Gene expression microarray technology (MAT) is a rapidly developing analytical tool in basic and clinical research. The number of publications using microarrays (MAs) has been growing exponentially over the past four to five years (1, 2). In the 1990s, MAT was considered to be one of the most promising new tools for the science of the 21st century, when fast, high-throughput technologies would be used to grasp the complexity of biological systems (3–5). Since MAT is potentially capable of looking at all of the cellular processes at the mRNA level at a given moment, it was expected to deliver not only a comprehensive quantity of data about the transcriptional level, but also to shed light on novel processes, pathways, and molecular interactions in the living cell. MAT data, providing “freeze-frame” views of the transcriptome, could help us understand the role and weight of known, and yet to be discovered, molecular mechanisms in the “big picture” (6–8).
In clinical research, MAT is expected to help us better understand the molecular basis of diseases, and the difference between healthy and diseased cells, tissues, and organs, giving us clues about potential treatment (9–11). At the same time, microarrays (MAs) can be used to monitor the effects of these treatments, especially of new (and old) drugs (12–14). The unique view that MAT offers of the cells and tissues also generated high hopes in the field of pathology. It was expected to revolutionize and automate the analysis of tissue sections and the classification of diseases, disease subtypes, and disease stages. Surprisingly, the latter areas (especially the analysis and classification of different types, subtypes, and stages of tumors) are the ones in which MAT has proven to be immediately useful, delivering very promising and convincing results (15–17).
However, in basic research, initial MA studies often caused disappointment, failing to generate or confirm new hypotheses. On many occasions, MAT has generated more confusion than comprehension, and more questions than answers (14, 18, 19). Despite this fact, MAT has improved greatly in recent years. Most of the early problems (e.g., reproducibility, sensitivity, high background, standardization, preparation of samples, data analysis) have been addressed, and greatly improved; however, MAT still has problems breaking out of a relatively narrow field of applicability. One of the main sources of the remaining problems is that while MA analysis requires 0.5–5.0 million cells per sample, biological systems (tissues, organs) with this amount of cells are almost always mixtures of several different cell types (20)—all of which may behave differently in a given experiment. The two main exceptions are cell lines and tumors, in which one can have more than enough cells of the same type for MA analysis. Not surprisingly, these two are the main fields of successful MA studies (21–23). However, tumors only represent a narrow segment of pathological processes in humans, and immortalized cell lines have been repeatedly shown to differ significantly from in vivo cells, even the ones from which they originated (24, 25). In many other MA studies of unsorted cells, success could be achieved because most of the cells in the sample behaved similarly, thus their reaction to experimental condition changes could be detected (26–28). One would suspect, though, that even in these cases, many subtle effects might have remained undetected, overshadowed by the nonreacting cells. Even major changes in gene expression levels of a minor cell subset of the mixture might have been lost in the background of the more numerous unchanged cells.
The obvious solution to the problem of cell mixtures showing mixed gene expression profiles (GEPs) is cell sorting. Delivering sorted, more homogenous cell samples to MAs is expected to produce much clearer results than studying cell mixtures. Several recent studies have taken this approach and many of them have shown promising results (29–31). However, it is still not well understood how different cell sorting methods and sample handling protocols affect the GEP (18, 32, 33). It is also not well known how much effect different cell types have on each other's GEPs when they are mixed together. On the other hand, in the real world of both basic and clinical applications, 100% purity of a given cell type is not always achievable or feasible. So the question is, how pure is pure enough?
To address these very basic questions, we studied the overall GEP of defined cell mixtures to model heterogeneous biological samples. For the “overall GEP,” we used the unprocessed microarray readout of the cell sample with no values excluded. We evaluated the effects of cell labeling, fixation, and sorting on the overall GEP. We also analyzed how well we could recover the GEP of a pure cell type by sorting these cells from a mixture. Since different types of microarrays do not necessarily produce the same data (7, 34, 35), we used both spotted (Clontech, Palo Alto, CA) and short-oligonucleotide arrays (Affymetrix, Santa Clara, CA) to compare the results of the same experiments on these different platforms.
Microarray data analysis is not an obvious exercise; in fact it has developed into its own new field, with quite a few competing methods. The complexity of these methods often creates a communicational gap between the data-producing biologists and the data-analyzing mathematicians and biostatisticians (7, 8, 36). To avoid this gap, and to keep the presentation of our results as directly connected to the samples they represent as possible, we have deliberately used very simple, straightforward methods to compare the overall GEP of different samples, rather than any of the more sophisticated software packages existing today.
This article demonstrates why “cytomics” approaches, as discussed by Valet et al. (37) are important. Without the ability to analyze and purify cell subpopulations using cytomics technologies, much of the power of MAT is compromised or even lost.
- Top of page
- MATERIALS AND METHODS
- LITERATURE CITED
In this project, we set out to examine both the capabilities and limitations of the microarray approach in studying gene expression in real biological samples and defined, model cell mixtures. To determine if meaningful data could be obtained by this method, even in the unfortunate scenario when the investigated cell type was a small minority in the sample, we tested the effects of cell-subset ratios and sample processing methods on the overall GEP, as well as on individual gene expression levels. We modeled biological samples by cell mixtures of two cell types, with different ratios of each type, and analyzed their GEP by both spotted (Clontech) and short-oligonucleotide microarrays (Affymetrix).
We found that without applying any cell separation, the cell type in the majority dominated the overall GEP, while the GEPs of minor cell subsets were washed out. Looking at the overall GEP when investigating a minor cell subset of a cell mixture is like “only seeing the tip of the iceberg.” The differences between the GEPs of the gradually changing cell mixtures convincingly mirrored the changes in cell ratios. Summarizing our model cell mixture experiments, we concluded that the gene expression profiles for mixed cell populations are, as expected, the combined expression profiles for each cell subpopulation, weighted according to its relative frequency in the cell mixture.
When trying to determine the sensitivity of the MA approach in analyzing cell mixtures, we found that, in our model, the overall GEP of a more than 75% pure sample (PS2 and PS3) was indistinguishable from a 100% pure sample (PS1). The number of outlier genes within these samples as seen in the Trellis plots was very small (fewer than 10 out of more than 12,000 genes), considering that the raw data was not preprocessed prior to scatterplot analysis. However, these outlier genes might indicate that purity requirements can be very different for monitoring individual genes, depending on whether those same genes are expressed at high or low levels in the contaminating cell types.
Our results indicated that, in the case of minor cell subsets, to be able to see more than just the tip of the iceberg, cell purification is necessary. Any purification method takes time, and mRNA is a “moving target,” with possible degradation during the experiment. RNAs can be produced very rapidly and some mRNAs may be degraded in a matter of minutes in live cells, while others may be stable over several hours (32, 42). Is it possible to “freeze the GEP in time” by cell fixation until the cells get purified and delivered to the microarray? Most purification methods require labeling of the target cells. How much will the labeling process alter the GEP of the labeled cells? To address these questions, along with the reasonable concern that the more a sample is processed the more distorted its GEP might get, we also tested handling effects on overall sample GEP. We showed that after antibody labeling and methanol fixation the overall GEP remained unaltered, and even omitting steps traditionally used to improve RNA quality did not have a significant effect on the overall GEP. Again, the presence of a few outlier genes indicated that individual genes might be much more affected by certain processing steps; obviously, antibody labeling of a surface receptor on a live cell might trigger certain pathways altering the expression levels of the genes involved. Nevertheless, we were able to conclude that the overall GEP of a sample (representing the vast majority of all genes) is more robust and resistant to sample processing than has been generally appreciated.
Methanol fixation of the antibody-labeled cells prior to cell purification and MA analysis turned out to be a rather fortunate choice. It did not unfavorably alter the detection/selection process by either immunomagnetic cell separation or flow cytometric cell sorting. Both PE- and FITC-labeled antibodies used in these experiments maintained good separation characteristics after methanol postfixation. In summary, this fixation method not only preserved the GEP of labeled cells, but also allowed fluorescence-based labeling for cell sorting.
To address the question of how much purity we need in a sample, we showed that generally it is not necessary to achieve 100% purity. In our model for the overall GEP, anything above 75% pure was found to be indistinguishable from the pure sample. This level of purity can be achieved by two rounds of magnetic bead sorting. One round typically results in about 70% purity; two rounds raises the purity to approximately 90%; while after three rounds, it is generally above 95%. One round of magnetic bead sorting followed by one round of flow cytometry/cell sorting results in about 90–98% purity as well. Studying individual genes, however, might require much higher or lower sample purity, depending on the gene's relative expression levels in the cell subsets. With good cell biomarkers and techniques, multiparameter flow cytometry/cell sorting can be used to obtain purities of more than 99%. This degree of purity may be needed for correct GEP analysis of low-expressing genes in which even 90% purity may be insufficient to obtain accurate GEP results for those specific genes.
To test just how much of the “iceberg” can be revealed, we purified the minority cell subset of a 10% cell mix in which the GEP of the minor cell subset was shown to be covered by the background cells. Using both magnetic bead cell purification and flow cytometry/cell sorting, we managed to recover the “hidden” GEP virtually perfectly, also proving that the sort process itself did not distort the profile. The almost perfectly recovered profiles suggest that both magnetic bead and conventional flow cytometry/cell sorting purification methods are capable of purifying cells without significantly distorting their GEP. Results from the control KG-1a cells from the cord blood experiment confirmed this finding, since the cells that went through the sort process matched the unsorted cells, with an R2-value of 0.99. We concluded that for meaningful gene expression microarray profiling a minor cell subset of a cell mixture, purification of these cells is not only necessary, but also very much achievable, recovering the “pure profile” without any significant distortion, despite the concerns expressed previously in the literature (18, 33, 39). In our hands, following the procedures described in this work, the effects of sample handling on the GEP were minimal and not significant.
As a proof-of-principle experiment, we measured the GEP of purified, CD34+ cord blood stem/progenitor cells. Since these cells are present in cord blood at a less than 1% minority of all mononuclear cells (40, 41), their GEP had been heavily masked by the overwhelming presence of mature, contaminating, cell types. The true stem/progenitor cell-GEP was “invisible” without purification. We showed that the recovered GEP of these cells was characteristically different from both CBMCs and KG-1a cells. This result seriously questions the use of KG-1a cells as a model cell line for stem/progenitor cells in gene expression studies, even though that cell line was originally established from a bone marrow tumor.
For the cord blood experiment, we needed to pool several samples. Individual cord blood samples could not be directly analyzed, simply because they did not provide enough purified CD34+ cells necessary for one microarray analysis. This problem would be even more serious if we wanted to further purify this cell subset based on the cells' other surface antigen properties. Unfortunately, many biological samples do not provide enough purified cells of a certain cell type for direct gene expression profiling. For these samples, nondistorting RNA-amplification is necessary prior to microarray analysis.
In summary, we found the results presented here very promising. Both Clontech and Affymetrix arrays performed at a very high level of reproducibility, the generated profiles proved to be surprisingly robust, and hidden GEPs could be accurately recovered from cell mixtures by cell separation techniques. MAT, based on specific cell subpopulations, could truly become a driving technology not only in genomics, but also in the emerging field of cytomics, which aims at the understanding of the molecular architecture and functionality of cell systems (cytomes) by single-cell analysis in combination with exhaustive bioinformatic knowledge extraction (37).