Using the cDNA microarrays spotted with duplicated 7,334 cDNA clone, we analyzed RNA specimens successfully amplified in 44 peripheral blood collected from 25 confirmed SARS patients (age ranged from 23 to 80 years old, mean = 41.8, SD = 17.2, median = 34), of whom 24 survived. Except for one patient who died on the 4th day, duration of hospitalization in this cohort ranged from 12 to 51 days (n = 24, mean = 24.5, SD = 10.1, median = 21) (Additional file 1). We defined 11 specimens as acute SARS (AS) using the following criteria: (i) the whole blood RNA from a hospitalized patient was PCR positive for SARS-CoV, or (ii) the specimen was collected within 10 days after the disease onset in patients whose blood was later diagnosed ELISA-positive for anti-SARS IgG. The rest of 33 RNA specimens from SARS patients were labelled as recovering SARS (RS). Our study included 11 normal control (NC) volunteers and 11 patients with bacterial infections (IN) as healthy and non-SARS infection controls, respectively (Additional files 2 and 3).

When we compared the gene expression profiles among acute SARS (AS), recovering SARS (RS), bacterial infection (IN), and normal control (NC) groups, we observed the variances of gene expression in both SARS (AS, AS+RS) and bacterial (IN) groups to be equally higher than that in healthy controls (NC) (Fig. 1a). This result indicates that gene expression profiles of either SARS or bacterial groups differed significantly from that of normal controls.

A probe set of 885 genes with standard deviations greater than 0.5 across 66 arrays was selected for further analyses. An average linkage hierarchical clustering tree with Pearson correlation proximity was built on the 33 arrays (11 NC, 11 IN, and 11AS) using these 885 genes (Fig. 1b). The AS and NC groups were well separated into two opposite coherent clusters. Singular value decomposition (SVD) analysis, a dimension reduction method to project gene expression profiles to fewer representative eigenvectors 23, also successfully separated AS, IN, and NC specimens into three clusters with first two eigenvectors (Fig. 1c). Interestingly, the recovering SARS (RS) samples are interspersed among the AS and IN samples.

To identify which genes were specifically regulated by SARS-CoV, we performed two sets of two-sample Student t-test for means with an unequal-variance assumption. In the first set, we contrasted 11 AS versus 22 non-AS (NC and IN) specimens on all 885 genes. The genes with significant testing results were considered to be specifically induced by SARS-CoV (Fig. 2a,b). For the second set of t-tests, we compared 11 NC with 22 non-NC (IN and AS) specimens. We considered that the change in significant genes identified by the second t-test was induced by both bacterial and viral infections (Fig. 2c,d). Genes identified from these two sets of test were then ranked separately according to the corresponding sets of P-values. Gene expression profiles for the top 20 and the bottom 20 genes from both sets are displayed as Figure 2.

Unexpectedly, most of the genes specifically upreguated by SARS-CoV are ESTs (13/20 genes) that were not annotated previously (Fig. 2a). On the other hand, SARS-CoV stimulated the host innate immunity by upregulating genes including defensins 2425 and eosinophil derived neurotoxin 2627, similar to that of bacterial infections (Fig. 2c).

A simple k-nearest-neighbour method was used to obtain a near optimal number of 30 genes from the 885 filtered gene set for discriminating specimens between acute SARS (AS) and non-SARS (NC and IN) (Additional file 4). The selected top 30 upregulated (P < 6 × 10^-6) and the top 30 downregulated genes (P < 4 × 10^-7) from the AS versus non-AS (IN and NC) Student's t-test were used as the specific probe set to assess the status of SARS infection. Eight genes that were also significant in the NC versus non-NC (AS and IN) t-test were excluded, resulting in a specific AS probe set of 52 genes. For the GAP analysis, we calculated pair-wise Euclidean distances among 55 samples (11 AS, 33 RS, 11 NC) using these 52 genes, aiming to identify a one-dimensional order that could reflect the severity structure of the disease (Fig. 3a). Using this GAP elliptical arrangement of 55 specimens (columns), we observed a transition of gene expression patterns of 52 genes (rows) from the left side where NC clustered to the right side where AS accumulated (Fig. 3b). Hierarchical clustering trees guided by self-organized map (SOM) and other clustering methods were also performed to sort the SARS patient samples using the same 52 genes. The Robinson criterion 222829 is often employed to assess the performances of different seriation algorithms. Table 1 (and Additional file 5) shows that the GAP algorithm derived a smoother transition pattern than other methods in the Robinson sense. Thereby, we have derived the SARS severity rank according to the expression profile of 52 signature genes as a whole in each patient, as demonstrated by the smooth transition of expression levels in each (row) of these genes from NC to RS to AS (Fig. 3b).

For validation purposes, we further tested the stability of the rank (order) derived from GAP analysis on the 52 genes for the 55 specimens. The same GAP procedure was repeatedly applied to the top 20 to 200 genes (among the filtered 885 genes) with significant p-values (Student's t-test) between the AS versus non-AS (IN and NC). While the ranks for the 55 specimens obtained from the most significant 20 to 200 genes are highly correlated to each other, they are significantly different from the ranks derived from the 52-gene sets that were randomly selected from the 885 genes (data not shown).

We scrutinized the clinical courses of patients who donated the 10 RS specimens that were scattered among AS (Fig. 3a) and found evidence of underlying severity of the disease in the majority of patients. For example, sample RS43 from a patient who had been discharged from hospital for 2 weeks was still PCR-positive for SARS-CoV; RS54, a PCR-positive sample was not grouped as AS because of the negative ELISA result. RS38, RS40, and RS42 still represented acute SARS infections because they were collected only 1, 2, and 3 days after AS37, AS39, and AS41, respectively. Patients with RS78 and RS91 who had severe SARS courses were hospitalized for 41 and 51 days, respectively. The patient for RS8 was in the second week of disease. The only two unexplained specimens, RS18 and RS71 from the same patient, may represent a unique biological variability, accounting for the misclassification using this 52-gene molecular signature.

To test the efficacy of using these 52 genes as the molecular signature for the severity of SARS patients, we identified a significant correlation (P < 1 × 10^-6) between the derived rank of SARS severity and the number of days after the onset of disease (Fig. 4a). We further used this rank of SARS severity to examine the recovery trend in 17 recovering patients who had donated multiple specimens (Fig. 4b). Except for the one patient (5.3 % = 1/19, shown as the red line in Fig. 4b), similar trends existed in 18 out of 19 lines (94.7 %). Pugin et al. combined body temperature, white blood cell count, volume and appearance of tracheal secretions, oxygenation, chest X-ray, and tracheal aspirate cultures into a clinical pulmonary infection score (CPIS) as a diagnostic tool for pneumonia 30. We observed that the rank of SARS severity was also significantly (P < 0.001) correlated with the CPIS (Fig. 4c). Collectively, these results demonstrate a correlation between the molecular severity rank and clinical factors, suggesting the usefulness of the molecular signature as a genome-wide parameter for gauging the severity of SARS patients.

Blood specimens of 25 SARS patients (Additional file 1) were collected from 10 May to 4 July 2003 at Kaohsiung Medical Center of Chang Gung Memorial Hospital (CGMH) in Kaohsiung City of southern Taiwan. Two additional blood samples (RS94 and RS97) were collected from apparently healthy individuals who had recovered from SARS infection 3 months later. Diagnosis of SARS was based on the guidelines of World Health Organization (WHO) 34. More comprehensive data of the SARS cohort were previously published 33. This study was approved by the Institute Review Board of CGMH. Total RNA was isolated with the PAXgene Blood RNA System (Qiagen, USA) and stored at -80°C. After RNAs were further purified and concentrated into 15 μl BR5 solution with RNeasy MinElute kit (Qiagen, USA), 2 μl were used for linear RNA amplification using RiboAmp RNA Amplification Kit (Arcturus, California USA). Before the first Strand Nuclease Mix was added to the RNA samples, all of the RNA purification and amplification were performed inside a Biosafety Level 3 laboratory located in Lin-Kou Medical Center of CGMH. We analyzed the quality and quantity of amplified RNA with Bioanalyzer 2100 (Agilent, CA, USA).

The antigen used for the SARS detection ELISA was the detergent-extracted and gamma irradiated Vero E6 cells infected with SARS-CoV. Identical preparations from uninfected Vero E6 cells were used as the control. Patients' sera were 1:10 diluted and added to the ELISA plates, and goat anti-human IgG antibody conjugated with horseradish peroxidase (DAKO, Cambridgeshire, UK) was added for enzymatic reaction. After adding the substrate, O-phenylenediamine, the optical density (O.D.) was measured at 450 nm wavelength. The cutoff value of O.D. for SARS-CoV IgG ELISA was 0.15. Sensitivity of this method was 100% (28/28 in confirmed SARS cases) and specificity was 98.4% (790/803) in the healthy control group.

Real-time quantitative PCR analysis for SARS-CoV was performed with Cor-p-F4, Cor-p-R4 and Cor-probe developed by CDC (GA, USA) with HT 7900 Sequence Detection System (Applied Biosystems, CA, USA).

In this study, we used the GMRCL Human 7K set, Version 2 chips as previously described 35. Twelve amplified RNA samples from healthy donors (Additional file 3) were pooled as the common reference for every array in this study. A total of 66 aRNA samples including 11 acute SARS (AS), 33 recovering SARS (RS), 11 non-SARS infection (IN), and 11 normal controls (NC) were analyzed with cDNA microarrays as tests against the pooled aRNA (the common reference). Among 66 aRNA preparations, 28 were analyzed with the dye-swapping microarray design. We averaged the log ratios of the duplicated spots on each slide. In the dye swapping experiments, we further averaged the log ratios derived from two slides. We used 400 ng of aRNA for labeling and hybridization using a 3DNA Array 350RP Detection kit (Genisphere, PA, USA), and scanned slides with a confocal scanner ChipReader (Virtek, Canada). We acquired the spot and background intensities with GenePix Pro 4.1 software (Axon Instruments, Inc., CA, USA), and carried out within-slide normalization using programs written with MATLAB 6.0 software (The MathWorks, Inc., MA, USA). To assure the reproducibility of our microarray system, we got the similar gene expression profiles from replicated samples (RS88) using the hierarchical clustering analysis and also got the highly correlated results (r² = 0.84) from two specimens (AS37 and RS38) that were collected from the same patient at a time interval of only one day. We consistently obtained identical results in each of 28 pairs dye-swapping experiments. The complete microarray data is available in Additional file 7.

We performed hierarchical clustering using Cluster and TreeView software 36 with the following parameters: (i) a standard deviation > 0.5 as the filtering cutoff point (885 genes with marked changes selected among 66 arrays), (ii) mean-centered genes and normalized genes, (iii) cluster analysis carried out with uncentered correlation of arrays. We also performed a singular value decomposition (SVD) 23 analysis of the correlation matrix for all 66 samples. The first two eigenvectors weighted by the corresponding singular values (eigenvalues) of the 66 samples were plotted against each other.

Robinson criterion 2228 is frequently used to assess the performances of sorting algorithms on symmetric proximity matrices. A Robinson Matrix, R = [r_ij], is a symmetric matrix such that r_ij ≤ r_ik if j<k<i and r_ij ≥ r_ik if i<j<k. The GAP elliptical seriation 22 utilizing the ellipse structure from a singular value decomposition of a converged correlation coefficient matrix usually identifies permuted matrix with a near Robinson form. A brief review on GAP and some details of its applications are available 37.