Genomic responses in mouse models poorly mimic human inflammatory diseases

Junhee Seoka,1, H. Shaw Warrenb,1, Alex G. Cuencac,1, Michael N. Mindrinosa, Henry V. Bakerc, Weihong Xua, Daniel R. Richardsd, Grace P. McDonald-Smithe, Hong Gaoa, Laura Hennessyf, Celeste C. Finnertyg, Cecilia M. Lópezc, Shari Honarif, Ernest E. Mooreh, Joseph P. Mineii, Joseph Cuschierij, Paul E. Bankeyk, Jeffrey L. Johnsonh, Jason Sperryl, Avery B. Nathensm, Timothy R. Billiarl, Michael A. Westn, Marc G. Jeschkeo, Matthew B. Kleinj, Richard L. Gamellip, Nicole S. Gibranj, Bernard H. Brownsteinq, Carol Miller-Grazianok, Steve E. Calvanor, Philip H. Masone, J. Perren Cobbs, Laurence G. Rahmet, Stephen F. Lowryr,2, Ronald V. Maierj, Lyle L. Moldawerc, David N. Herndong, Ronald W. Davisa,3, Wenzhong Xiaoa,t,3, Ronald G. Tompkinst,3, and the Inflammation and Host Response to , Large Scale Collaborative Research Program4 aStanford Genome Technology Center, Stanford University, Palo Alto, CA 94305; Departments of bPediatrics and Medicine, sAnesthesiology and Critical Care Medicine, and tSurgery, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114; cDepartment of Surgery, University of Florida College of Medicine, Gainesville, FL 32610; dIngenuity Inc., Redwood City, CA 94063; eDepartment of Surgery, Massachusetts General Hospital, Boston, MA 02114; fDepartment of Surgery, Harborview Medical Center, Seattle, WA 98195; gShriners Hospitals for Children and Department of Surgery, University of Texas Medical Branch, Galveston, TX 77550-1220; hDepartment of Surgery, University of Colorado Anschutz Medical Campus, Denver, CO 80045; iDepartment of Surgery, Parkland Memorial Hospital, University of Texas, Southwestern Medical Center, Dallas, TX 75390; jDepartment of Surgery, Harborview Medical Center, University of Washington School of Medicine, Seattle, WA 98195; kDepartment of Surgery, University of Rochester School of Medicine, Rochester, NY 14642; lDepartment of Surgery, University of Pittsburgh Medical Center Presbyterian University Hospital, University of Pittsburgh, PA 15213; mDepartment of Surgery, St. Michael’s Hospital, University of Toronto, Toronto, ON, Canada M5B 1W8; nDepartment of Surgery, San Francisco General Hospital, University of California, San Francisco, CA 94143; oDivision of Plastic and Reconstructive Surgery, Department of Surgery, University of Toronto, Toronto, ON, Canada M4N 3M5; pDepartment of Surgery, Stritch School of Medicine, Loyola University, Chicago, IL 60153; qDepartment of Anesthesiology, Washington University, School of Medicine, St. Louis, MO 63110; and rDepartment of Surgery, University of Medicine and Dentistry of New Jersey-Robert Wood Johnson Medical School, New Brunswick, NJ 08903

Contributed by Ronald W. Davis, January 7, 2013 (sent for review December 6, 2012) A cornerstone of modern biomedical research is the use of mouse obtained from serial blood draws in 167 patients up to 28 d after models to explore basic pathophysiological mechanisms, evaluate severe (15), 244 patients up to 1 y after injury, new therapeutic approaches, and make go or no-go decisions to and 4 healthy humans for 24 h after administration of low-dose carry new drug candidates forward into clinical trials. Systematic bacterial endotoxin (14) and expression analysis on analogous studies evaluating how well murine models mimic human inflam- samples from well-established mouse models of trauma, , matory diseases are nonexistent. Here, we show that, although acute and endotoxemia (16 treated and 16 controls per model) (16–18). inflammatory stresses from different etiologies result in highly In humans, severe inflammatory stress produces a genomic storm similar genomic responses in humans, the responses in correspond- affecting all major cellular functions and pathways (15) and ing mouse models correlate poorly with the human conditions and therefore, provided sufficient perturbations to allow comparisons fi also, one another. Among genes changed signi cantly in humans, between the genes in the human conditions and their orthologs in the murine orthologs are close to random in matching their human the murine models. R2 counterparts (e.g., between 0.0 and 0.1). In addition to improve- In this article, we report on a systematic comparison of the ge- ments in the current animal model systems, our study supports nomic response between human inflammatory diseases and murine higher priority for translational medical research to focus on the models. First, we compared the correlations of gene expression more complex human conditions rather than relying on mouse mod- changes with trauma, burns, and endotoxemia between human els to study human inflammatory diseases. subjects and corresponding mouse models. Second, we character- ized and compared the temporal gene response patterns seen in human disease | translational medicine | inflammation | these human conditions and models. Third, we also identified the immune response | injury major signaling pathways significantly regulated in the inflammatory response to human and compared them with the human urine models have been extensively used in recent decades in vivo endotoxemia model and three murine models. Fourth, we Mto identify and test drug candidates for subsequent human trials (1–3). However, few of these human trials have shown suc- cess (4–7). The success rate is even worse for those trials in the field fl Author contributions: M.N.M., S.F.L., R.V.M., L.L.M., D.N.H., R.W.D., W. Xiao, R.G.T., and of in ammation, a condition present in many human diseases. To I.H.R.I.,L.S.C.R.P. designed research; J. Seok, H.S.W., A.G.C., M.N.M., H.V.B., W. Xu, H.G., date, there have been nearly 150 clinical trials testing candidate L.H., C.C.F., C.M.L., S.H., E.E.M., J.P.M., J.C., P.E.B., J.L.J., J. Sperry, A.B.N., T.R.B., M.A.W., agents intended to block the inflammatory response in critically ill M.G.J., M.B.K., R.L.G., N.S.G., B.H.B., C.M.-G., S.E.C., P.H.M., J.P.C., L.G.R., R.V.M., L.L.M., – D.N.H., W. Xiao, and R.G.T. performed research; J. Seok, W. Xu, D.R.R., H.G., R.W.D., and patients, and every one of these trials failed (8 11). Despite com- W. Xiao contributed new reagents/analytic tools; J. Seok, A.G.C., H.V.B., W. Xu, D.R.R., mentaries that question the merit of an overreliance of animal H.G., C.C.F., C.M.L., and W. Xiao analyzed data; and J. Seok, H.S.W., A.G.C., M.N.M., H.V.B., systems to model human immunology (3, 12, 13), in the absence of W. Xu, G.P.M.-S., L.L.M., W. Xiao, and R.G.T. wrote the paper. systematic evidence, investigators and public regulators assume The authors declare no conflict of interest. that results from animal research reflect human disease. To date, Freely available online through the PNAS open access option. MEDICAL SCIENCES there have been no studies to systematically evaluate, on a mo- 1J. Seok, H.S.W., and A.G.C. contributed equally to this work. lecular basis, how well the murine clinical models mimic human 2 fl Deceased June 4, 2011. in ammatory diseases in patients. 3 fl To whom correspondence may be addressed. E-mail: [email protected], wxiao1@ The In ammation and Host Response to Injury, Large Scale partners.org, or [email protected]. Collaborative Research Program has completed multiple studies 4A complete list of the Inflammation and Host Response to Injury, Large Scale Collabo- on the genomic responses to systemic inflammation in patients and rative Research Program can be found in SI Appendix. – human volunteers as well as murine models (14 18). These data- This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. sets include genome-wide expression analysis on white blood cells 1073/pnas.1222878110/-/DCSupplemental. www.pnas.org/cgi/doi/10.1073/pnas.1222878110 PNAS | February 26, 2013 | vol. 110 | no. 9 | 3507–3512 Human Burn 0.91 0.47 0.08 0.05 0.00

Human Trauma 0.47 0.08 0.05 0.00 −3 0 3

Human Endotoxemia 0.08 0.09 0.01 −3 0 3

Mouse Burn 0.13 0.01 −3 0 3 Fig. 1. Correlations of the gene changes among human burns, trauma, and endotoxin and the corresponding Mouse mouse models. Scatter plots and Pearson correlations (R2) Trauma 0.00 of the log twofold changes of 4,918 human genes re- sponsive to trauma, burns, or endotoxemia (FDR < 0.001; −3 0 3 fold change ≥ 2) and their murine orthologs in the murine models. As shown in the upper left, the genomic responses 2 = Mouse to human trauma and burns are highly correlated (R Endotoxemia 0.91). In contrast, as shown in the lower right, the murine models correlate poorly with each other (R2 = 0.00–0.13)

−3 0 3 and almost randomly with the corresponding human con- ditions (R2 = 0.00–0.09). Similar results were seen with rank −3 0 3 −3 0 3 −3 0 3 −3 0 3 −3 0 3 correlation (SI Appendix, Fig. S1). sought and evaluated representative patient and murine studies of are highly similar in humans, but these responses are not repro- several additional acute inflammatory diseases. These results show duced in the current mouse models. New approaches need be ex- that the genomic responses to different acute inflammatory stresses plored to improve the ways that human diseases are studied.

A Human Mouse B Burn Trauma Endotoxemia Burn Trauma Endotoxemia Year Month Day Week Week Day 0h 1y 0h 1m 0h 1d 0h 1w 0h 1w 0h 1d 1

e

2 miT

yrevoc

e

R 3

4 0h 1d 7d 1m 3m 1y Human Human Human Mouse Mouse Mouse Burn Trauma Endotoxemia Burn Trauma Endotoxemia

5 C HLA-DRA

6 log2(Expr) 7 Human Burn Human Trauma Human Endotoxemia Mouse Burn Mouse Trauma

8 −3 −2 −1 0 Mouse Endotoxemia

0h 6h 12h 1d 3d 7d 14d 1m 3m 6m 1y 3- 0 3 Time

Fig. 2. Comparisons of the time-course gene changes for human burns, trauma, and endotoxin and the murine models. (A) K-means clustering of 4,918 genes responsive to human systemic inflammation over time. Time intervals for the human conditions: up to 1 y (burns), 1 mo (trauma), and 1 d (endotox- emia); time intervals for mouse models: 1 wk (trauma and burns) and 1 d (endotoxemia). (B) Box plots of recovery times of gene changes in the human and mouse conditions. (C)Log2 expression changes vs. time of HLA-DRA as an example where genes changed significantly over a long time in human injuries but minimally in the murine models.

3508 | www.pnas.org/cgi/doi/10.1073/pnas.1222878110 Seok et al. Results increase to R2 = 0.11–0.28 among the top 105 genes with the Correlation of Gene Changes to Trauma, Burns, and Endotoxemia greatest magnitude of changes (fold change ≥ 6). Between Human Subjects and Murine Models. Significantly changed To probe further and address the possibility of bias against the genes were identified for each dataset; 5,544 genes were identi- murine models by requiring murine orthologs to correlate with fied as significant [false discovery rate (FDR) < 0.001 and fold human genes responsive in the human conditions, we examined change ≥ 2] (Materials and Methods) between patients and the reverse (human orthologs corresponding to the responsive SI Appendix healthy subjects (4,389 genes in trauma, 3,250 genes in burns, and mouse genes) ( ,Fig.S3). High correlation was main- R2 = 2,251 genes in endotoxemia). Among these 5,554 genes signifi- tained in the human orthologs ( 0.88), whereas there were R2 ≤ cantly changed in human conditions, 4,918 genes had mouse only minimal increases in correlations among mouse genes ( orthologs assayed, which were included in the subsequent analysis. 0.22). To address the potential confounding feature of different sample sizes between the studies, we randomly resampled (SI The maximum fold changes of gene expression were calcu- Appendix A lated in log scale between patients and healthy subjects for each ,Fig.S4 ) comparable small numbers of humans and mice and showed that the correlations remained high between dataset of human burn, trauma, and endotoxemia and between R2 = treated and control groups for the corresponding murine models. human injuries ( 0.85). In addition, with comparable sam- Between two datasets, the agreements of the maximum gene fold ple sizes, the human conditions introduced a more profound genomic response (more than or equal to three times more changes (Pearson correlation and rank correlation between the genes) than the corresponding mouse models (SI Appendix,Fig. two datasets) as well as the directionality of the changes (the S4B). However, intragroup variances remained comparable (SI percentages of genes changed to the same direction between Appendix,TableS1). the two datasets) were compared. Among the 4,918 genes, correlations of the maximum gene Comparison of the Temporal Response Patterns Between the Human R2 changes (Pearson correlation, ) (Fig. 1) and percentages of Conditions and Murine Models. An inherent assumption in the use SI Appendix genes changed in the same direction ( , Fig. S1) in all of murine models to mimic human systemic inflammation is that six conditions indicate that there was very high similarity in gene the time course of injury and repair between the species is sim- R2 = response between human trauma and human burns ( 0.91, ilar. Decisions about timing of sample acquisition, endpoints for 97%), and moderate similarity between human endotoxemia and analysis, or dosing of drugs are dependent on this assumption. R2 = human injuries ( 0.47, 88%). In contrast, there was very low We queried gene changes in the human conditions and mouse correlation of expression changes between human genes in any models over different time spans (days, weeks, months, and years). of the conditions and their mouse orthologs in the mouse models We then analyzed the time-course pattern of the expression of 2 (R ≤ 0.09, ∼47–63%). By random chance, 50% of the genes each gene by clustering (Fig. 2A) and further characterized by between two uncorrelated conditions are expected to change in response time (time for the gene change to reach one-half of its the same direction. Furthermore, there was poor correlation of maximum value) and recovery time (time for the gene to decrease the mouse gene orthologs with each other among the three to one-half its maximum value) (Fig. 2B and SI Appendix,Fig.S5). murine models (R2 ≤ 0.13, ∼39%–63%). Similar results were Although in all conditions, the gene response time occurred within seen when the rank correlation was calculated as in SI Appendix, the first 6–12 h, the recovery time differed dramatically. Genomic Fig. S1. We further examined subsets of 4,918 genes that showed disturbances recovered in the murine models within hours to 4 d an even greater magnitude of changes (SI Appendix, Fig. S2) and but lasted for 1–6 mo or more in patients. For example, HLA- observed a modest increase of correlation between murine DRA changed over a long time period in human injuries but only models and human conditions. For example, the correlations minimally in the murine models (Fig. 2C). In addition, although MEDICAL SCIENCES Fig. 3. Pathway comparisons between the human burns, trauma, and endotoxin and mouse models. Shown are bar graphs of Pearson correlations (R2) for the five most acti- vated and suppressed pathways between the four model systems (human endotoxemia and the three murine mod- els) vs. human trauma and burns. Negative correlations are shown as negative numbers (−R2). Human burn is shown as the reference. In every pathway, human endotoxemia had much higher similarity to human injury than mouse models.

Seok et al. PNAS | February 26, 2013 | vol. 110 | no. 9 | 3509 genes within each cluster changed in same directions over the time course between human conditions (SI Appendix, Fig. S6), there was Human Diseases B great variability between the three murine models, even within the T clusters, suggesting that mouse responses differed not only from S the human conditions but also from one another (SI Appendix, E Fig. S7). E S SS I R Comparison of the Significantly Regulated Pathways Between the Human Conditions and Murine Models. We identified the major signaling pathways significantly regulated in human injuries and compared them with a human endotoxemia model and three T/t: Trauma murine models. In human injuries, most of the pathways up-reg- ulated were related to innate immune response, and those path- B/b: Burn E/e: Endotoxemia ways down-regulated were related to adaptive immunity. Fig. 3 i tb S/s: Sepsis shows the correlations of human burn injury with the three mouse s model systems and human trauma for the five most activated and s R/r: ARDS most suppressed pathways. The correlations depended on the in- ss I/i: Infection dividual pathway compared (SI Appendix, Table S2). For example, e 50% 60%r 70% 80% 90% 100% the mouse endotoxin model, which is often used as a proxy for e Mouse Models inflammation in humans, had a maximum correlation of 0.43 for Percentage of genes changed to the same direction (%) Percentage IL-10 signaling, but it had no or negative correlations for all five of 0.0 0.2 0.4 0.6 0.8 1.0 the most down-regulated pathways. In every pathway, human endotoxemia had much higher similarity to human injury than seen Correlation of fold changes of genes ( R 2) in the mouse models. fl Additional examination of the top 40 pathways shows that the Fig. 4. Comparison of the genomic response to severe acute in ammation median correlations and percentages of genes changed in the from different etiologies in human and murine models. GEO was queried for studies in the white blood cells of additional severe acute inflammatory same direction were 0.16 and 67% for mouse burns, 0.03 and diseases (sepsis, ARDS, and infections) in human and mouse. The resulting 60% for mouse trauma, and 0.00 and 52% for mouse endotox- datasets are listed in Table 1. Shown are correlations (R2; x axis) and di- emia. Therefore, individual gene activation in the human con- rectionality (%; y axis) of gene response from the resulting multiple pub- ditions was not necessarily predicted by the ortholog in the lished datasets in GEO compared with human burn injury. corresponding mouse model in either direction or magnitude [e.g., the toll-like receptor (TLR) pathways] (SI Appendix, Fig. S8). worldwide prevalence of the use of mice to model human in- Comparison of the Genomic Responses in Several Additional Acute flammation. We, therefore, sought and evaluated additional patient Inflammatory Diseases and Mouse Models. We were surprised by the and corresponding mouse model studies from Gene Expression poor correlation between the genomic responses in the mouse Omnibus (GEO) (19) for several other severe acute inflammatory models and those responses in human injuries, especially given the diseases [sepsis, acute respiratory distress syndrome (ARDS), and infections] as listed in Table 1. The fold change of each gene measured was calculated be- Table 1. Studies available in GEO on severe acute inflammatory tween patients and controls in a human study or between treated diseases in human and murine models group and control group in a murine model study. The gene Disease GEO accession R2 Percent response in each dataset was then compared with the 5,554 genes significantly changed in human trauma, burns, and endotoxemia. Human Correlations and directionality of gene response were calculated Burns (as reference) GSE37069 1.00 100 using these common genes between each additional dataset and Trauma GSE36809 0.91 97 the burn injury that was used as the reference. Endotoxemia (test) GSE3284 0.47 88 As shown in Fig. 4 and Table 1, the genomic responses in patients Endotoxemia (verification) GSE3284 0.59 90 correlated well with each other, whereas this response was mim- ARDS GSE10474 0.55 84 icked poorly by the mouse models. The same relationship held for Sepsis GSE13904 0.76 93 the directionality of the gene response. For example, in humans, Sepsis GSE9960 0.64 87 correlations of 0.91 (trauma), 0.55 (ARDS), 0.54–0.79 (sepsis), and Sepsis GSE13015 0.61 86 0.50 (infection) were observed with percentage of the genes Sepsis GSE28750 0.63 85 changing in the same direction of 97%, 84%, 86–92%, and 83%, Acute Infection GSE6269 0.50 83 respectively. For murine orthologs, there were correlations be- Mouse tween 0 and 0.08, with 47–61% of genes changing in the same Burns GSE7404 0.08 60 direction—random chance would predict 50%. In addition, pub- Trauma GSE7404 0.05 61 lished data with radiation injury (20) showed similar results (SI Endotoxemia GSE7404 0.00 47 Appendix, Fig. S9). Endotoxemia GSE5663 0.00 50 ARDS GSE19030 −0.01 48 Discussion Sepsis (CLP) GSE5663 0.03 53 Studying disease in patients is much more complex than study- Sepsis (CLP-Mild) GSE5663 0.02 52 ing model systems. In the trauma patients that we studied, pa- Sepsis GSE19668 0.05 58 tient heterogeneity existed in the most relevant characteristics, Sepsis GSE26472 0.02 55 including demographics, such as age [years, median (M) = 33, Infection GSE20524 0.08 61 middle quartiles (MQ) = 25–44] and sex (male = 63%), severity of = R2 represents Pearson correlation. Negative correlations are shown as −R2. injury as indicated by maximum Abbreviated Injury Scale (M 4, Percent represents the percentages of genes changed to the same direction MQ = 3–5), (M = 33, MQ = 22–41), and between the two datasets. CLP, cecal ligation and puncture. worst base deficit (M = −9.1, MQ = −12.0 to −6.4), variable

3510 | www.pnas.org/cgi/doi/10.1073/pnas.1222878110 Seok et al. amounts of blood received (total transfusion in milliliters; M = the model mimics or fails to mimic the molecular behavior of key 1,900, MQ = 1,050–3,000), and patient clinical outcomes as in- genes, key pathways, or the genome-wide level thought to be im- dicated by survival (96%), multiorgan failure (maximum Marshall portant for the relevant human disease. score; M = 5, MQ = 3–7), time to recovery (days; M = 7, MQ = 4– New approaches could be explored to improve current models. 15), nosocomial infections (54.5%), and hospital length of stay For example, by first requiring comprehensive genomic descrip- (days; M = 21, MQ = 12–32) (15, 21). Similarly, in the burn patient tions in patient studies to define the human disease, the disease- population, significant variations existed in the extent of burn altered pathways could be used as a guide to develop the animal injuries, for example, in terms of total body surface area of burns model. The quality of the animal model could then be determined (TBSA; M = 53%, MQ = 32–74%), percentages of flame burns by how well it reproduces the human disease on a molecular basis (86%) and inhalation injury (52%), and survival (85%). In addi- rather than simply phenotype. In addition, the development of tion, patients received a veritable pharmacopeia of drugs that can synthetic human models by in vitro reconstitution of disease-re- affect their pathophysiologic and genomic response. However, lated cell types or tissues (31) might similarly improve current despite of these heterogeneities, we observed highly consistent disease models. Furthermore, new genomic information, such as genomic response in patients between trauma and burns in con- the availability of personal genomes (32) or exomes (33) to cap- trast to the lack of correlations in the murine models. ture the disease heterogeneity directly from patients or systematic Critically ill patients with different underlying acute injuries interpretation of genome-wide signatures in human diseases (34, have similar-appearing physiological reactions, a condition known 35), will likely complement or short circuit the need for mouse by consensus definition as Systemic Inflammatory Response models in disease discovery and drug development. Notwith- Syndrome (22, 23). An unproven hypothesis central to the pursuit standing, our study supports higher priority to focus on the more of drug targets for this syndrome has been that the molecular complex human conditions rather than relying on mouse models mechanisms underlying this syndrome are similar regardless of to study human inflammatory diseases. initiating etiology. The very high correlation in response between trauma, burns, and endotoxemia in humans strongly supports this Materials and Methods hypothesis and suggests that such an approach may be possible. Patient Enrollment and Sampling. The Inflammation and the Host Response to There are multiple considerations to our finding that transcrip- Injury, Large Scale Collaborative Research Program was funded to study the tional response in mouse models reflects human diseases so poorly, early inflammatory response to serious injuries. Between 2003 and 2009, 1,637 including the evolutional distance between mice and humans, the patients were enrolled at one of seven US Level I trauma centers if they ex- complexity of the human disease, the inbred nature of the mouse perienced a blunt injury associated with prehospital or emergency department model, and often, the use of single mechanistic models (3, 12, 13, systolic hypotension (<90 mmHg) or an elevated base deficit (>6 mEq/L), ≤ 24). In addition, differences in cellular composition between mouse blood transfusion requirement 12 h, and had an Abbreviated Injury Scale score > 2 for any body region, exclusive of the brain. Of these patients, serial and human tissues (24, 25) can contribute to the differences seen in blood samples were obtained from 167 patients for genomic analysis. The the molecular response. Additionally, the different temporal spans first blood sample was taken within 12 h of the injury and 1, 4, 7, 14, 21, and of recovery from disease between patients and mouse models are 28 d after injury. Study subjects were treated under the guidance of standard an inherent problem in the use of mouse models. Late events re- operating procedures developed, implemented, and audited across all par- lated to the clinical care of the patients (such as fluids, drugs, sur- ticipating centers to minimize treatment. Details of the study design are de- gery, and life support) likely alter genomic responses that are not scribed in ref. 15. Similarly, between 2000 and 2009, 244 burn patients were captured in murine models. enrolled at one of four burn centers if they were admitted to a participating The evolution of the immune system for any species is, at least in burn center within 96 h after injury, had burns over at least 20% of the TBSA, part, a direct consequence of the microbe-exerted selection pres- and required at least one excision and grafting procedure. After admission, sure for that species. Recent articles have highlighted tradeoffs study subjects were treated under the guidance of standard operating pro- that species make to balance often opposing evolutionary strate- cedures developed, implemented, and audited across all participating centers to minimize treatment variation. Demographics, clinical outcomes, and gies for resistance vs. tolerance or resilience to infection (26, 27). complications occurring in the intensive care unit and within 28 d posttrauma Relative to the human response, mice are highly resilient to in- and up to 1 y postburns were recorded and described in ref. 15, and they are flammatory challenge (28, 29). For example, the lethal dose of available at http://TRDB.gluegrant.org. In addition, 35 healthy control sub- endotoxin is 5–25 mg/kg for most strains of mice, whereas a dose jects (16–55 y) were recruited between 2004 and 2007. Two control subjects that is 1,000,000-fold less (30 ng/kg) has been reported to cause were studied two times approximately 2 y apart, yielding 37 samples. The shock in humans (30). The extreme sensitivity of humans relative Institutional Review Board of each participating center approved the study. to mice to massive inflammation may result in genomic responses that reach an upper threshold in each human disease, whereas the Human Endotoxemia Model. Eight healthy male and female subjects between resilience of mice may prevent maximum responses and lead to 18 and 40 y of age provided written informed consent. Subjects were i.v. greater heterogeneity. This attenuated murine response also administered either National Institutes of Health Clinical Center Reference suggests that a higher level of injury in mouse models might result Endotoxin Escherichia coli 0113 (CC-RE-Lot 2) at a dose of 2 ng/kg body weight (n = 4, one female and three males) or 0.9% sodium chloride (n = 4, one fe- in a transcriptome response that better mimics the response seen male and three males) over a 5-min period. Blood samples were collected in the patients. For example, a limitation of the current study is before endotoxin infusion (0 h) and 2, 4, 6, 9, and 24 h after infusion. that, in the burn patients, the TBSA of burns (TBSA > 20%, M = 53%, MQ = 32–74%) is, on average, higher than the TBSA of the Murine Injury Models. Male C57BL/6J mice were purchased from Jackson murine model (TBSA = 25%). However, currently, there is no Laboratory. The study was approved by the Institutional Animal Use and Care existing murine model with intensive care support that would al- Committees of each participating institution. The mice were acclimated for at low consistent survival of mice with substantially greater than 25% least 1 wk before use in these experiments at 8 wk of age. Details of the TBSA burn injury. protocols have been described previously. Groups of 12 mice undergoing the MEDICAL SCIENCES As a cornerstone of modern biomedical research, the use of injury protocols were given inhalation anesthesia with isofluorane and then mouse models has dominated scientific literature (3, 13). The subjected to 25% TBSA scald burn or trauma/hemorrhage (T/H). After burn prevailing assumption—that molecular results from current mouse injury, mice were resuscitated with 1 mL normal saline injected i.p. T/H consisted of laparotomy followed by withdrawal of sufficient blood from an models developed to mimic human diseases translate directly to — arterial line to decrease and maintain mean arterial blood pressure at 35 human conditions is challenged by our study. Because virtually mmHg for 90 min, after which times these mice were resuscitated with Ringer every drug and drug candidate functions at the molecular level, lactate (four times the shed blood volume). Groups of 12 mice also underwent one practical approach forward is to raise the bar by requiring sham burn or sham T/H procedures under anesthesia. In separate experi- molecular detail in the animal model studies indicating whether ments, groups of six injured or sham mice were anesthetized with isoflurane

Seok et al. PNAS | February 26, 2013 | vol. 110 | no. 9 | 3511 and exsanguinated by cardiac puncture at 2 h, 1 d, 3 d, or 7 d after injury. Mice were assayed on Affymetrix mouse MOE430 arrays. Among 5,554 genes receiving LPS, 10 ng E. coli 0113, the same endotoxin applied to the human significantly changed in human conditions, 4,918 genes had mouse study, in 200 μL PBS or saline control by i.p. injection were not anesthetized. orthologs assayed on the mouse arrays, which were included in the The amount of LPS was chosen to generate similar cytokine induction pat- subsequent analysis. terns as seen in a human LPS study (17). Comparison of Gene Response Between Datasets. The maximum fold changes Gene Expression Profiling. Total blood leukocytes were isolated according to of gene expression were measured in log scale between patients and healthy protocols published previously. Total cellular RNA was extracted and hy- subjects for each dataset of human burn, trauma, and endotoxemia and be- bridized onto an Affymetrix HU133 Plus 2.0 GeneChip according to the tween treated and control groups for the corresponding murine models. Be- ’ manufacturer s recommendations. Gene expression data have been de- tween two datasets, the agreements of the maximum gene fold changes (R2)as posited in GEO under the accession numbers listed in Table 1. well as directionality of the changes (%) were compared. R2 represents the square of Pearson correlation. Similar results were seen when the rank corre- Analysis of Time-Course Gene Expression Data. The trajectory of longitudinal lation was calculated as in SI Appendix, Fig. S1. Percent represents the per- expression of each gene was obtained by a cubic spline function, and the centages of genes changed to the same direction between the two datasets. mean expression of controls was considered as the base line expression. The Representative patient and corresponding mouse model studies were maximum deviation of the trajectory line from the base line was referred to as identified from GEO for several additional severe acute inflammatory diseases the maximum fold change or simply, the fold change of the gene between (sepsis, ARDS, and infections) (Fig. 4 and Table 1). The fold change of each patients and healthy controls. For each gene, the time required for the gene gene measured was calculated between patients and controls in a human to change to one-half of its maximum fold change was defined as its response time, and the time required for the gene to return back to one-half after it study or between treated and control groups in a murine model study, and for reached maximum point was defined as its recovery time. The significance a time-course dataset (GSE20524), the maximum fold change was calculated. of the longitudinal gene expression change was estimated using EDGE (Ex- The gene response in each dataset was then compared with the 5,554 genes fi traction of Differential Gene Expression) by 1,000 random permutations. signi cantly changed in human trauma, burns, and endotoxemia. Significant genes were selected by FDR < 0.001 and fold change ≥ 2 for each dataset; 5,544 genes were identified as significant between patients and ACKNOWLEDGMENTS. The datasets reported in the paper are available on healthy subjects (4,389 genes in trauma, 3,250 genes in burns, and 2,251 GEO with accession numbers listed in SI Appendix. This work was supported genes in endotoxemia). by Defense Advanced Research Projects Agency Grant W9111NF-10-1-0271 (to H.S.W.), National Institute on Disability and Rehabilitation Research Grant H133A070026 (to D.N.H.), National Institute of General Medical Sci- – Human Mouse Orthologs. Entrez genes (20,273) were assayed on Affymetrix ences Grants 5P50GM060338 (to D.N.H.), 5R01GM056687 (to D.N.H.), human HU133 plus v2 arrays, among which 16,646 genes had mouse R24GM102656 (to W. Xiao), and 5U54GM062119 (to R.G.T.), and National orthologs according to the Mouse Genome Database (17), and 15,686 genes Human Genome Research Institute Grant P01HG000205 (to R.W.D.).

1. US Department of Health and Human Services Food and Drug Administration, Center for 18. Lederer JA, et al. (2008) Comparison of longitudinal leukocyte gene expression after Drug Evaluation and Research (2006) Guidance for Industry: Nonclinical Safety Evaluation burn injury or trauma-hemorrhage in mice. Physiol Genomics 32(3):299–310. of Drug or Biologic Combinations. Available at http://www.fda.gov/OHRMS/DOCKETS/ 19. Edgar R, Domrachev M, Lash AE (2002) Gene Expression Omnibus: NCBI gene ex- 98fr/05d-0004-gdl0002.pdf. pression and hybridization array data repository. Nucleic Acids Res 30(1):207–210. 2. European Commission (2010) Sixth Report from the Commission to the Council and 20. Dressman HK, et al. (2007) Gene expression signatures that predict radiation exposure the European Parliament on the Statistics on the Number of Animals Used for Ex- in mice and humans. PLoS Med 4(4):e106. perimental and Other Scientific Purposes in the Member States of the European 21. Cuschieri J, et al. (2012) Benchmarking outcomes in the critically injured trauma pa- Union COM2010. Available at http://ec.europa.eu/environment/chemicals/lab_animals/ tient and the effect of implementing standard operating procedures. Ann Surg 255 pdf/com_2010_511.pdf. (5):993–999. fl 3. Woodcock J, Woosley R (2008) The FDA critical path initiative and its in uence on new 22. Bone RC (1991) Let’s agree on terminology: Definitions of sepsis. Crit Care Med 19(7): – drug development. Annu Rev Med 59:1 12. 973–976. 4. Pound P, Ebrahim S, Sandercock P, Bracken MB, Roberts I; Reviewing Animal Trials 23. Rangel-Frausto MS, et al. (1995) The natural history of the systemic inflammatory Systematically (RATS) Group (2004) Where is the evidence that animal research ben- response syndrome (SIRS). A prospective study. JAMA 273(2):117–123. fi – e ts humans? BMJ 328(7438):514 517. 24. Mestas J, Hughes CC (2004) Of mice and not men: Differences between mouse and 5. Hackam DG, Redelmeier DA (2006) Translation of research evidence from animals to human immunology. J Immunol 172(5):2731–2738. humans. JAMA 296(14):1731–1732. 25. Cabrera O, et al. (2006) The unique cytoarchitecture of human pancreatic islets has 6. van der Worp HB, et al. (2010) Can animal models of disease reliably inform human implications for islet cell function. Proc Natl Acad Sci USA 103(7):2334–2339. studies? PLoS Med 7(3):e1000245. 26. Schneider DS, Ayres JS (2008) Two ways to survive infection: What resistance and 7. Rice J (2012) Animal models: Not close enough. Nature 484(7393):S9. tolerance can teach us about treating infectious diseases. Nat Rev Immunol 8(11): 8. Hotchkiss RS, Opal S (2010) Immunotherapy for sepsis—a new approach against an 889–895. ancient foe. N Engl J Med 363(1):87–89. 27. Råberg L, Sim D, Read AF (2007) Disentangling genetic variation for resistance and 9. Hotchkiss RS, Coopersmith CM, McDunn JE, Ferguson TA (2009) The sepsis seesaw: tolerance to infectious diseases in animals. Science 318(5851):812–814. Tilting toward immunosuppression. Nat Med 15(5):496–497. 28. Schaedler RW, Dubos RJ (1961) The susceptibility of mice to bacterial endotoxins. 10. Wiersinga WJ (2011) Current insights in sepsis: From pathogenesis to new treatment – targets. Curr Opin Crit Care 17(5):480–486. J Exp Med 113:559 570. 11. Mitka M (2011) Drug for severe sepsis is withdrawn from market, fails to reduce 29. Warren HS, et al. (2010) Resilience to bacterial infection: Difference between species – mortality. JAMA 306(22):2439–2440. could be due to proteins in serum. J Infect Dis 201(2):223 232. 12. Davis MM (2008) A prescription for human immunology. Immunity 29(6):835–838. 30. Sauter C, Wolfensberger C (1980) Interferon in human serum after injection of en- – 13. Hayday AC, Peakman M (2008) The habitual, diverse and surmountable obstacles to dotoxin. Lancet 2(8199):852 853. human immunology research. Nat Immunol 9(6):575–580. 31. Huh D, et al. (2010) Reconstituting organ-level lung functions on a chip. Science 328 14. Calvano SE, et al. (2005) A network-based analysis of systemic inflammation in hu- (5986):1662–1668. mans. Nature 437(7061):1032–1037. 32. 1000 Genomes Project Consortium, et al. (2010) A map of human genome variation 15. Xiao W, et al. (2011) A genomic storm in critically injured humans. J Exp Med 208(13): from population-scale sequencing. Nature 467(7319):1061–1073. 2581–2590. 33. Ng SB, et al. (2009) Targeted capture and massively parallel sequencing of 12 human 16. Brownstein BH, et al. (2006) Commonality and differences in leukocyte gene ex- exomes. Nature 461(7261):272–276. pression patterns among three models of inflammation and injury. Physiol Genomics 34. Dudley JT, et al. (2011) Computational repositioning of the anticonvulsant topiramate 24(3):298–309. for inflammatory bowel disease. Sci Transl Med 3(96):96ra76. 17. Copeland S, Warren HS, Lowry SF, Calvano SE, Remick D (2005) Acute inflammatory 35. Sirota M, et al. (2011) Discovery and preclinical validation of drug indications using response to endotoxin in mice and humans. Clin Diagn Lab Immunol 12(1):60–67. compendia of public gene expression data. Sci Transl Med 3(96):96ra77.

3512 | www.pnas.org/cgi/doi/10.1073/pnas.1222878110 Seok et al. Supplementary Methods

Analysis of time course gene expression data. The time course data of the expression level of a representative gene is shown in the below figure. The trajectory of longitudinal expression of each gene was obtained by a cubic spline function, and the mean expression of controls was considered as the base line expression. The maximum deviation of the trajectory line from the base line was referred to as the maximum fold change, or simply the fold change of the gene between patients and healthy controls. For each gene, the time required for the gene to change to the half of its maximum fold change was defined as its response time, and the time required for the gene to return back to the half after reached maximum point was defined as its recovery time. The significance of the longitudinal gene expression change was estimated using EDGE (1) by 1,000 random permutations. Significant genes were selected by FDR<0.001 and fold change ≥2 for each data sets. 5,544 genes were identified as significant between patients and healthy subjects (4,389 in trauma, 3,250 in burns, and 2,251 in endotoxemia).

Analysis of the longitudinal gene expression. Shown is the time course data of the measured expression level of a representative gene. On the x-axis, points of samples of controls are displayed at 0 hour, and points of patient samples are displayed according to their sampling time after injury. On the y-axis, each point shows the log2 fold change of the expression value of the gene in one sample compared with the mean expression of the controls. The longitudinal trend line (in green) was obtained by spline smoothing of the gene expression values. The maximum fold change is defined as the largest deviation of the trend line from the base level, i.e. the mean expression of the controls. The response time and recovery time are defined as the time when the trajectory reaches half of the maximum fold change before and after reaching the maximum respectively.

Human-mouse orthologs. 20,273 Entrez genes were assayed on Affymetrix human HU133 plus v2 arrays, among which 16,646 had mouse orthologs according to the Mouse Genome Database (2), and 15,686 were assayed on Affymetrix mouse MOE430 arrays. Among the 5,554 genes significantly changed in human conditions, 4,918 genes had mouse orthologs assayed on the mouse arrays, which were included in the subsequent analysis.

Comparison of gene response between datasets. The maximum fold changes of gene expression were measured in log-scale between patients and healthy subjects for each data set of human burn, trauma and endotoxemia, and between treated group and control group for the corresponding murine models. Between two datasets, the agreements of the maximum gene fold changes (R2) as well as directionality of the changes (%) were compared. R2 represents the square of Pearson’s correlation. Similar results were seen when the rank correlation was calculated as in Fig. S1. % represents the percentages of genes changed to the same direction between the two datasets. In Fig. S9, gene expression changes of patients subjected to irradiation were compared with those of mouse models (GSE10640) (3, 4). Gene expression data was obtained from GSE10640, which includes two independent sets of patients under irradiation as part of their pre- transplantation conditioning (n=24 and 10, each set was assayed on a different microarray platform) and 11 mouse models. 487 significantly changed genes (FDR<0.05) between patients under irradiation and controls were identified. R2 and % were calculated in these genes between the first data set of patients (n=24) and the second set of patients (n=10) and the mouse models.

Fig. S1. Rank correlations (R2) and percentages of genes changed to the same direction (%) between the human and mouse studies. The lower left half of the figure shows the scatter plots of the log2 fold changes of 4,918 human genes responsive to trauma, burns or endotoxemia (FDR<0.001; fold change >2), the same as shown in Fig. 1. The upper right half lists the rank correlations (R2) and percentages of genes changed to the same direction (%) between the corresponding studies. Note that by random chance 50% of the genes between two uncorrelated conditions are expected to change in the same direction. The results show that the genomic responses to human trauma and burns are highly correlated, while the murine models correlate poorly with human conditions and each other.

Fig. S2. Correlations of gene changes between the human and murine studies in subsets of genes significantly changed in human injury and showed a greater magnitude of fold changes (FC). Among 4,918 human genes significantly changed in human trauma, burns or endotoxemia (FDR<0.001; FC≥2), 336 genes had FC≥4, 171 had FC≥5, and 105 had FC≥6 in human burn. Shown are bar graphs of Pearson correlations (R2) between human trauma, endotoxemia, three murine models versus human burns as a reference. While the murine models correlate poorly among the 4,918 genes significantly changed in human injuries (FDR<0.001; FC≥2) with the corresponding human conditions (R2=0.00-0.09), the correlations increase to (R2=0.11-0.28) among the top 105 genes with the greatest magnitude of changes (FDR<0.001; FC≥6).

Fig. S3. Scatter plots and Pearson correlations (R2) of genes significantly changed in the murine models. In the mouse models, 1,519 genes were significant at FDR<0.001 and ≥2 fold changes between the treated and control groups in either one of the three murine models, among which 87% (1,318) have human orthologs. In contrast to the 4,918 genes responsive to one of the human conditions as shown in Fig. 1, here the scatter plots depict fold changes of these 1,318 genes. The axes represent log2 fold changes. The high correlation was maintained in these genes between human trauma and burn (R2=0.88), whereas correlation between models was low (R2≤0.22). A.

B.

Fig. S4. Robustness of the analysis with respect to the sample sizes. The two human injury studies included 167 trauma patients sampled over the first 28 days after injury, 244 burn patients over the first year after injury, and 37 healthy controls, while the human endotoxemia study included 4 healthy volunteers received bacterial endotoxin and 4 controls each measured at four time points. The mouse models of trauma, burns, and endotoxemia each included 16 mice as the treated group and 16 as the control group (four time points, and 4 treated and 4 controls at each time point). To address the potential confounding feature of different sample sizes between the studies, comparable numbers of humans and mice were compared by random resampling. Since significant genes in the mouse burn and trauma models were identified from the time course changes over four time points after injury, to simulate a similar environment in the human injury studies, we randomly selected four burn subjects at 3 days (0-3d), 7 days (3d-7d), 14 days (7d-14d), and 21 days (14d-21d) as well as four healthy controls at each time point, and then measured the significance of each gene with the same method applied to the mouse data. For human trauma, four trauma subjects were randomly selected at 1 day (0-1d), 4 days (2d-4d), 7 days (4d-7d), and 14 days (7d-14d) as well as the controls. The comparison was repeated by 1,000 times. (A) Box plots of the correlations (R2) between burn injury and the other human conditions and murine models. The maximum fold changes of the randomly selected burn patients were compared with those of randomly selected trauma patients as well as the maximum fold changes from human endotoxemia and murine models. The correlations remained high between human injuries (R2 =0.85). (B) The number of significantly changed genes detected in each human condition and murine model. For the human injuries, the time course of the randomly selected burn and trauma patients and controls were analyzed and the number of significant genes were identified (FDR<0.001 & FC≥2). The median of the results from the 1,000 repeats is shown for human burns and trauma. Human conditions cause much more profound genomic response (≥ 3x more genes) than the corresponding mouse models.

Fig. S5. Box plots of the response and recovery times of gene expression in human burns, trauma and endotoxemia as well as corresponding mouse models. The response time is the time when the gene fold change reaches half of the maximum change. The recovery time is the time when the gene fold change returns back to half after it reached the maximum. While most of the gene response (75% percentile) occurred within the first 12 hours in all of the human injuries (human burns and trauma) and models (human endotoxemia and the three mouse models), the recovery differed dramatically in the time scale. The genomic disturbance in the models recovered within hours to 4 days (75% percentile), but lasted for 1-6 months in trauma and burn patients.

Fig. S6. Centroid expression of each gene cluster in human as shown in Fig. 2A. Each panel describes the centroid expression of a gene cluster for the three human conditions. The longitudinal trend lines were obtained by spline smoothing. Clusters 1-3 depict the common early activation (35% of the genes in Fig. 2A), clusters 5-8 are for the common early suppression (60%), and cluster 4 presents the late activation specific to burn and trauma (5%). Most of the genes changed in the same direction between burns and trauma (97%) and between the two injuries and endotoxemia (88%). In addition, the gene response time was within hours to burns, trauma and endotoxemia, while gene recovery differed dramatically between hours for endotoxemia, one month for trauma, and six months for burns. A. B.

Fig. S7. Gene changes over time for the genes responsive in the murine models. 1,519 genes were differentially expressed between the treated and control groups (FDR< 0.001 and ≥2-fold change). (A) K-means clustering of genes over time. Red indicates increased and blue indicates decreased expression relative to the mean (white). (B) Centroid expression of each gene cluster in the murine models. Each panel describes the centroid expression of a gene cluster for the three murine models. The longitudinal trend lines were obtained by spline smoothing. There was great variability among the three murine models within each cluster for the time span, suggesting that mouse gene responses differed from one another. A. Toll-like receptor expression in human burns

Log2(2)

Log (1.5) 2 0 -Log (1.5) 2

Fold Change -Log (2) 2 0 0.001 0.01 1 FDR

B. Toll-like receptor expression in the murine burns model

C. Toll-like receptor expression in human trauma

D. Toll-like receptor expression in the murine trauma model

E. Toll-like receptor expression in human endotoxemia

F. Toll-like receptor expression in the murine endotoxemia model

Fig. S8. The changes of genes in the TLR pathways for human injuries and the mouse models. Genes with FDR<0.01 are colored, and the rest genes are in white. Genes with significant changes (FDR<0.001 and fold change ≥2) are colored in red if activated or blue if suppressed. Genes with mild changes (FDR<0.01 and fold change ≥1.5) are colored in light red or light blue. Other genes (FDR<0.01 but fold change<1.5) are colored in gray. Toll-like receptor expression in (A) human burns, (B) the murine burns model, (C) human trauma, (D) the murine trauma model, (E) human endotoxemia, and (F) the murine endotoxemia model.

Fig. S9. Comparison of gene expression changes by radiation exposure in mice and patients. Gene expression data was obtained from GSE10640 (3, 4), which includes two independent sets of patients under irradiation as part of their pre-transplantation conditioning (n=24 and 10) and 11 mouse models. 487 significantly changed genes (FDR<0.05) between patients under irradiation and controls were identified. Plotted are the correlations (R2, left, pink) of the fold changes of these genes and the percentages of genes changed in the same direction (%, right, green) between the first data set of patients (n=24) and the second set of patients (n=10) and the mouse models. Table S1. Reproducibility of human and mouse studies. Standard deviations (SD) and coefficients of variation (CoV) were calculated from the biological replicates at each time point and averaged over the time points for mouse experiments and human endotoxemia. For human trauma and burn patients, the SD and CoV were calculated from samples of each binned time as described in Fig. S4 and averaged over the bins.

Table S1.

SD CoV

Human Controls 0.2986 0.053

Human Burn 0.3606 0.060

Human Trauma 0.2474 0.042

Human Endotoxemia 0.1456 0.061

Mouse Controls 0.2389 0.058

Mouse Burn Model 0.2378 0.058

Mouse Trauma-Hemorrhage Model 0.3204 0.078

Mouse Endotoxemia Model 0.2090 0.051

Table S2. Pathway comparisons between human injuries and mouse models. Shown are Pearson correlations (R2) and percentages of genes changed to the same direction (%) for (A) the 20 most up-regulated and (B) the 20 most down-regulated pathways between the four model systems (human endotoxemia and the three murine models) versus human burn injury. Negative correlations are shown as negative numbers (-R2). Human trauma is shown as reference. In every pathway, human endotoxemia had much higher similarity to human injury than mouse models had.

)

) ) 2 ) 2 % %

)

) ) 2 ) 2 %

% ) ) 2

Table S2A %

( Endotoxemia use # Genes (R Human Trauma ( Trauma Human (R Endotoxemia Human ( Endotoxemia Human (R Burn Mouse ( Burn Mouse Mouse Trauma (R Mouse Trauma ( (R Endotoxemia Mouse Mo Fcγ Receptor-mediated Phagocytosis in Macrophages 46 0.96 100% 0.63 93% 0.04 63% -0.04 37% 0.16 72% and Monocytes IL-10 Signaling 33 0.95 100% 0.68 91% 0.49 79% 0.22 67% 0.43 79% Integrin Signaling 77 0.93 99% 0.47 90% 0.10 66% 0.01 51% 0.04 61% B Cell Receptor Signaling 56 0.96 100% 0.65 96% 0.22 70% 0.00 50% 0.03 59% Toll-like Receptor Signaling 25 0.97 100% 0.79 100% 0.06 68% 0.02 56% 0.28 80% Production of Nitric Oxide and Reactive Oxygen Species in 62 0.93 100% 0.60 92% 0.18 63% -0.02 45% 0.06 61% Macrophages Role of Pattern Recognition Receptors in Recognition of 34 0.93 100% 0.22 82% 0.04 65% 0.01 50% -0.02 50% Bacteria and Viruses TREM1 Signaling 26 0.88 96% 0.33 92% 0.06 69% -0.03 35% 0.02 54% GM-CSF Signaling 30 0.94 100% 0.76 97% 0.25 70% 0.02 60% 0.00 43% IL-6 Signaling 35 0.95 100% 0.68 100% 0.39 74% 0.22 71% 0.30 71% IL-15 Signaling 33 0.96 100% 0.74 97% 0.11 58% 0.00 45% 0.04 58% IL-8 Signaling 61 0.94 100% 0.48 87% 0.21 67% -0.01 41% 0.03 51% IL-1 Signaling 34 0.92 97% 0.74 100% 0.10 65% 0.03 59% 0.29 71% IL-3 Signaling 34 0.95 100% 0.76 97% 0.12 56% -0.02 44% 0.00 56% PI3K/AKT Signaling 42 0.94 100% 0.36 86% 0.27 62% 0.05 60% 0.04 69% p38 MAPK Signaling 37 0.94 100% 0.70 97% 0.34 73% 0.20 70% 0.29 65% PPARα/RXRα Activation 55 0.94 98% 0.55 95% 0.16 53% 0.06 60% 0.19 64% Leukocyte Extravasation 65 0.93 97% 0.67 95% 0.16 60% -0.01 40% 0.29 71% Signaling JAK/Stat Signaling 31 0.95 100% 0.73 97% 0.09 48% 0.00 52% 0.02 52% Role of PKR in Interferon Induction and Antiviral 19 0.95 100% 0.71 95% 0.24 68% 0.10 63% 0.00 53% Response Summary median 0.94 100% 0.67 95% 0.16 65% 0.01 51% 0.04 61%

minimum 0.88 96% 0.22 82% 0.04 48% -0.04 35% -0.02 43% maximum 0.97 100% 0.79 100% 0.49 79% 0.22 71% 0.43 80%

)

) 2 2

)

) 2 2

) Table S2B 2

# Genes (R Human Trauma (%) Human Trauma (R Endotoxemia Human (%) Endotoxemia Human (R Burn Mouse (%) Burn Mouse Mouse Trauma (R Mouse Trauma (%) (R Endotoxemia Mouse (%) Endotoxemia Mouse iCOS-iCOSL Signaling in T 53 0.94 100% 0.72 96% 0.20 68% 0.00 49% -0.02 49% Helper Cells CD28 Signaling in T Helper 57 0.94 100% 0.69 98% 0.18 68% 0.04 61% -0.02 47% Cells Calcium-induced T 30 0.95 100% 0.77 100% 0.26 70% 0.03 63% -0.11 40% Lymphocyte Apoptosis PKCθ Signaling in T 53 0.94 100% 0.68 96% 0.26 75% 0.05 60% 0.00 51% Lymphocytes T Cell Receptor Signaling 47 0.96 100% 0.74 98% 0.32 68% 0.00 51% -0.03 47% Role of NFAT in Regulation 72 0.94 100% 0.74 100% 0.22 74% 0.04 60% 0.00 49% of the Immune Response Primary Immunodeficiency 20 0.95 100% 0.38 80% 0.11 55% 0.01 60% -0.02 35% Signaling B Cell Development 12 0.90 100% 0.48 83% 0.14 67% 0.20 75% 0.01 42% CTLA4 Signaling in Cytotoxic 46 0.92 100% 0.67 85% 0.18 70% 0.04 54% 0.00 48% T Lymphocytes Antigen Presentation Pathway 18 0.75 89% 0.59 89% 0.11 67% 0.36 67% -0.04 50% EIF2 Signaling 67 0.92 97% 0.67 96% 0.05 61% 0.04 66% 0.00 52% IL-4 Signaling 35 0.91 97% 0.61 91% 0.15 60% 0.03 54% 0.00 54% NF-κB Activation by Viruses 30 0.96 100% 0.62 100% 0.16 63% -0.01 47% 0.00 57% Type I Diabetes Mellitus 48 0.93 100% 0.74 98% 0.24 67% 0.07 69% 0.00 46% Signaling Cytotoxic T Lymphocyte- mediated Apoptosis of Target 27 0.88 100% 0.79 96% 0.40 70% 0.07 67% -0.29 30% Cells T Helper Cell Differentiation 32 0.94 100% 0.80 100% 0.03 53% 0.03 66% -0.03 38% Purine Metabolism 73 0.93 100% 0.38 85% 0.03 56% 0.12 66% -0.08 33% Regulation of eIF4 and 52 0.94 96% 0.59 92% 0.12 58% 0.00 60% 0.04 60% p70S6K Signaling mTOR Signaling 61 0.92 97% 0.39 84% 0.07 59% 0.00 57% -0.02 52% OX40 Signaling Pathway 26 0.89 100% 0.75 100% 0.42 73% 0.30 77% -0.05 35% Summary median 0.93 100% 0.68 96% 0.17 67% 0.04 61% -0.02 48%

minimum 0.75 89% 0.38 80% 0.03 53% -0.01 47% -0.29 30% maximum 0.96 100% 0.80 100% 0.42 75% 0.36 77% 0.04 60%

Acknowledgements

List of additional participants and affiliations of the Inflammation and Host Response to Injury Large Scale Collaborative Research Program:

Amer Abouhamze, Dept. of Surgery, University of Florida College of Medicine, Gainesville, FL 32610. Ulysses G.J. Balis, Dept. of Pathology, University of Michigan Medical School, Ann Arbor, MI 48109 David G. Camp II, Pacific Northwest National Laboratory, Richland, WA 99352 Asit K. De, Dept. of Surgery, University of Rochester Medical Center, Rochester, NY 14642 Brian G. Harbrecht, Dept. of Surgery, University of Louisville, Louisville, KY 40292 Douglas L. Hayden, Biostatistics Center, Massachusetts General Hospital, Boston, MA 02114 Amit Kaushal, Stanford Genome Technology Center, Stanford University, Palo Alto, CA 94305 Grant E. O'Keefe, Dept. of Surgery, University of Washington, Seattle, WA 98104 Kenneth T. Kotz, Dept. of Surgery, Massachusetts General Hospital, Boston, MA 02114 Weijun Qian, Pacific Northwest National Laboratory, Richland, WA 99352 David A. Schoenfeld, Biostatistics Center, Massachusetts General Hospital, Boston, MA 02114 Michael B. Shapiro, Dept. of Surgery, Northwestern University, Chicago, IL 60611 Geoffrey M. Silver, Dept. of Surgery, Loyola University School of Medicine, Maywood, IL 60153 Richard D. Smith, Pacific Northwest National Laboratory, Richland, WA 99352 John D. Storey, Dept. of Molecular Biology, Princeton University, Princeton, NJ 08544 Robert Tibshirani, Dept. of Statistics, Stanford University, Stanford, CA 94305 Mehmet Toner, Dept. of Surgery, Massachusetts General Hospital, Boston, MA 02114 Julie Wilhelmy, Stanford Genome Technology Center, Stanford University, Palo Alto, CA 94305 Bram Wispelwey, Biostatistics Center, Massachusetts General Hospital, Boston, MA 02114 Wing H Wong, Dept. of Statistics, Stanford University, Stanford, CA 94305

Supplemental References:

1. Storey JD, et al. (2005) Significance analysis of time course microarray experiments. Proc Natl Acad Sci U S A., 102(36):12837-42. 2. Blake JA, et al. (2011) The Mouse Genome Database (MGD): premier model organism resource for mammalian genomics and genetics. Nucleic Acids Res., 39(Database issue):D842-8. 3. Dressman HK, et al. (2007) Gene expression signatures that predict radiation exposure in mice and humans. PLoS Med., 4(4):e106. 4. Meadows SK, et al. (2008) Gene expression signatures of radiation response are specific, durable and accurate in mice and humans. PLoS One, 3(5):e1912.