Frozen glioma specimens and a portion of the paraffin-embedded tissues were obtained from the Neurosurgery Tissue Bank at the University of California, San Francisco. These patients were on clinical protocols and follow-up data were available. Additional samples were from an ongoing research trial on subtotally resected GBM at University of Texas, M. D. Anderson Cancer Center, for which follow-up was also available. Postmortem specimens from normal brains were obtained from Stanford University Hospital. Each frozen tumor tissue was sectioned and histopathologically evaluated by a single neuropathologist (A.B.), and specimens containing less than approximately 25% tumor cells were not included for study. All samples were obtained with informed consent and their use was approved by the Committee on Human Research.

For the four tumors for which paired samples were acquired, patients underwent contrast-enhanced T1-weighted magnetic resonance imaging. Two regions of each tumor were marked by a neurosurgeon based upon radiographic appearances that suggested unique histopathological features (e.g. contrast-enhancing vs. non-contrast-enhancing). The selected regions were subsequently sampled under stereotactic guidance prior to removal of the remaining tumor.

The microarray methods followed closely those of previous studies (1, 2) (also see http://brownlab.stanford.edu). Briefly, total RNA was extracted using Trizol followed by mRNA purification using FastTrack (Invitrogen). Messenger RNA was reverse transcribed to cDNA and directly labeled with Cy dyes (Amersham Biosciences) before hybridization to the arrays. Arrays contained ~23,000 elements (representing ~18,000 different UniGene clusters) and were scanned using a 4000B GenePix scanner at 10 mm resolution (Axon Instruments Inc.). Equal amounts of the experimental and reference probes were pooled and hybridized to microarrays. The reference pool of mRNA was prepared from a combination of human cell lines that has been previously described (2). Comparison of all experimental samples to the same reference allowed the relative expression level of each gene to be compared across all of the experiments. The resulting images were processed using ScanAlyze (available here.) The data were then normalized and indexed in the Stanford Microarray Database (SMD).

For Figure 1A, array elements that were not contaminated with mitochondrial DNA (as annotated in SMD) were extracted if their Intensity/Background ratio was greater than 1.5 in either channel. (For Figure 1B, arrays from non-GBM tumors were removed at this point.) The array vectors were then centered to the median, followed by the gene vectors. Only elements with data on at least 80% of the arrays and which had at least 5 measurements greater than 2 fold from the gene median were included in subsequent analyses. Since the tumor samples were hybridized to arrays from three different production runs, we removed batch specific artifacts by only considering elements that had correlation coefficients between -0.5 and 0.5 when compared to standard vectors encoding array batch using 1s and 0s. Agglomerative hierarchical clustering was performed using the Cluster program (3). Stand-alone Perl scripts were written to facilitate a number of the additional analyses.
     We used a two step algorithm to identify survival-associated genes: First, a Cox regression coefficient was calculated for each well-measured gene and a moving average (window size: 71) of these values was calculated based on hierarchical cluster order. Only one sample was used from each tumor for this calculation. Using 1,000 random permutations of the array labels, the average Cox statistics above and below which clusters were significantly associated with survival at p<0.01 were identified. Specifically, for each permutation of the array labels, we calculated the Cox regression coefficient for every gene and performed a moving average analysis as above. For each round, the maximum and minimum moving average values were stored. The p value cut-offs were determined from the frequencies of the minimum and maximum averaged Cox statistics. Permutation analysis was done using the R statistical software package. In order to further limit the genes on which to focus follow-up studies, we used a second algorithm to identify "tumor intrinsic" genes that were expressed most consistently in geographically-distinct samples from the same tumor, but varied in samples from different tumors. The intrinsic score of a given gene is the ratio of the mean squared pair wise difference in that gene's transcript levels between multiple samples from the same tumor, to the mean squared pair wise difference in the gene's transcript levels between samples from different tumors. Kaplan-Meier survival analysis was carried out using WINSTAT plug-in software for Microsoft Excel (www.winstat.com).

Paraffin-embedded tissues were de-waxed and re-hydrated, followed by high-temperature antigen retrieval in 10 mM sodium citrate (pH 6) (4). Immunoreactivity was visualized by biotinylated secondary antibody and Vectastain Elite ABC kit (Vector). Rabbit polyclonal antibodies against FABP7 were generous gifts from Drs. Godbout and Heintz (5, 6). Scoring was semi-quantitative based on the extent and intensity of nuclear staining by a single neuropathologist (K.A.).

SF767MG cells were transfected with pcDNA3 or pcDNA3-FABP7 using FuGENE (Roche). Cells were incubated for 36 hours followed by selection in 1 mg/ml G418 for 10 days with media changes every 1-3 days depending on cell death. Cells were incubated in serum free media for 24 hours and then isolated using 5 mM EDTA in Ca²⁺/Mg²⁺-free PBS. For migration assays, cells were washed twice and resuspended in serum free media. Next, 1x10⁴ cells were seeded into the upper chamber of TransWell Fluoroblocks (8 mm pore size, Corning BD Biosciences) in 300 ml final volume. 700 ml of 10% FBS in DMEM were placed in the lower chamber to serve as chemoattractant. At each time point transwell inserts were removed and soaked in 250 ml of serum free media containing 5mM of CFSE for 15 minutes at 37ºC, then washed twice in 10% FBS in DMEM for 5 minutes each. Unmigrated cells were removed using cotton swabs and the numbers of migrated cells in 5 randomly chosen fields under 200X magnification were counted using a fluorescence inverted microscope. Each data point was obtained by combining 3 of the 5 cell numbers with the closest values. At each time point, the data consisted of 3 separate transfections with triplicates within each transfection.

1. Alizadeh, A. A., Eisen, M. B., Davis, R. E., Ma, C., Lossos, I. S., Rosenwald, A., Boldrick, J. C., Sabet, H., Tran, T., Yu, X., Powell, J. I., Yang, L., Marti, G. E., Moore, T., Hudson, J., Jr., Lu, L., Lewis, D. B., Tibshirani, R., Sherlock, G., Chan, W. C., Greiner, T. C., Weisenburger, D. D., Armitage, J. O., Warnke, R., Staudt, L. M. & et al. (2000) Nature 403, 503-11.
2. Perou, C. M., Sorlie, T., Eisen, M. B., van de Rijn, M., Jeffrey, S. S., Rees, C. A., Pollack, J. R., Ross, D. T., Johnsen, H., Akslen, L. A., Fluge, O., Pergamenschikov, A., Williams, C., Zhu, S. X., Lonning, P. E., Borresen-Dale, A. L., Brown, P. O. & Botstein, D. (2000) Nature 406, 747-52.
3. Eisen, M. B., Spellman, P. T., Brown, P. O. & Botstein, D. (1998) Proc Natl Acad Sci U S A 95, 14863-8.
4. Shi, S. R., Key, M. E. & Kalra, K. L. (1991) J Histochem Cytochem 39, 741-8.
5. Feng, L., Hatten, M. E. & Heintz, N. (1994) Neuron 12, 895-908.
6. Godbout, R., Bisgrove, D. A., Shkolny, D. & Day, R. S., 3rd (1998) Oncogene 16, 1955-62.