All three microarray generations, the TAG3, TAG4 and S. cerevisiae
whole genome tiling arrays, identified ALG7
as the primary target of tunicamycin, as expected (Figure ). The tiling array also identified several other genes as additional potential targets. This list of targets includes ADO1
, all of which have been shown to be sensitive to tunicamycin when knocked out, as well as BCK1
, a gene which has previously been shown to be resistant to tunicamycin when overexpressed [19
]. In particular, ADO1
is a prime example of a gene deletion strain exhibiting increased sensitivity on the tiling array, since it is detected at a log2
ratio of 2.59 in the tiling array data, but at 0.50 and 0.66 in the TAG3 and TAG4 data, respectively. In addition to known sensitive strains, our screen identified COP1
, which are involved in ER to Golgi vesicle-mediated transport (see Table for summary of sensitive strains) [25
]. As with most sensitive strains, these genes were detected at slightly higher levels on the tiling array than on the other array generations. The tiling array appears to have slightly higher variance in its log2
ratios than the other arrays (standard deviation of 0.58 in tiling, compared to 0.37 and 0.43 in TAG4 and TAG3 arrays, respectively). We determined this to be due to its increased sensitivity to hybridized barcode abundance since sometimes strains that appear sensitive on the tiling array, fall into the background signal of the other arrays, as with ADO1
. It is reassuring to observe both the primary target of tunicamycin and genes annotated as sensitive to tunicamycin in our results. Additionally, we also identified genes associated with the endoplasmic reticulum and involved in the unfolded protein response because tunicamycin promotes protein misfolding.
Figure 1 Identifying tunicamycin targets on three microarray generations. Barcode intensity data are normalized according to a DMSO reference treatment. Blue dots represent non-essential genes, red dots represent essential genes and grey dots are genes that are (more ...)
Gene targets of tunicamycin identified in the tiling array experiment.
Because the tiling array has millions of probes, only a few thousand of which are barcode probes, we hypothesized that non-specific hybridization of barcode DNA to the genome tiling probes could potentially contribute to noise in target identification. This may have been problematic because the tiling probes were not designed for explicit use with the barcode probes, which could lead to unanticipated cross-hybridization of barcode samples to tiling probe features. To determine if non-specific binding was a factor in our experiments, we co-hybridized barcode DNA with unlabeled digested genomic DNA (gDNA). The digested gDNA (20-150 bp) competitively hybridized to tiling probes of the array to which barcodes may have had a non-specific affinity. We asked if the addition of gDNA could result in an increase of specific binding of barcodes to barcode probes, yielding a HIPHOP profile with greater dynamic range and more distinct targets (making the millions of tiling probes unavailable for barcode hybridization) analogous to the addition of salmon or herring sperm to a Southern blot to prevent non-specific hybridization [27
]. However, in practice, we found that the addition of gDNA did not improve resolution of the target ALG7
when compared to a microarray without competitive gDNA co-hybridization (Additional File 2
Our initial experiments used protocols for each microarray that were optimized for that particular technology. For example, each array type has particular hybridization, washing and staining protocols. To minimize the effect of these subtle variations and to accurately compare intensity data across array generations, we hybridized a reference sample (treated with 2% DMSO) to TAG3, TAG4 and tiling microarrays and applied TAG4 wash protocols to each array type. The hybridization conditions were fixed so that we could be certain that any changes we observed were attributed solely to feature size and not protocol variation. We scanned the microarrays following this protocol, and subsequently applied the tiling array antibody stain wash step to all three chips and, once again, scanned them. In this manner, each array was treated identically. In general, we observed median downtag intensity was higher than median uptag intensity (Figure ), an observation that was also reported by Pierce et al
]. In addition, the median intensities differed across generations, with TAG3 intensity lower than TAG4 intensity, which was lower than tiling intensity.
Figure 2 A) Median intensity for all barcodes, including downtags, uptags and unassigned barcodes (used to measure background). Median is used due to non-normal intensity distributions (see B). Tiling intensities are consistently higher than TAG4, which are higher (more ...)
We found that TAG4 and tiling array intensities were very highly correlated (Tables and ; example in Figure ). This correlation increased slightly once the arrays had been antibody stained during the tiling wash protocol. In contrast, TAG3 intensities did not correlate as well with either TAG4 or tiling, and this decreased significantly after antibody staining. However, this low correlation is unlikely to affect identification of drug targets on TAG3 arrays, as these strains are often the most distinguishable from the background, as shown previously (Figure ).
Pearson correlation coefficients (r) across microarray generations without antibody (Ab) stain.
Pearson correlation coefficients (r) across microarray generations with antibody (Ab) stain.
Figure 3 TAG4 and tiling array data correlation after antibody staining. This example shows that the signal intensity for common barcodes between TAG4 and tiling arrays are highly correlated (r = 0.952), demonstrating that tiling arrays are as accurate as TAG4 (more ...)
The relatively recent design of the TAG4 microarray includes five replicates of each barcode probe [11
]. However, we noticed that intensity values do not vary greatly between these replicates, and, therefore, a minimum of three replicates should be included to allow for appropriate trim mean calculations and masking of unusable barcode probes [14
]. This finding confirms an earlier assertion by Pierce et al
. that suggests that the minimum number of replicates required to achieve high correlation is three replicates, and that the increase in correlation from the fourth and fifth replicates is marginal [11
]. Although the TAG3 and tiling results contain only single data points for each barcode and are able to determine ALG7
as the primary target of tunicamycin (Figure ), replicate data points are advised to accommodate hybridization, washing and staining inconsistencies.