Expression profiling of blood samples from an SU5416 Phase III metastatic colorectal cancer clinical trial: a novel strategy for biomarker identification

Study population

Patient samples were derived from 2 randomized, open-label, multicenter Phase III clinical trials comparing standard of care chemotherapy alone or combined with SU5416 in patients with metastatic colorectal cancer. In both trials SU5416 was delivered twice weekly at a dose of 145 mg/m² via I.V. infusion. In the first trial (designated Trial A), the standard of care chemotherapy consisted of weekly administration of 5-FU and leucovorin (Roswell Park regimen); in the second trial (designated Trial B), the standard of care chemotherapy consisted of weekly or bi-weekly administration of 5-FU, leucovorin, and Irinotecan (CPT-11). A total of 23 pairs of patient samples were included in Affymetrix microarray expression profiling analysis (2 females and 9 males in the SU5416 treatment arm, and 2 females and 10 males in the control arm). The median patient age was 66 and 65 years for the SU5416 treatment arm and control arm, respectively. For RT-PCR verification experiments, samples from 12 females and 24 males from the SU5416 treatment arm, and 14 females and 17 males from the control arm were used. The median age for these patients was 62 and 60 years, respectively. Clinical response criteria were defined according to RECIST guidelines; briefly, complete response (CR) is defined as complete disappearance of all measurable and evaluable clinical evidence of cancer; partial response (PR) is defined as at least a 50% reduction in the size of all measurable tumor areas, progressive disease (PD) is defined as an increase of ≥25% (compared to baseline or best response) in the size of all measurable tumor areas; and stable disease (SD) is defined as neither sufficient shrinkage to qualify for PR nor sufficient increase to qualify for PD

Patient samples

All clinical samples for biomarker analysis were harvested and handled in accordance with full Institutional Review Board-approved protocol, and study participants had signed the study informed consent prior to any study-related procedures. All blood samples were collected into Vacutainer tubes containing sodium heparin. Ten ml of blood was withdrawn from patients prior to receiving any treatment on day 1 and also prior to dosing at end of cycle 1 (day 56 in Trial A, day 42 in Trial B). For PBMC preparations, blood samples were shipped overnight at ambient temperature to a central processing facility (Quest Diagnostics, Inc., Collegeville, PA, USA) for PBMC isolation via Ficoll gradient method [15]. Purified PBMCs were shipped in RNA lysis buffer (Clontech, Palo Alto, CA, USA) to SUGEN where isolation of total RNA was performed. For Trial B, whole peripheral blood samples were directly frozen at the clinical sites and shipped on dry ice to SUGEN for RNA isolation.

RNA sample processing

Total RNA was purified from PBMC samples using Clontech Nucleospin RNA II kit reagents (Clontech, Palo Alto, CA) and from whole blood samples using MRC TRI Reagent BD (Molecular Research Center, Cincinnati, OH, USA), an adaptation of the Chomczynski single-step method [16], according to the manufacturer's instructions. All sample preparations included a treatment with RNAse-free Dnase. RNA yields were measured by UV absorbance and RNA quality was assessed by agarose gel electrophoresis with ethidium bromide staining for visualization of ribosomal RNA band integrity.

Affymetrix high-density oligonucleotide microarray analysis of PBMC expression profiles

In general, the standard RNA processing and hybridization protocols as recommended by Affymetrix (Santa Clara, CA, USA) were followed in this study; these protocols are available in the Genechip^® Expression Analysis Technical Manual. Yields of total RNA for PBMC samples were generally low and for the majority of patients it was not possible to use the standard amount of total RNA (≥ 5 μg) as recommended in the standard protocol. Therefore a double linear amplification approach [17] was used in the generation of cRNA for hybridization. In these experiments, equal amounts of starting material were used for pre- and post-treatment samples from each donor (typically 2 μg). Briefly, the protocol was as follows: double-stranded cDNA was synthesized from total RNA, with Invitrogen Life Technologies SuperScript Choice system reagents (Invitrogen, Carlsbad, CA, USA). The T7-(dT)₂₄ oligomer was used for priming first-strand cDNA synthesis. Double-stranded cDNA product was purified via phenol-chloroform extraction, and then used as template in first round of in vitro transcription (IVT) of cRNA. The IVT reaction was performed with BioArray HighYield RNA Transcript Labeling Kit (Affymetrix) according to manufacturer's protocol, but with substitution of non-biotinylated ribonucleotides for biotinylated ribonucleotides. The cRNA product was then purified with Qiagen spin column clean-up protocol and used as template in second round of cDNA synthesis. This second round of cDNA synthesis was similar to the first round except that random hexamers were used in priming of first-strand synthesis, with T7-(dT)₂₄ oligomer priming the second-strand. The second round of IVT of cRNA was as in the first round but with biotinylated ribonucleotides rather than non-biotinylated ribonucleotides. Purified cRNA was quantitated, chemically fragmented according to Affymetrix protocol, and then hybridized overnight on Human Genome U95A Arrays (which contain probe sets for the detection of approximately 12,600 transcripts). Hybridized arrays were washed and stained with phycoerythrin-conjugated streptavidin detection chemistry in an Affymetrix Fluidics station. Images were scanned with a Hewlett-Packard GeneArray scanner.

Data Analysis

Data files were generated from scanned array images in the Affymetrix Microarray Suite Version 4.0 program. In this program, Average Difference (AD) values serve as relative indicators of the expression level of transcripts represented on the arrays. Average Difference determination relies on difference between background-subtracted signal from perfect match (PM) oligos and corresponding mismatch control (MM) oligos within a probe set representing a given transcript. To enable comparison of all hybridization data, global scaling was applied by multiplying the output of each experiment by a Scaling factor (SF) to make its average intensity equal to a user-defined Target Intensity (which was set at 1500 for these experiments). For comparisons between time points from a single patient, batch files were generated with Microarray Suite. These files contain calculated fold change (FC) values, which represent differential expression between day 56 compared to baseline, and also Difference Calls (DC), which represent a more conservative estimate of differential expression, with qualitative scores assigned to each transcript measurement according to the following system: Increased (I), Marginally Increased (MI), No Change (NC), Marginally Decreased (MD), and Decreased (D).

Subsequent data analysis was performed primarily with Spotfire DecisionSite for Functional Genomics software (version7.0) package and its Array Explorer component (Spotfire, Somerville, MA). Hierarchical clustering analysis and statistical comparisons were included in this step. Further refinement of the data, including filtering by Difference Call scores, was done with the Microsoft Access 97 database analysis program.

SYBR Green quantitative RT-PCR verification of array results

Primers were designed with Primer Express 1.5 software (Applied Biosystems). In all cases, primers were designed to bind within the sequence represented by Affymetrix probe sets (target sequence information available at http://www.affymetrix.com/analysis/index.affx). Total RNA samples (1 μg) were reverse transcribed to yield first-strand cDNA using the Applied Biosystems Reverse Transcription Reagents protocol (Applied Biosystems, Foster City, CA, USA). The reverse transcription reactions were then diluted 1:5 in distilled H₂O. SYBR Green PCR reactions were performed in 96-well optical plates and run in an ABI PRISM^® 7700 Sequence Detection System (SDS) machine. For individual reactions, 10 μl of each sample were combined with 15 μl of SYBR Green PCR Master Mix (Applied Biosystems) containing the appropriate primer pair at 350 nM. Data were extracted and amplification plots generated with ABI SDS software. All amplifications were done in duplicate and threshold cycle (C_t) scores were averaged for subsequent calculations of relative expression values. The C_t scores represent the cycle number at which fluorescence signal (ΔR_n) crosses an arbitrary (user-defined) threshold. Heat dissociation curve analysis was performed after each SYBR Green run as a test of whether a single product had been generated in each PCR reaction; multiple peaks in the dissociation curves are indicative of multiple PCR products and thus reduced specificity and sensitivity.

Quantitation and statistical analysis of SYBR Green RT-PCR data

The C_t scores for genes of interest for each sample were normalized against C_t scores for the corresponding endogenous control gene, which was the β-glucuronidase (GUS) gene. Relative expression for day 56 compared to day 1 was determined by the following calculation, as described in the Applied Biosytems users bulletin on Relative Quantitation of Gene Expression and in [18]:

Rel Exp = 2^-ΔΔCt,

Where ΔΔ C_t = (Ct _Target - Ct _GUS)_{day 56} - (Ct _Target - Ct _GUS)_{day 1.}

The relative expression data for a selected subset of potential biomarkers were tested for differences between the SU5416 (treatment) and the standard of care (control) arms. The Mann-Whitney U Test with a critical alpha level of 0.05 was used for statistical significance. Individual genes observed to be significantly different by Affymetrix analysis and in both sets of SYBR Green RT-PCR experiments were screened as potential biomarker candidates. This subset of potential biomarker candidates was tested subsequently for utility as class predictors to discriminate between the SU5416 and the standard of care arms. Discriminant analysis [19], a multivariate statistical technique, was used for this purpose. The genes were tested individually, using all possible combinations, and by reducing dimensions (Principal Component Analysis) in order to determine the subset of genes (predictor variables) that yielded highest classification accuracy. Cross-validation was used to test the robustness of classification accuracy. Results from three different cross-validations were evaluated to select the best set of predictor biomarkers: (1) jackknife method (dropping one case at a time), (2) randomly splitting the pooled data into two halves, prediction (for building model) and validation (for testing model) sets, and (3) using one trial as prediction and the second trial as validation sets, respectively. All statistical analyses were carried out after natural-log transformation of the data; SYSTAT 9.01 (SPSS, Inc., Chicago, IL, USA) software was used in statistical analysis.

Affymetrix expression profiling of pre- and post-treatment matched PBMC samples

Total RNA was isolated from PBMC prepared from patient's blood samples taken before (pre-dose day 1) and after (pre-dose day 56) administration of SU5416 or corresponding control regimen in a Phase III trial, designated as Trial A. Due to typically low RNA yields, many sample pairs were of insufficient quality for further use (failure rate of at least 1 of 8 samples, or 25% of sample pairs). Only sample pairs in which both day 1 and day 56 samples RNA yields were of 1 μg or greater were used in expression profiling analysis. Samples were hybridized to U95A high-density oligonucleotide arrays. A total of 11 sample pairs from the SU5416 treatment arm and 12 from the control arm were analyzed in the primary dataset. The change in expression (ratio of day 56 measurements to day 1 for each patient as calculated by Affymetrix software) was defined for each transcript; these are referred to as fold change (FC) values. The FC values for the 23 cases were analyzed with Spotfire Decision Site software tools, to compare patients in the SU5416 arm and control arm.

A t-test analysis was used to identify transcripts that were statistically significantly different between the two treatment arms. Over 100 genes with p-values less than 0.02 were identified; however, because there are over 12,000 transcripts measured on the U95A arrays, some of the genes would potentially be identified by chance. To further refine this subset of genes, queries based on Difference Call (DC) status were performed; Difference Calls offer a more stringent but non-numerical measure of differential expression, and are derived from a decision matrix that weighs comparison results from four metrics used in the Affymetrix analysis platform. The data were filtered to identify genes that were 'Increased' (I) or 'Decreased' (D) in a majority of the SU5416 arm cases but not in the control arm. A group of 13 genes that frequently showed increased expression was identified. Figure 1 displays a schema of the DC scores assigned to each gene for each patient sample pair; all cases from the SU5416 arm show induction in at least 6 of the 13 genes. Table 1 lists the number of cases in each arm in which an 'Increased' call was assigned and includes a brief description of putative functions of the gene products. The average fold change of all of these transcripts was higher in the SU5416 arm (the lowest average fold change was 2.6 for hypothetical protein FLJ13052, the highest was 33 for lactoferrin); the range of fold changes was also broader in this category, presumably reflecting variability among patients.

Quantitative RT-PCR validation of differentially expressed transcripts

SYBR Green RT-PCR was used to validate the microarray expression profiling data. SYBR Green is a dye that fluoresces when bound to double-stranded DNA, thus signal is directly proportional to the amount of product formed during PCR amplification [18, 23]. This method allows rapid and inexpensive comparison of gene expression across a large number of samples. The qRT-PCR validation was performed with a total of 31 sample pairs, 8 of which had previously been analyzed on Affymetrix U95A arrays and thus allowed a direct comparison of the correlationbetween the 2 transcript profiling methods. Data for each gene was normalized to expression of a housekeeping gene, β-glucuronidase (GUS). Direct comparison of SYBR Green RT-PCR results and Affymetrix results from the same RNA samples (n = 8 day 1-day 56 pairs) showed an overall qualitative agreement (i.e., same trend of induction or no change detected by both methods for each target) of 73%, as 35 of 48 comparisons were concordant. This number may be an underestimate since results for one patient were inconsistent for all 6 transcripts.

Figure 2 summarizes the results from the RT-PCR validation and compares them to those from Affymetrix analysis. Mann-Whitney U test comparison of SU5416 and control results from both analyses indicates that three of the six genes displayed statistical significance (p-values less than 0.05) based on the SYBR Green RT-PCR data (Table 3); these are CD24, lactoferrin, and LCN2. MMP-9 exhibited a p-value that was close to the significance cutoff and thus was also selected for further analysis. Defensin α3 and TSP-1 were not pursued further. Thus, four of the six initial transcripts selected from Affymetrix expression profiling were confirmed as differentially expressed using a PCR-based approach, in a larger sample set from the same clinical trial.

Quantitative RT-PCR validation of differentially expressed transcripts with samples from a second Phase III SU5416 trial

To further confirm these transcripts as potential biomarkers of SU5416 administration, SYBR Green RT-PCR analysis was carried out in samples from a second Phase III trial using SU5416 (Trial B). This trial was also a randomized metastatic colorectal cancer study, comparing a new chemotherapy standard of care, 5-FU/leucovorin/CPT [24, 25] to standard of care plus SU5416. RNA was isolated from patients' peripheral blood samples (rather than PBMC) harvested at the beginning (pre-dose day 1) and end (day 42) of cycle 1. In order to test if similar gene expression changes occurred, analysis was performed on 36 sample pairs, 18 from SU5416 treatment arm and 18 from control arm.

Figure 3 summarizes the overall frequency of induction of 2-fold or greater in each arm of the trial. It is clear that these transcripts are more frequently induced at day 42 of treatment in the SU5416 arm than in the control arm. This is supported by statistical analysis, as indicated in results of the Mann-Whitney U Test (Table 4). A visual representation of hierarchical clustering analysis of the qRT-PCR relative expression values from both trials for the each of the transcripts is displayed in Figure 4. This clustering pattern displays the distinction between the SU5416 and control arms based on relative expression data, and also indicates further distinctions among subsets of patients as well as the degree of overlap between trial arms in the clustering pattern. The extent of similarity between the relative expression patterns for each transcript (represented in columns) is also indicated; the pattern of MMP-9 is distinct from the others as it appears in a separate branch in the dendrogram structure. A table containing the raw relative expression measurements included in this dataset can be viewed in Additional file 1.

Discriminant analysis of the classification power of candidate biomarkers

An important next step was to test whether relative expression data from these samples could be used in a predictive fashion to classify samples to the appropriate trial arm. In order to assess this, discriminant analysis [19] of the RT-PCR data was performed. The relative expression values from both trials were combined into a single dataset and then natural log-transformed to reduce the scale of the values, making control and treated arms more comparable. When the cases were pooled (n = 67) and subjected to classification prediction, the overall accuracy of assignment to the appropriate trial arm was 84% when lactoferrin, CD24, and LCN2 were used as the predictor gene set (inclusion of MMP-9 slightly reduced the accuracy of cross-validation). Further cross-validation was performed by the jackknife method (which does a series of predictions, randomly removing 1 case from the total each time), and by splitting the data set into 2 random halves (one a 'training' set and the other a 'testing' set).