Human Biomarker Discovery and Predictive Models for Disease Progression for Idiopathic Pneumonia Syndrome Following Allogeneic Stem Cell Transplantation*

Allogeneic hematopoietic stem cell transplantation (SCT) is the only curative therapy for many malignant and nonmalignant conditions. Idiopathic pneumonia syndrome (IPS) is a frequently fatal complication that limits successful outcomes. Preclinical models suggest that IPS represents an immune mediated attack on the lung involving elements of both the adaptive and the innate immune system. However, the etiology of IPS in humans is less well understood. To explore the disease pathway and uncover potential biomarkers of disease, we performed two separate label-free, proteomics experiments defining the plasma protein profiles of allogeneic SCT patients with IPS. Samples obtained from SCT recipients without complications served as controls. The initial discovery study, intended to explore the disease pathway in humans, identified a set of 81 IPS-associated proteins. These data revealed similarities between the known IPS pathways in mice and the condition in humans, in particular in the acute phase response. In addition, pattern recognition pathways were judged to be significant as a function of development of IPS, and from this pathway we chose the lipopolysaccaharide-binding protein (LBP) protein as a candidate molecular diagnostic for IPS, and verified its increase as a function of disease using an ELISA assay. In a separately designed study, we identified protein-based classifiers that could predict, at day 0 of SCT, patients who: 1) progress to IPS and 2) respond to cytokine neutralization therapy. Using cross-validation strategies, we built highly predictive classifier models of both disease progression and therapeutic response. In sum, data generated in this report confirm previous clinical and experimental findings, provide new insights into the pathophysiology of IPS, identify potential molecular classifiers of the condition, and uncover a set of markers potentially of interest for patient stratification as a basis for individualized therapy.

Allogeneic hematopoietic stem cell transplantation (SCT) 1 is the only curative therapy for many individuals with malignant and non malignant conditions. Unfortunately, a number of complications limit successful outcomes. Diffuse lung injury occurs in 25-55% of SCT recipients and can account for ϳ50% of all SCT related mortality (1,2). Lung injury following SCT can be infectious or noninfectious in nature. Idiopathic pneumonia syndrome (IPS) is defined as widespread alveolar injury following SCT without an active lower respiratory tract infection (1,2). The incidence of IPS ranges from 5-15% depending upon donor type and conditioning regimen intensity. The median time to onset of IPS is ϳ18 days, and the mortality rate is Ͼ70% despite aggressive treatment and supportive care (2,3).
Several laboratories have used rodent models to elucidate the molecular mechanisms at play in the development of IPS (2). These models recapitulate many aspects of the clinical histopathology and physiological observations that are associated with human disease (reviewed in (2)). For example, histochemical staining of mouse lung tissue reveals dense mononuclear cell infiltrates consisting of macrophages and lymphocytes around pulmonary vessels and bronchioles. Increases in these mononuclear infiltrates are accompanied by high levels of tumor necrosis factor alpha (TNF␣) in both lung tissue and broncho-alveolar lavage (BAL) fluid (4). Neutralization of TNF␣ via a number of strategies reduces the progression and severity of experimental IPS demonstrating a casual role of this protein in the pathophysiology of disease (5)(6)(7)(8). These experimental insights have been translated into clinical studies that are currently evaluating the use of etanercept (Enbrel ® ; Amgen, Thousand Oaks, CA) a soluble TNF binding molecule for the treatment of IPS in both pediatric and adult SCT recipients (9).
Preclinical studies have helped refocus IPS away from an idiopathic clinical syndrome and toward a SCT complication wherein the lung is a target of both innate and adaptive immune dysregulation (2). To date, the majority of clinical research efforts involving IPS have focused primarily on describing disease characteristics with a few reports measuring immuno-modulatory proteins in BAL fluid and/or plasma via enzyme linked immunosorbent assays (ELISA) (9,10). Although these studies have provided important insights into understanding the possible contribution of soluble cytokines to the inflammation engendered during IPS, the approaches taken thus far have not been comprehensive (they target only a few known inflammatory proteins) and are limited with respect to expanding our understanding of the disease pathway. Nevertheless these focused studies provide a set of expected results that should be replicated in unbiased, comprehensive studies.
Discovery-based, quantitative, proteomics provides a powerful, unbiased approach for identification and quantification of hundreds of proteins, and multiple reliable proteomic techniques are currently available (11)(12)(13)(14). Label-free protein expression analysis is one such peptide based ("bottom up") proteomic technique, which capitalizes on the highly reproducible chromatography and parts-per-million mass accuracy available in current liquid chromatography/mass spectrometry (LC-MS) systems. Peptide species in these approaches are quantified by ion intensity, while individual peptides are grouped across samples based on precise mass and retention time measurements (15,16). Here we describe two labelfree protein expression analysis studies where key samples were derived from a clinical trial for the treatment of IPS. The goals of these two studies were different and thus the study designs and verification strategies varied, although both relied on label-free mass spectrometry as the main experimental platform. In the first experimental study we identified disease progression-associated proteins by analyzing peptide intensity changes over two time points (day of SCT and day of diagnosis) for cases who developed IPS compared with control patients studied at two comparable time points (day of SCT and 14 -21 days later). In this case, the specific goal was to explore and potentially expand the disease pathway using molecular network analysis. The peptides/proteins detected as significantly changing in this analysis were verified by a number of methods including: pathways analysis, comparison to literature results including mouse models and human data, and verification of a specific protein of interest (LBP). The second label-free expression experiment was performed in order to build a predictive model of disease and treatment. Thus, we focused on predictors of both disease and treatment response using additional samples collected only at Day 0 of SCT. In this case a cross-validation (leave-one-out) strategy was used to understand classification power of the predictive biomarkers. The results of the first study both confirm previous clinical and experimental findings and provide new insights into the pathophysiology of IPS. The results of the second study suggest a predictive model that can identify, at the time of SCT, the patients who will progress to IPS and subclassify those IPS patients into responders and nonresponders to cytokine neutralization therapy. These markers are suitable for further study in a wider clinical setting.

EXPERIMENTAL PROCEDURES
Patients and Patient Controls-Samples analyzed in the label-free protein expression experiments were collected from patients with IPS who were enrolled on a clinical study conducted at the University of Michigan and the Dana Farber Cancer Institute. Information regarding patient exclusion criteria, the procedures used to diagnosis IPS, and therapeutic intervention with corticosteroids and etanercept have been described in detail (9). In brief, study patients were Ն1 year in age, recipients of allogeneic SCT and within 100 days of stem cell infusion. All patients had evidence of IPS as defined by NIH working group criteria (1,2). There were no restrictions based upon underlying disease, stem cell source, transplant conditioning regimen, or donor/ recipient HLA match. SCT control subjects received myelo-ablative conditioning and subsequently exhibited no signs of acute graft versus host disease, hepatic veno-occlusive disease, significant infection, or pulmonary dysfunction through day 100 following SCT. All patients (or their surrogates) and controls gave written, informed consent in accordance with the Declaration of Helsinki and the trial was approved by the respective Institutional Review Boards of the University of Michigan and the Dana-Farber Cancer Center.

Label-free Expression
Plasma Sample Collection and Preparation-Blood was obtained for label-free expression immediately after stem cell transplantation (Day 0) and at either the time of diagnosis (for patients with IPS) or between 14 and 21 days after transplant for patient controls. Specimens were collected in heparinzed tubes and plasma was separated and stored at Ϫ80°C until analysis. Aliquots for individual plasma samples for both studies were thawed and subsequently depleted of the seven most abundant proteins using a 4.6 ϫ 100 mm multiple affinity removal system (MARS Hu7, Agilent Technologies, Santa Clara, CA) according to manufacturer's instructions. Each depleted sample was concentrated (5,000 molecular weight cutoff, Millipore, Billerica, MA) and buffer exchanged with 50 mM Tris pH 8.8 to a final volume of ϳ100 l. Total protein concentrations were determined by 2D Quant Kit as described by the manufacturer (GE Healthcare Piscataway, NJ) and 10 g of flow through and bound fractions for each sample were loaded onto a one dimensional SDS-PAGE gel (4 -20% Tris-HCL) as a quality control measure to confirm success of the depletion step. Subsequent to digestion, each sample was adjusted to 60 g in 50 l volume. Twenty microliters of 0.2% Rapigest (Waters, Milford, MA) was added followed by dithiotheritol to a final concentration of 5 mM. The samples were reduced at 80°C for 15 min and cooled to room temperature prior to alkylation with iodacetamide at a final concentration of 10 mM for 30 min. Proteolytic digestion was performed with bovine trypsin (Promega, Madison, WI) with a final enzyme to protein ratio of 1:10 (w/w) for 18 h at 37°C. Elisa methods for LBP were performed by the Immunology Core Laboratory at the University of Michigan as previously published (9) and per manufacturer's recommendations (Cell Sciences, Canton, MA).
Discovery Label-free Expression Data Set-The categorical factors under study consist of a Disease Factor (DF) crossed with a Time factor (TF). The two levels of the DF were the two groups of patients both of whom received allogeneic SCT and were either subsequently diagnosed with IPS (Disease -D) or experienced no complications (Controls -C). The two levels of the TF included the two time-points studied: day of SCT (Day-0) and either the day of IPS diagnosis (for IPS patients) or day 14 to 21 (Day-14) for controls. The later corresponds to the median time of onset of IPS. With this design, the goal of the first analysis was tightly focused on identifying proteins changing over both time and disease development (interaction). Eleven patients were included in this initial study. Specimens were collected at each time point, therefore a total of 22 observations (within 4 experimental groups) were analyzed with n ϭ 5 to 6 biological replicates per group (see statistical section for power analysis as well as supplemental Fig. S1). Samples were randomized across the analysis (without blocking). There is no common reference sample. No technical replicates or sample pooling was performed.
Predictive Model Label-free Expression Data Set-The experimental design for establishing the disease predictive model included SCT patients who subsequently developed IPS (Disease-D) or those who experienced no complications (Controls-C). The experimental design for establishing the responsiveness predictive model included SCT patients who did not develop IPS (Controls-C), and those with IPS who ultimately responded to etanercept therapy (Responders-R) or not respond (Nonresponders -NR). Samples utilized in this model were those only collected at day of SCT with the goal established to define predictive and prognostic biomarker variables within the context of the aforementioned clinical trial. Technical replicates for 23 individual samples (14 from IPS patients; nine from controls) were utilized giving a total of 46 observations that were randomized across the analysis.
Liquid Chromatography and Mass Spectrometry-Six hundred nanograms of each sample were analyzed by LC/MS/MS using a Dionex Ultimate 3000 capillary LC system (Dionex Sunnyvale, CA) and a LTQ-Orbitrap XL (Thermo, Waltham, MA). The order of sample injections was randomized over all samples for each study. Separation and detection of peptides was performed as previously published (17). The instruments were mass calibrated immediately before the analysis using the instrument protocol. Raw LC/MS/MS data were processed via Proteomarker software version 1 (Infochromics, Toronto, Canada) for the discovery set and Rosetta Elucidator version 3.3.01 (Rosetta Biosoftware, Seattle, WA) for the predictive model set. The MS/MS peak lists were subsequently searched by Mascot version 2.2.0 (Matrix Science London, UK). The database used was the human International Protein Index (IPI) (68020 sequences). Search settings were as follows: trypsin enzyme specificity; mass accuracy window for precursor ion, 10 ppm; mass accuracy window for fragment ions, 0.8 Daltons; variable modification including carbamidom-ethylation of cysteines, 1 missed cleavage and oxidation of methionine. The criteria for peptide identification were a mass accuracy of Յ10 ppm and an expectation value of p Յ 0.05 and an estimated False Discovery Rate (FDR) of 1% (18,19). Proteins that had Ͼ2 peptides matching the above criteria were considered confirmed assignments whereas proteins identified with one peptide regardless of the Mascot score were highlighted as tentative assignments. Automated differential quantification of peptides in a set of samples was accomplished with Proteomarker or Rosetta Elucidator as previously described (17,20,21).
Data Quality Control-The discovery label-free experiment was processed for data quality control and dimension reduction in order to reduce missing value rates and error rates as previously described (17). Briefly, the data set was a matrix of n ϭ 22 observations/ individuals (rows) times p ϭ 14,253 aligned m/z intensities (columns). From empirical evidence within the data, we assumed a bimodal distribution of consensus peptide sequence Mascot scores. A diffset for analysis was rejected if the consensus peptide sequence score was below the 85.7th percentile (Յ 25.35), which represents the lowest mode of distribution of Mascot scores (see supplemental Fig. S2). Out of these, 2024 diffsets were identified. An additional prefiltering step was carried out to select those diffsets for which the observed count of missing values per peptide satisfied two criteria simultaneously: (1) it had to be less than the total sample size n minus half of the minimum sample size per experimental group, (2) it had to maximize the difference between the total number of remaining peptides after selection and the total number of missing values. Here with n ϭ 22 and an initial number of peptides p ϭ 14253, taking the maximum mean value from the above criteria, the maximum number of missing events allowed per peptide was bounded to 19, and the final number of peptides retained was p ϭ 1093. Despite the previous QC filtering steps, missing values still remained in the preselected data. These missing values represent a mixture of truly absent peptides whose expressions are to be taken as background "noise," and truly low abundance peptides whose levels of expression are close to the limit of detection. To account for this nonignorable left-censoring issue (i.e. the nonrandom nature of the missingness mechanism at play) and its extent, we used a probability model adapted from Wang et al. (22), which describes "artifactual missing events" in this type of data. This model makes inferences on the missing values of one sample, based on the information from other "similar" samples (technical replicates or nearest neighbors). It substitutes a missing measurement of intensity with its expected value of the true intensity given that it is unobservable. The imputation model enabled quantification of low-abundance peptides and reduced the proportion of missing data to less than 10%. The remaining missing values after imputation represent truly absent peptides in the samples and were imputed by taking an estimate of the background noise. This amount of missing data compared very favorably to previously published label-free data sets (17).
The predictive model label-free experiment was processed using Rosetta Elucidator as previously described (20,21,23). The predictive data set was a matrix of n ϭ 23 rows (observations/individuals) times p ϭ 12,588 columns (aligned m/z intensities). Peptide and protein assignments were made using peptide and protein teller algorithms (24,25) in Rosetta Elucidator with a false discovery rate of 1% (Rosetta Biosoftware, Seattle, WA). Quantification and imputation were performed only on those peptides that passed the above identification criteria, which yielded 1722 peptides.

Statistical Analysis
Data Transformation-To help remove sources of systematic variation (bias and variance) in the measured intensities of the above data sets because of experimental artifacts and to ensure that the usual assumptions for statistical inferences are met (normality), we first applied a log-transformation on the variables (peptides). In addition, because the homoscedasticity assumption in multi-group designs is also required e.g. in ANOVA models, we also applied our recently developed "joint adaptive mean-variance regularization" procedure as described in (26,27), now available as an R package called "MVR" (28). Briefly, the joint adaptive regularization procedure simultaneously overcomes the lack of degrees of freedom and the variancemean dependence issue in this type of data set where the number of variables hugely dominates the number of samples (26,28). Although, the procedure is designed to stabilize the variance across variables (peptides), we observed that it also translates into good variance stabilization effect across sample groups in a multi-group design, as is the case in this study.
Power Analysis and Sample Size Calculation-We determined a sample size per group that would guarantee to detect a minimum effect of interest, at a certain power, while simultaneously controlling the positive False Discovery Rate (pFDR) (29,30), as described in Liu et al. (31). After variance stabilization and normalization of the data (26), the usual distributional assumptions of test statistics become satisfied. Under the above assumption, all peptides are assumed to have approximately a common standard deviation across samples/ groups and a common effect size ⌬/ (31). In the case of a balanced design, we calculated the group sample size n required to detect e.g. at least a 2.3-fold change (effect size ⌬/ ϭ log 2 (2.3)/ ϭ 2) in the interaction effect between the Disease and Time factors as a function of power 1 Ϫ ␤, the parameter 0 , interpreted as the proportion of nondifferentially expressed peptides, the common standard deviation , and for a fixed level ␣ of pFDR. Based on preliminary data in our hands, estimates of the parameters 0 and of the common standard deviation of the peptides across experiments were 0 Ϸ 0.47-0.77 and ϭ 0.84 -1.00, respectively, which is consistent with estimates found in other platforms in high dimensional data (31). Under above assumptions and estimations, the sample size per group, required to detect at least a 2.3-fold change in the interaction effect, providing at least 80% power ͑1 Ϫ ␤ ϭ 0.8͒, and controlling for at most 5% pFDR ͑ pFDR Յ ␣ ϭ 0.05͒ was n ϭ 5 (supplemental Fig. S1). Therefore, a minimum number of individual subjects to enroll in our study had to be at least 2n ϭ 10 (for a total of 4n ϭ 20 observations).
Discovery Label-free Protein Expression-Unsupervised Analyses-Potential groups and outliers among the samples were checked by a principal component analysis (PCA) (32). Clustering analysis was also performed using complete linkage hierarchical clustering and gap statistics to estimate the real number of clusters in the data (33).

Discovery Label-free Differential Expression Profiles of Peptide-Linear Mixed Model and Empirical Bayes Estimators for Statistical
Inference-The thousands of retained peptides were analyzed by well-established techniques that have been developed for analyzing high-dimensional data. If we let y jkl be the intensity signal on the original scale of i-th individual and the j-th peptide, and z jkl ϭ t͑Y jkl ͒ be the outcome on the transformed scale, where t(.) is the appropriate transformation mentioned earlier (26), a linear mixed ANOVA model for each individual peptide j is fitted as follows: where Time ͑T ͒ jk and Disease ͑D͒ jl and their interaction ͑TD͒ jkl are taken as the fixed effects of interest, whereas a Patient effect ͑P͒ ij is taken as a random effect, and where j represents the average signal intensity for that peptide across all factors and observations, and the error term ijkl is assumed to be independent from the random factor ͑P͒ ij , both assumed to be normally distributed with mean 0 and some (unknown) variance components: ijkl xN͑0,͚͒; and ͑P͒ ij xN͑0, P 2 ͒. In this experimental design, contrasts were built for each of the fixed effects of interest, with coefficients to be estimated. Variables were ranked in order of evidence of differential expression. Corresponding p values were adjusted for multiple testing using a recent extension of the standard Benjamini-Hochberg procedure, which controls the expected False Discovery Rate (FDR) (34). This error rate, called the positive FDR (denoted pFDR) (29,30), resulted in a procedure less conservative than the conventional FDR. As well as statistical significance, we report the estimated log 2 transformed Fold Change (log 2 FC or M ϭ log 2 (FC)). This represents the log-ratio (log 2 FC) for individual peptide between two experimental conditions in the case of a main effect and to a difference in log-ratios in the case of an interaction effect. An estimated average log 2 FC of -1 and 1 correspond to a 1 ⁄2 and twofold change respectively. Moderated t-, and Bstatistics are also reported, these represent different measures of statistical significance. The Moderated t-statistic corresponds to the usual t-statistic except that information has been borrowed across variables (peptides), whereas the B-statistic is the empirical Bayes log 2 of the posterior odds that the peptide is differentially expressed (35). Finally raw and adjusted p values (with FDR correction) are listed. Note that in every list all the peptides are ranked by adjusted p value and then by B-statistic. Implementations and algorithms of were from the freely available consortium CRAN. For linear modeling and supervised inferences, we used the package "limma" (36). Finally, for the control of the positive FDR, we used the package "qvalue" (37).
Predictive Model Building-Our multi-class prediction model uses a combination of "Spike and Slab" linear regression and Random Forest multiclass prediction models (38). Briefly, a multiclass random forest set of predictor variables (peptides) was built by following Ishwaran and Rao's approach for categorical response variable, where a combination of Spike and Slab Variable Selection (38 -40) and Random Forest Multiclass Classifier (41) is used. A pseudo Y response vector of response values is generated using each of the top N principal components (n ϭ 10) of the peptide expression matrix X. The Generalized Elastic Net (42-44) obtained from using Spike and Slab linear regression (38 -40, 45) is used to select the peptides, one regression fit at a time for each of the pseudo Y response vector. The final combined set of peptides is passed unto a random forest classifier to construct a multiclass random forest set of predictor peptides. This procedure is repeated B ϭ 100 times to get "Variable Signatures" and "Model Sizes," and estimate total and within-class "Prediction Error Rates." Each Monte-Carlo replicate is based on balanced K-fold cross-validation (K ϭ 3) with the data randomly split into a (KϪ1)/K set used for training and 1/K set for testing. For classification, the Prediction Error Rate is measured by the Misclassification Error Rate computed over the hold-out data (1/K). Full cross-validation was performed by including all steps of peptide selection and model construction (classifier) in order to avoid risks of selection bias and Prediction Error Rates bias. Classification models were selected by simultaneously minimizing the misclassification error rate (MER) and maximizing the area under the ROC curve (AUC). Note that the performance of a model peaks at a certain model size (models III and V in our results, Fig. 5) as overly parsimonious models tend to underfit the data (excess bias), and overly large model tend to overfit the data (excess variance) because of a degradation of the signal to noise ratio when marginally relevant proteins (variables) are added to the model. Computations were performed using the recent implementation of Spike & Slab models in the R library called "Spike-Slab," available from the CRAN repository.
Ingenuity Pathway Analysis (IPA)-Following statistical analysis, significant peptides (p Ͻ FDR ϭ 0.1) and their corresponding estimated log-ratios were imported into Ingenuity Pathway Analysis (Ingenuity Systems, Redwood City, CA). Once imported, a single log ratio value was generated at the protein level by determining the mean for all peptides identified for a specific protein.

RESULTS
Label-free Expression Analyses-The sample depletion protocol was reproducible across individual samples and yielded sufficient protein for label-free expression. Reproducible protein patterns via 1D-SDS-PAGE were observed across all samples for both flow-through and bound fractions (data not shown). These depleted samples were subsequently digested and analyzed by liquid chromatography tandem MS (LC/MS/MS). Fig. 1 highlights the chromatographic reproducibility in terms of intensity and retention time among samples collected from controls on Day 0 of SCT and 14 and IPS patients on Day 0 of SCT and the drastic differences in the chromatographic profile between these 3 groups compared with samples obtained from IPS patients at diagnosis. In addition, correlation analysis was performed on the peptide intensities analyzed in the predictive model label-free study for both biological and technical replicates. Excellent reproducibility was observed in this analysis for both technical and biological samples with mean correlations among control patient peptides of 0.94 and mean correlations of 0.90 for IPS patient peptides whereas technical replicates had mean correlations of 0.99.
Additional detail for these correlations can be found in supplemental Fig. S3.
Discovery Principal Component Analysis, Statistical Inferences and FDR Corrections-Grouping and outlier detection among the samples were evaluated by a PCA and clustering analysis (Fig. 2). The PCA biplot illustrates the separation of data into two distinct groups: one that included samples collected from patient controls at Day 0 and Day 14 and the IPS group at Day 0 whereas the other group contained all samples from the IPS patients at diagnosis. In addition, two potential outliers were observed in the IPS groups from Day 0 and diagnosis. These two samples were identified as biological outliers as they were collected from the same patient, but they were not removed from the analysis.
The goal of the discovery study was to observe changes of protein expressions over time in the disease versus the control group. We therefore fitted the same linear mixed-effect model of a two-way analysis of variance univariately to each individual peptide. In this design, the interaction effect is reported as this contrast evaluates the Disease effect over Time. supplemental Table S1 highlights the peptides that were determined as having a significant change. Using a pFDR correction cutoff of 10%, 290 peptides were up (160) or down (130) regulated and no more than 29 of these selected peptides are expected to be false positives across the interaction contrast. These peptides represent 81 proteins whose abundance significantly changes over time with disease (supplemental Table S1). The volcano plot in Fig. 3 summarizes the distribution of peptides that were found to be significantly down (green) or up (red) regulated in this interaction. Table I highlights the subset of 81 proteins (from the peptides in supplemental Table S1) that met our confirmation criteria for protein identification described in Methods. Specific examples of significant immune-related proteins seen in the interaction set include lipopolysaccharide binding protein (LBP), lumican (LUM), attractin (ATRN), and multiple negative and positive acute phase reactant proteins like complement, apolipoprotein H (APOH), and C-reactive protein (CRP).
Ingenuity Pathway Analysis (IPA)-All significant proteins (interaction) were analyzed by IPA to understand pathways associated with the development of IPS in human subjects. Of the 81 proteins from Table I, the software mapped 72 of the proteins. Table II highlights the top canonical pathways, the number of proteins identified in each pathway, and their significance in the analysis. Many pathways and dysregulated proteins relate to inflammation and immune-dysregulation with the acute phase response being the top ranked canonical pathway. Fig. 4 summarizes the major connections in the acute phase response (TNF␣/IL-6) signaling pathway and illustrates the corresponding acute phase proteins identified in the discovery study as significantly changing. The data indicates strong correlations observed for changes in plasma abundance of both negative (e.g. those that decrease) and FIG. 2. PCA analysis plots for labelfree analysis. Two-dimensional bi-plot that uses the first two principal components to display information about the variables (peptides) and the relationship between samples as indicated by intersample distances. The position of any sample point relative to the variables reflects the score of that sample on the variables. IPS patients at diagnosis (D.14d) are distinctly separated from all other experimental groups: IPS at day 0 (P.0d), controls at day 0 (C.0d), and controls at day 14 (C.14d). Variables contributing very little to the principal components have small loading values and are plotted around the center of the plot. Variables that contribute most to distinguishing the sample groups are plotted around the borders of the plot.

FIG. 3. Volcano plots of interaction effect.
The t test volcano plot arranges peptides by statistical significance (vertical axis) and magnitude of change (horizontal axis). The horizontal axis represents the estimated log fold change (denoted log 2 (FC) or M) and the vertical axis represents the log odds of differential expression (denoted log 2 (Odds) or B). The most significant peptides are those that have the largest absolute log fold change and the highest log odds score. Among them, the most significant peptides found by regularized ANOVA (highlighted in red (up) and green (down)) are those distributed in the top right or left of the plot. positive (e.g. those that increase) acute phase reactants as a function of disease progression. For example, six of the eight negative acute phase reactants identified as significant in the interaction effect were observed to decrease in abundance whereas the plasma expression for 16 of 28 positive acute phase reactants identified were increased. Notably, this pathway is known to be dysregulated in mouse models of IPS (see above), thus this network analysis confirms the relevance of this response in humans. Additional canonical pathways identified in this analysis that are associated with immune modulation included: complement and coagulation system and pattern recognition of bacteria and viruses (Table II). Overall, the data suggests coordinate dysregulation in these associated pathways occurs during the development of IPS. Predictive Models for Disease Progression to IPS and Response to Etanercept Therapy-The label-free expression analysis to derive prognostic and predictive biomarkers included 1722 peptides that passed identical prefiltering steps. These peptide variables were subjected to predictive model building as outlined in Methods. A total of eight different models, in which the number of peptide classifiers was varied, were evaluated for each prediction (disease and responsiveness). MER and AUC values are reported in Fig. 5. The performance ranged from a maximum AUC of 97% with 48 proteins to a minimum AUC of 80% using just three proteins in the model. Model III had the highest overall performance for prognosis (AUC ϭ 97%) and included 95 peptides corresponding to 48 different proteins. Model V had the highest predictive (drug responsiveness) performance (AUC ϭ 95%) and utilized 24 peptides corresponding to nine proteins. Moreover, these models accurately classified all patients at day of SCT who would progress to IPS and identified the subset that would respond to etanercept treatment (no false negatives). Supplemental Tables S2 and S3 highlight the peptides that were used in both models. Table I highlights the proteins that were identified as significant in the discovery analysis and were also found in the predictive models for progression to IPS and/or response to therapy. Interestingly, a number of proteins in the acute phase response signaling (APRS) pathway, which was the top canonical pathway in the discovery study, also showed significant differences in abun- FIG. 4. The acute phase canonical pathway and corresponding acute phase reactants identified as significant for interaction effect. Proteins highlighted in green correspond to a decrease in abundance and proteins highlighted in red correspond to an increase in abundance. dances at day 0 when comparing samples from control versus IPS patients as well as responders versus nonresponder subgroups. For instance both negative (APOH and histidine-rich glycoprotein-HRG) and positive (serum amyloid P component-APCS, complement, CRP, and Alpha 1 antichymotrypsin-SERPINA3) acute phase reactants identified in the discovery study were found in the diseases and response prediction models. Fig. 6 highlights the mean abundance differences for selected proteins analyzed in the predictive model experiments that were interesting from the standpoint of inflammatory biology. For these proteins, T-tests were performed on the mean intensities between control and IPS patients and the responder versus nonresponder subset. Acute phase proteins APCS, inter-alpha trypsin inhibitor heavy chain 2 (ITIH2), and transrthyretin (TTR) were very significantly different in the control versus IPS comparison (p Ͻ 9e Ϫ04 ). Moreover, C3, ATRN, and LUM proteins, which are involved in the innate immune response were significantly different when comparing the responder versus nonresponder subset (p Ͻ 1e Ϫ08 ). The significant changes in abundance of these proteins suggest that dysregulation of innate immunity (and specifically the APRS pathway) occurs in patients who develop IPS and can be observed as early as the day of SCT. Therefore, these proteins have significant clinical value as they are candidate predictive biomarkers for patients with respect to both the development of disease (IPS) and/or responsiveness to (future) treatment. DISCUSSION IPS remains a leading cause of treatment-related mortality following allogeneic SCT, but our understanding of the immune dysregulation responsible for pulmonary inflammation engendered in this context remains limited. Murine IPS mod-FIG. 5. Predictive models for disease and response to etanercept therapy. A, Performance versus model size plots of disease (IPS) and responsiveness predictive models. Classification models were optimized by simultaneously minimizing the misclassification error rate (MER) and maximizing the area under the curve (AUC). Red highlights the AUC and green highlights the MER for each model. Loess stands for the scatter plot smoothing curve. B, Disease and responsiveness modeling summary results. A total of 46 samples from nine control patients and 14 disease patients were analyzed in the disease model. A total of 28 samples from nine patients who responded to therapy (responders) and five IPS patients who did not respond to therapy (nonresponders) were analyzed in the responsiveness models. els suggest that the lung is a target of two distinct, but inter-related pathways of immune-mediated injury: one involving soluble inflammatory effectors including TNF␣, and the other driven by antigen specific, donor T-cell effectors (46). Based directly upon this preclinical work, we recently completed an early phase clinical trial using a soluble TNF␣ binding protein, etanercept (Enbrel®, Amgen, Thousand Oaks, CA) for the treatment of IPS. Results from this trial showed that the combination of etanercept and corticosteroids was associated with high response rates, improved survival, and acceptable toxicity in SCT recipients with IPS (9). The current report extends these initial biologic findings and FIG. 6. Box-whisker plots for protein abundance at day of SCT for selected proteins across the clinical groups. The x axis for each plot represents the different groups of patients: Cont (red) ϭ controls, Resp (green) ϭ IPS patients who responded to etanercept therapy and NonResp (blue) ϭ IPS patients who did not respond to therapy. The y axis is the mean intensity of peptides (resulting in a protein value) after data transformation in a log 2 scale as described in Materials and Methods. The full range of data shown for each protein includes the median, quartile limits, and variability in the group medians being compared (small angle of the indentation reflects small variance). In each box plot, the notches give an estimate of the 95% confidence interval for the median. If the notches of two box plots do not overlap there is strong evidence that the two medians differ significantly. The students t test was used for two-group comparisons of selected proteins and p values for control versus IPS (responders and nonresponders combined) and responders versus nonresponders are reported. (nonsignificant ϭ "p Ͼ 0.05"; * ϭ "0.01 Ͻ p Ͻ 0.05"; ** ϭ "0.001 Ͻ p Ͻ 0.01"; *** ϭ "0.0001 Ͻ p Ͻ 0.001"; **** ϭ "0.00001 Ͻ p Ͻ 0.0001"; ***** ϭ "p Ͻ 0.00001"; n ϭ 5 to 9 per group). APCS ϭ Serum amyloid P-component, ITIH2 ϭ Interalpha trypsin inhibitor heavy chain 2, TTR ϭ Transthyretin, ATRN ϭ Attractin, C3 ϭ Complement 3 and LUM ϭ Lumican. further characterizes and compares the plasma proteome of patients with IPS to control patients who have an uncomplicated course early post-SCT.
One canonical pathway identified from the IPA network analysis as disease associated was the APRS pathway. The APRS pathway is a key component of the innate immune system that is triggered by disturbances in homeostasis resulting from injury and/or infection. The APRS pathway involves the altered expression of proteins synthesized in the liver whose plasma concentrations increase (positive acute phase reactants) or decrease (negative acute phase reactants) in response to circulating inflammatory cytokines such IL-6, IL-1, and TNF␣ (47)(48)(49). For example, TNF␣ binds to its receptors, TNFRI and TNFRII, and initiates series of signaling cascades, which culminates in the activation of nuclear factor kappa B (NF-kB) a key transcription factor in the acute phase response. The activation of NF-kB alters the transcription of inflammatory molecules that is ultimately reflected by changes in their plasma concentrations (Fig. 4). Similarly, IL-6 can also serve as potent inducer of the APRS pathway. Upon binding to the IL-6 receptor, IL-6 stimulates hepatocytes to produce acute phase response proteins through a series of signaling cascades including the activation of STAT3. Thus, specific serum proteins can serve as "readouts" of the APRS pathway and provide insights into the status of key cytokines involved in its regulation (50). ELISA assays have shown significant increases in both BAL fluid and plasma concentrations of TNF␣ and IL-6 in IPS patients (9), and the proteome readout observed in this report provides additional evidence for broad-based activation of the APRS pathway as seen by the dysregulated expression of many relevant downstream molecules (Fig. 4).
Activation of the APRS pathway in the clinical setting as revealed here also reinforces experimental data establishing a causal role for TNF␣ in the development of IPS. TNF␣ levels are elevated in the BAL fluid of mice with IPS, and the administration of a soluble, dimeric, TNF binding protein (rhTNFR: Fc; Amgen Corp, Thousand Oaks, CA) from week 4 to week 6 after SCT reduced the progression of lung injury during this time period (4,5). The use of TNF␣ deficient mice has shown that lung injury after allogeneic SCT is dependent upon donor, rather than host derived TNF␣ and that cytokine production from both donor accessory cells (macrophage/monocytes) and T cells significantly contributes to this toxicity (6). Donor derived TNF␣ serves as both an effector and facilitator of lung injury. TNF␣ contributes to pulmonary vascular endothelial cell apoptosis that accompanies pulmonary inflammation during the development of IPS (8). In addition, TNF␣ secreted by donor T cells regulates the chemokine milieu in the lung within the first 2 weeks after SCT, which directly contributes to the subsequent recruitment of monocytes and macrophages as lung injury progresses (6). Recent studies have revealed that receptor/ligand interactions between TNF␣ and TNFRII contribute to leukocyte recruitment during IPS by regulating pulmonary ICAM-1 expression (7).
The innate immune response is initiated in large part by antigen presenting cells (APCs) and myeloid-derived accessory cells (macrophages, monocytes, and neutrophils) both of which require activation via pattern recognition of evolutionary conserved structures on pathogens (51). These structures are also known as pathogen-associated molecular patterns (PAMPs) and are detected via pattern recognition molecules such as Toll like receptors (TLRs) and LBP (52,53). Once detected, the engagement of PAMPs with TLRs on the cell surface triggers a cascade of signaling events that culminate in the production of proinflammatory cytokines (54). Examples of PAMPs include lipopolysaccharide (LPS), lipoteichoic acid, peptidoglycan, and mannan (51,53). LPS, also known as endotoxin, is part of the outer membrane of Gram-negative bacteria and is a potent PAMP responsible for the inflammatory response during endotoxic shock (55). Circulating LPS liberated from Gram-negative bacteria binds to LBP in the blood stream and forms a complex with its co-receptor CD14 on the cell surface enabling transfer of LPS to MD2 and subsequently to TLR4, which is followed by TLR4 oligomerization and signaling (54). Experimental studies in animal models have suggested a critical role for LPS in cytokine amplification observed during the development of IPS. LPS levels are elevated in the BAL fluid of mice with IPS, and the intravenous administration of LPS to mice with advanced GVHD significantly increases the severity of lung injury (4,5). The enhanced inflammation is associated with large increases of TNF␣ and LPS in the BAL fluid and the development of alveolar hemorrhage (5). Furthermore, direct antagonism of LPS early in the time course of SCT reduces systemic levels of TNF␣ and significantly decreases the severity of GVHD and IPS (56), a finding that is linked to the responsiveness of donor (rather than host) cells to LPS stimulation (57).
As a number of pattern recognition molecules were identified as significant (for interaction) in the label-free expression study (Table I), and pattern recognition of bacteria and viruses was a significant annotated pathway identified by the network analysis (Table II) we illustrate specific examples of protein level changes of pattern recognition molecules in Fig. 7. In particular, LBP levels are seen to have increased significantly for IPS patients. To verify the changes in this protein, we conducted ELISA experiments on LBP for a cohort of 24 patients from the clinical trial (9). These data in Table III show a roughly fivefold increase in LBP values at diagnosis versus either controls (day 14) or IPS patients at day 0. Although Fig.  7 also shows a significant increase in LBP as a function of IPS diagnosis, we wished to also evaluate the level of fold change in intensity from the proteomics data for more direct comparison to ELISA data. Thus, we summed the individual unnormalized intensities for the four LBP peptides to create a surrogate protein level intensity of LBP for the same three groups as examined for ELISA. These results are also shown in Table III and reflect a similar fold change to the ELISA data. Thus, by multiple means of measurement, LBP is seen to be significantly increased in IPS patients, and may represent a robust molecular diagnostic for the condition with a straightforward assay method already available. Biologically, LBP is a well characterized acute phase protein that binds endotoxin (58,59). TLR4 activation and sensitivity to endotoxin relies on a series of interactions starting with LBP that enables optimal presentation of endotoxin to TLR4 (60,61). In addition to bacterial recognition, LBP also can indirectly regulate the inflammatory response. Numerous studies in LBP knockout mice have shown that the absence of LBP leads to reduced LPS responsiveness and immune activation. Thus, it is an interesting candidate for future evaluation.
Peptidoglycan recognition protein 2 (PGLYRP2) and lumican are pattern recognition proteins that had similar interaction profiles in our analysis; both showed increases in abun-dance in IPS patients at day 0 and a significant decrease at diagnosis when compared with controls at the same time point (Fig. 7). Whereas LBP has a well-established role in the innate immune response, the biologic functions of PGLYRP2 and lumican have only recently been identified. PGLYRPs participate in the innate immune response and contain a well conserved peptidoglycan binding type 2 amidase domain across most animals (62)(63)(64). Mammalian PGLYRP2 is produced by the liver, and initial studies suggested that it functioned as an anti-inflammatory scavenger of peptidoglycan (65). However, recent studies in a peptidoglycan-induced arthritis animal model have demonstrated that PGLYRP2 has pro-inflammatory properties; PGLYRP2 was found to locally modulate the inflammatory response via cooperation with other pattern recognition molecules specifically Nod2 and TLR4 both of which are important regulators of inflammatory cytokine production (66). In our study, four peptides of PGLYRP2 were identified in the label-free analysis as significant for the interaction effect (Fig. 7).
Lumican is a small, leucine rich, extracellular matrix glycoprotein that has recently been shown to be involved in the innate immune response (67,68). A study that investigated the role of lumican in wound healing found that mice deficient in lumican (Lum Ϫ/Ϫ ) had poor recruitment of macrophages to the site of injury and a significant decrease in the induction of inflammatory cytokines (69). When challenged with LPS, Lum Ϫ/Ϫ mice are resistant to septic shock and death and are poor inducers of TNF␣ and IL-6 (70). This study also showed that lumican expression is induced during the innate immune FIG. 7. Interaction profile plots of top selected proteins. Interaction plots of four selected proteins for ATRN ϭ Attractin, LBP ϭ Lipopolysaccharidebinding protein, LUM ϭ Lumican and PGLYRP2 Peptidoglycan recognition protein 2. The x axis for each plot represents the time point and the y axis is the intensity after data transformation in a log 2 scale as described in Materials and Methods. Individual peptide intensities can be found in supplemental Table S1. The solid line represents the median intensity value for peptide (s) (resulting in a protein value) and the dotted line is their estimated 95% confidence interval for the median difference. Blue represent values for patients with IPS and red represents values for control subjects. response; LPS binds to lumican and co-precipitates with CD14, which is a key regulator of LPS sensing. Taken together, these data suggest lumican may enhance host sensitivity to microbial products like LPS. In the label-free analysis, four peptides for lumican were identified as significant for the interaction effect. Five of the nine IPS patients had higher levels at day 0 than controls at the same time point. A ϳ50% increase in average peptide abundance (non log scale) was observed for these subjects when compared with other subjects analyzed in this study. Interestingly, these IPS subjects all responded to etanercept therapy. In addition to PAMP related proteins, attractin, a dipeptidyl peptidase IV-(DPP IV) like enzyme was significant in this analysis and shared a similar interaction profile with lumican and PLGYRP2 (Fig. 7). Attractin is a CUB (complement C1r/ C1s, Uegf, Bmp1) family protease that has activity similar to DPP IV, and its activity is conserved in both the secreted and transmembrane forms (71). Interestingly, attractin is not expressed on resting T cells but is highly expressed on activated T cell and peripheral blood monocytes (71,72). The enzymatic activity of attractin plays an important role in the cellular responses to stimuli as well as serving as modulator of T-cell activation and extravasation (71,73,74). Attractin may also aide in the formation of immuno-regulation clusters by a direct interaction with monocytes and T cells or by regulating the levels of circulating chemotactic cytokines (71); recent work conducted on peripheral blood monocytes demonstrate that attractin assists in adhesion to fibronectin and cytokine production in these cells (72). Many of the proteins discussed above, including the pattern recognition molecules, were significant for interaction in the discovery analysis and also good classifiers for disease and/or response to therapy. In this context, attractin peptides ranked in the top 10% of all peptide classifiers for response to therapy and had significant intensity values when comparing responder versus non responders (Fig. 6). Taken together, this data suggest that early immune dysregulation (as early as day 0 of SCT) can be identified in patients who ultimately develop IPS and when evaluated simultaneously, changes in key proteins can be leveraged to classify patients for effective treatment.
In sum, preclinical and early phase clinical findings supporting a role for TNF␣ in the development of IPS have paved the way for the development of two multicenter trails that are currently accruing pediatric (phase II trial) and adult (phase III trial) patients. The data reported herein extend this ongoing translational research effort utilizing label-free proteome expression analysis. We have characterized the plasma proteome of patients with and without IPS on day 0 and at the day of diagnosis of IPS or day 14 following SCT for controls. We have successfully identified a number of protein changes associated with IPS that are involved in innate immunity, pattern recognition and the APRS pathway many of which are consistent with observations from preclinical models whereas others have not previously been biologically linked to this disease process. Collectively, our results confirm and broaden what is known about the pathophysiology of human disease and firmly support mechanistic insights fostered by animal models. Moreover, statistical modeling identified a panel of proteins that can identify patients on day 0 of SCT who will ultimately develop IPS and predict their responsiveness to TNF␣ neutralization strategies. Importantly, insights generated from the proteomic analysis of samples collected at the bedside can now be (1) brought back to the bench to evaluate the specific contribution of novel proteins to the inflammation engendered during IPS using established animal systems, (2) used to direct validation efforts in other SCT patient cohorts, including those under treatment by etanercept, and ultimately (3) incorporated into the intelligent design of future clinical trials wherein stratification by anticipated response to treatment can form the basis of individualized therapy and improved outcomes for SCT recipients and a multitude of other patients with inflammatory conditions. The latter highlights the potential broader implications of our results.