the biopsy was clinically considered to be a sign of infection reactivation. Thus, although microarrays did not seem to discriminate perfectly between rejection and Chagas relapse, they were more sensitive to Chagas than histology or PCR (Table 1).
Having found that microarrays can be predictive for rejection, we asked whether they might also be able to detect infections before the parasites are seen by histology. We used the 87-gene R/Ch predictor set to analyze 12 nonrejection biopsies from patients who had been transplanted because of Chagas related heart failure but who showed no clinical or histological signs of disease relapse or of rejection at the time that the biopsies had been taken. Seven of the biopsies came from patients who relapsed 1 to 2 months later (pre-Chagas group), and the microarrays detected 5 of these as having gene expression profiles characteristic of Chagas (Table 1). The microarrays did not flag biopsies from patients that had remained clear of disease nor 2 samples with toxoplasma and leprosy infections. Thus, there were no false positives and 5 of 7 true positives that had been missed by histology. From the 2 ‘false-negative’ biopsies, one was taken 1 week after transplantation (Table 2, sample A7), and 65 days before Chagas relapse. The second was taken three weeks after transplantation and 2 months before the relapse (Table 2, sample F21). From this patient, we also had a second sequential biopsy, taken 34 days later. Although still quiescent by histology, this biopsy was Chagas positive by microarray. Thus, it seems that the sensitivity of the microarray technique is such that it can pick up ongoing relapse of Chagas disease up to 60 days earlier than histology (or PCR; Tables 1 and 2⇑).
In kidney transplants, as in hearts, it can sometimes be difficult to discriminate between infection and rejection.20 Encouraged by the results of the R/NR predictor set, we asked whether our 87-gene R/Ch predictor set might also be usefully applied to kidneys. Of the 2 studies of kidneys, one13 included data from urinary tract infections, and, of the 87 genes in our infection versus rejection predictor set, 61 were also represented on the chips used by this kidney study.
Using these 61 genes, we found only 11 that were similarly regulated in cardiac Chagas infections and in urinary tract infections in kidney transplant patients (supplementary Table VII), showing that kidneys respond somewhat differently to bacterial infections than hearts do to trypanosome parasites. As a predictor set, however, these 11 genes turned out to be quite useful, as they correctly identified 17 of 19 analyzable kidney samples (8 infections versus 11 uninfected rejection samples; Figure 2b). Thus, although the 2 types of infectious agents, and their target tissues, are quite different, the microarrays seem to reveal the existence of some universal features of infections in transplants.
In searching for predictive/diagnostic gene sets, we had set the statistical criteria quite strictly so as to find a relatively small set of genes whose patterns were definitive. To have a better look at the biology of the grafted tissues, we did pairwise comparisons between rejection, nonrejection, and infection groups, casting a wider net of genes to allow a more confident discovery of biological themes (details in the online data supplement, page 4). We found a set of 524 genes whose expression differed between rejection and nonrejection (R/NR), 820 that differed between Chagas and nonrejection (Ch/NR), and 318 genes that differed between rejection and Chagas samples (R/Ch) (supplemental Figure III).
Two hundred forty-one genes (supplemental Table VIII) were common to both the R/NR and Ch/NR sets, ie, “pathology”-specific genes whose expression in either pathological state similarly differed from the baseline of nonrejection. To reveal main biological pathways in a gene list, we performed gene-enrichment analysis, which compares the frequency of genes of a given category found in any particular gene list with the frequency of genes of the same category in the genome as a whole. Not surprisingly, the most significantly enriched category among upregulated pathology-specific genes (more than one-third of upregulated genes) was related to the immune response and included groups of genes involved in innate and adaptive immunity, antigen presentation, cytokine binding, and chemotaxis (Figure 3). Immunoglobulin genes were among those most up regulated (5- to 10-fold change; supplemental Table VIII), suggesting that B cells may have a local effect at the site of inflammation.
Among downregulated pathology genes, a substantial portion was concerned with energy-yielding metabolism (Figure 3). The most predominant among these were genes representing mitochondrial oxidative phosphorylation, hydrogen electron transport, and ATP metabolism. In addition, genes involved in apoptosis were enriched in this list (Figure 3).
Kidney and lung transplants showed similar patterns. The majority of upregulated genes in acute rejection were related to immune/inflammatory processes, and downregulated genes were related to energy-yielding metabolism (Figure 2a).
We next had a look at genes specific for each particular pathological state (infection and rejection), rather than common to the two. In the 173-gene “infection” profile (supplemental Table X), we observed a pronounced enrichment of genes encoding the proteasome–ubiquitin complex (Figure 3). The proteasome genes were upregulated, whereas the ubiquitin-related genes were downregulated. This may be the mechanism by which T cruzi regulates protein catabolism and antigen presentation that are known to be altered in this infection.21,22 The infection gene profile also contained at least 2 genes coding for molecules already known to be involved in T cruzi infection (α-2-macroglobulin and spermidine/spermine N1-acetyltransferase).23–25 These 2 genes represent 2 types of regulation involved in trypanosome infections: genes regulated by the host to mediate immune and tissue protective responses and host genes modulated by the parasite for its own survival objectives. The remaining 171 Chagas-specific genes are therefore likely to represent a wealth of previously unknown genes involved in T cruzi infection.
Finally, there were 44 rejection-specific genes (supplemental Table IX). Because this list was relatively small, the gene enrichment analysis did not reveal highly significant gene categories (Figure 3). In this case, we did a manual analysis, searching various databases (PubMed, GenBank, GeneCards),and we found that there were some upregulated immune/inflammatory genes and some downregulated genes involved in muscle function (supplemental Table XI).
In the present work, we revealed the existence of molecular expression profiles that are characteristic of acute rejection and infection in transplanted hearts. Overall, as a diagnostic tool, gene profiling by microarray analysis seemed to be superior to histology (the current gold standard), as it picked up both rejection crises and infections that were missed, or misdiagnosed, by histology, and it was often able to detect both these states in their early stages, weeks or months earlier than histology. Of a total of 76 samples, the microarrays classified 22 rejections, 8 infections, 4 incipient rejections, and 5 incipient infections. Of these, 5 rejections and prerejections and 6 infections and preinfections had been missed by the histology. These findings have enormous clinical potential, as they allow for timely and appropriate treatment.
In generating diagnostic gene profiles (known as gene predictors), one uses a portion of available samples as a “training” group and remaining samples as “test” groups to validate the gene predictors. Because there is some concern that the choice of training samples can bias the resultant predictor gene set, and that different gene sets will generate different classifications,26 we took 3 different groups of samples and used each group as a training set to generate rejection/nonrejection gene predictors, validating the predictor with the other 2 groups. Although the 3 training sets gave only partially overlapping gene lists, the prediction results were extremely consistent.
Having found that all 3 gene predictors gave excellent discrimination between rejection and nonrejection of heart transplants, we also tested them on the published data from kidney and lung transplants.13–15 This was not a perfect comparison to the heart data for several reasons. First, the microarrays used in the kidney and lung studies covered a somewhat different range of genes from the microarray we used for the hearts. For example, among the 98-gene R/NR predictor, we found that only 64 were represented on the chips from the lung study15 and 54 and 45 on the chips from the first13 and second14 kidney studies, respectively. Second, detailed descriptions of clinical information (which were extremely helpful for heart graft analysis) were not available for the kidney samples. Third, whereas the heart and kidney studies were done using biopsies of the target tissues, the lung microarray analyses were accomplished with bronchoalveolar lavage fluid. Fourth, the studies were done in different laboratories, using different procedures. In spite of these technical differences, our analysis showed that the different gene-predictor sets discovered on heart biopsies were also able to correctly diagnose rejections and infections in 90% to 95% of cases in kidney and lung samples. Therefore, the analysis of gene profiles allows for diagnosis of rejection versus nonrejection versus infection in several different tissues. Although it may seem surprising that a gene profile established using heart biopsies should be useful for diagnostics in other tissues, it is likely to reflect similarities of a given pathological process, rather than similarities of the physiology of a particular organ.
One might argue that taking biopsies (either for histology or for microarray analysis) is an invasive procedure and wish for a less invasive technique. To this end, there have been many studies on blood (reviewed previously27,28), the most recent of which used microarray and RT-PCR to generate an “AlloMap” RT-PCR test in an attempt to discriminate rejection from nonrejection in cardiac transplant patients.29 Unfortunately, this test gave a reasonable correlation with the clinical situation only from samples taken later than 6 months after transplant. During the period when patients are at greatest risk of acute rejection (the first 6 months),30 the RT-PCR tests on blood samples were not informative. When we reanalyzed their microarray dataset, either using our predictor sets or attempting to discover a new set based on their 285 blood samples, we found that the microarray data from blood did not give significant predictions (data not shown).
In contrast, our analyses of biopsies, where ≈90% of the samples were collected during the first 6 months after transplantation, gave correct classifications in ≈95% of cases, discriminating both ongoing and incipient rejections and infections. At the present time, therefore, in spite of its invasiveness, analyses based on biopsies may be preferable to those obtained from blood samples, at least for the first 6 months after transplantation.
Exploring the biological pathways, we found a clear common inflammatory signature of rejection in the three transplanted organs that also extended to hearts undergoing Chagas infection (Figures 1 and 2⇑). There were 2 sets of genes prominently enriched in this signature: a set of upregulated immune genes and another of downregulated energy related genes. The high levels of immune/inflammatory genes most likely result from both an increase in gene expression per cell and an increase in the number of infiltrating immune cells expressing these genes. The finding that immunoglobulin genes were the most upregulated was unexpected, because the location of antibody production has largely been attributed to the secondary lymphoid tissues. However, activation of immunoglobulin genes in the target tissue has also recently been seen in human renal and lung transplantation,13,15 as well as in experimental chronic Chagas cardiomyopathy.31
Although it may not be surprising that there exist similarities in the (upregulated) immune responses found in rejections and infections, the finding that the depression of energy metabolism is also a common feature of acute inflammation, regardless of the causative agent (rejection/infection) or tissue/organ (heart/kidney/lung), is quite astonishing. Interestingly, this pattern has also been seen in experimental cardiac infection with T cruzi,32 diabetes,33 and schizophrenia.34 It suggests that energy loss may be the first step in the destructive chain of events occurring in inflamed tissue during several different types of pathology, even before apparent organ dysfunction and structural changes. Such loss may occur because inflammatory intermediates (such as free radicals) can directly cause mitochondrial injury and respiratory chain dysfunction.35,36 This may be one of the reasons that therapeutic scavenging of free radicals enhances transplant acceptance.37,38 Therefore, our data indicate 2 major biologic pathways to which therapy could be directed: (1) the immune response, which is already widely targeted by diverse immunosuppression regimens and tolerance induction protocols39; and (2) energy metabolism, which has been given much less attention in transplants. Targeting the metabolic pathways involved in energy usage could potentially lead to substantial improvement of transplant outcomes.
Although the study has some limitations, such as the fact that we did not look for gene sets able to discriminate different grades of rejection and that we studied only 1 type of infection, the results are nevertheless statistically highly significant. Future fine tuning with a large number of samples from a multicentric study will allow for insights into these questions.
In summary, we found that microarrays can be used to create predictor sets of genes that can more accurately diagnose an incipient rejection and/or infection than the current gold standard (histology) used today for heart transplant diagnosis. Although we based our predictor sets on heart transplants and Chagas infections, the resulting predictor sets were also able to diagnose lung and kidney rejections and to discriminate kidney rejections from their accompanying bacterial urinary tract infections, suggesting that there are universal biological profiles that characterize rejections and infections in different types of transplants.
This study is dedicated to Katy Tomes, a kidney recipient who died of an infection that was diagnosed too late. We thank Allan Kirk and Louis Staudt for cogent advice on the data analysis and manuscript discussion. We also thank Mike Wilson for help with set up of microarrays, Franco Marincola and Ena Wang for teaching us RNA amplification, Marcia M. de Souza for discussion of histological results, Amador Goncalves-Primo and Carolina Fagundes de Sá Primo for excellent technical assistance in experiments, and Drs M. Hertz and J. Lande for providing additional information about lung transplants.
Sources of Funding
This research was supported by the Research Program of the Immunogenetics Division, Federal University of São Paulo and the Intramural Research Program of the NIH, National Institute of Allergy and Infectious Diseases. G.F.S. is a recipient of the Fellowship of the Fundação de Amparo à Pesquisa de São Paulo (FAPESP).
Centers for Disease Control and Prevention. Chagas disease after organ transplantation–United States, 2001. JAMA. 2002; 14: 1795–1796.
Stewart S, Winters GL, Fishbein MC, Tazelaar HD, Kobashigawa J, Abrams J, Andersen CB, Angelini A, Berry GJ, Burke MM, Demetris AJ, Hammond E, Itescu S, Marboe CC, McManus B, Reed EF, Reinsmoen NL, Rodriguez ER, Rose AG, Rose M, Suciu-Focia N, Zeevi A, Billingham ME. Revision of the 1990 working formulation for the standardization of nomenclature in the diagnosis of heart rejection. J Heart Lung Transplant. 2005; 24: 1710–1720.
Bacal F, Silva CP, Bocchi EA, Pires PV, Moreira LF, Issa VS, Moreira SA, das Dores Cruz F, Strabelli T, Stolf NA, Ramires JA. Mychophenolate mofetil increased chagas disease reactivation in heart transplanted patients: comparison between two different protocols. Am J Transplant. 2005; 8: 2017–2021.
Wright GW, Simon R. A random variance model for detection of differential gene expression in small microarray experiments. Bioinformatics. 2003; 19: 2448–2455.
Simon R, Radmacher MD, Dobbin K, McShane LM. Pitfalls in the analysis of DNA microarray data: class prediction methods. J Natl Cancer Inst. 2003; 95: 14–28.
Flechner SM, Kurian SM, Head SR, Sharp SM, Whisenant TC, Zhang J, Chismar JD, Horvath S, Mondala T, Gilmartin T, Cook DJ, Kay SA, Walker JR, Salomon DR. Kidney transplant rejection and tissue injury by gene profiling of biopsies and peripheral blood lymphocytes. Am J Transplant. 2004; 4: 1475–1489.
Vazquez M, Ben-Dov C, Lorenzi H, Moore T, Schijman A, Levin MJ. The short interspersed repetitive element of Trypanosoma cruzi, SIRE, is part of VIPER, an unusual retroelement related to long terminal repeat retrotransposons. Proc Natl Acad Sci U S A. 2000; 97: 2128–2133.
Shulzhenko N, Morgun A, Franco M, Souza MM, Almeida DR, Diniz RV, Carvalho AC, Pacheco-Silva A, Gerbase-Delima M. Expression of CD40 ligand, interferon-gamma and Fas ligand genes in endomyocardial biopsies of human cardiac allografts: correlation with acute rejection. Braz J Med Biol Res. 2001; 34: 779–784.
Shulzhenko N, Morgun A, Zheng XX, Diniz RV, Almeida DR, Ma N, Strom TB, Gerbase-DeLima M. Intragraft activation of genes encoding cytotoxic T lymphocyte effector molecules precedes the histological evidence of rejection in human cardiac transplantation. Transplantation. 2001; 72: 1705–1708.
Racusen LC, Solez K, Colvin RB, Bonsib SM, Castro MC, Cavallo T, Croker BP, Demetris AJ, Drachenberg CB, Fogo AB, Furness P, Gaber LW, Gibson IW, Glotz D, Goldberg JC, Grande J, Halloran PF, Hansen HE, Hartley B, Hayry PJ, Hill CM, Hoffman EO, Hunsicker LG, Lindblad AS, Marcussen N, Mihatsch MJ, Nadasdy T, Nickerson P, Olsen TS, Papadimitriou JC, Randhawa PS, Rayner DC, Roberts I, Rose S, Rush D, Salinas-Madrigal L, Salomon DR, Sund S, Taskinen E, Trpkov K, Yamaguchi Y. The Banff 97 working classification of renal allograft pathology. Kidney Int. 1999; 55: 713–723.
Morrot A, Strickland DK. Higuchi M, del Reis M, Pedrosa R, Scharfstein J. Human T cell responses against the major cysteine proteinase (cruzipain) of Trypanosoma cruzi: role of the multifunctional alpha 2-macroglobulin receptor in antigen presentation by monocytes. Int Immunol. 1997; 9: 825–834.
Piacenza L, Peluffo G, Radi R. L-arginine-dependent suppression of apoptosis in Trypanosoma cruzi: contribution of the nitric oxide and polyamine pathways. Proc Natl Acad Sci U S A. 2001; 98: 7301–7306.
Deng MC, Eisen HJ, Mehra MR, Billingham M, Marboe CC, Berry G, Kobashigawa J, Johnson FL, Starling RC, Murali S, Pauly DF, Baron H, Wohlgemuth JG, Woodward RN, Klingler TM, Walther D, Lal PG, Rosenberg S, Hunt S; CARGO Investigators. Noninvasive discrimination of rejection in cardiac allograft recipients using gene expression profiling. Am J Transplant. 2006; 6: 150–160.
Mootha VK, Lindgren CM, Eriksson KF, Subramanian A, Sihag S, Lehar J, Puigserver P, Carlsson E, Ridderstrale M, Laurila E, Houstis N, Daly MJ, Patterson N, Mesirov J, Golub TR, Tamayo P, Spiegelman B, Lander ES, Hirschhorn JN, Altshuler D, Groop LC. PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat Genet. 2003; 34: 267–273.
Prabakaran S, Swatton JE, Ryan MM, Huffaker SJ, Huang JT, Griffin JL, Wayland M, Freeman T, Dudbridge F, Lilley KS, Karp NA, Hester S, Tkachev D, Mimmack ML, Yolken RH, Webster MJ, Torrey EF, Bahn S. Mitochondrial dysfunction in schizophrenia: evidence for compromised brain metabolism and oxidative stress. Mol Psychiatry. 2004; 9: 684–697.
Tatsumi T, Matoba S, Kawahara A, Keira N, Shiraishi J, Akashi K, Kobara M, Tanaka T, Katamura M, Nakagawa C, Ohta B, Shirayama T, Takeda K, Asayama J, Fliss H, Nakagawa M. Cytokine-induced nitric oxide production inhibits mitochondrial energy production and impairs contractile function in rat cardiac myocytes. J Am Coll Cardiol. 2000; 35: 1338–1346.
Tanaka M, Mokhtari GK, Terry RD, Balsam LB, Lee KH, Kofidis T, Tsao PS, Robbins RC Overexpression of human copper/zinc superoxide dismutase (SOD1) suppresses ischemia-reperfusion injury and subsequent development of graft coronary artery disease in murine cardiac grafts. Circulation. 2004; 110 (suppl II): II-200–II-206.