A comprehensive platform for highly multiplexed mammalian functional genetic screens

Beginning in the late 1990's, construction of a systematic, barcoded gene deletion strain collection in S. cerevisiae set the stage for high throughput functional screening in eukaryotes [1, 2]. Researchers using that resource developed powerful functional genomics screening approaches, and significantly, introduced the concept of performing complex, pooled population screens to simultaneously interrogate an organism's entire genome to identify drug targets, synthetic genetic interactions and other phenomena [1, 3–7]. Similarly, the contemporary development of large libraries of short hairpin RNAs (shRNAs) and open reading frame (ORF) collections has significantly expanded the research toolkit available for performing mammalian functional genetics in a comprehensive manner [8–11]. The development of such tools, combined with the years of lessons from yeast-based screens suggested that similar genome-wide screens using a 'barcoded pool' screening strategy should be possible in a mammalian context [12, 13]. Indeed, several groups have demonstrated the impressive ability of pooled RNAi dropout screens to identify genes essential to cancer cell proliferation [14–18]. Based on these early reports, pooled RNAi screens hold great promise for identifying genes important for growth, metabolism, differentiation, DNA damage response, sensitivity/resistance to therapeutics and many other processes in both normal and diseased cells.

Linking phenotypic changes to gene dosage in a heterogeneous population of cells requires a stably inherited molecular tag, such as a DNA barcode, that uniquely associates the phenotype with the perturbagen. A decade of work with yeast and other microbes has shown that DNA barcodes are sensitive and quantitative genetic markers that permit cell-based screening in high complexity pooled formats, with subsequent deconvolution by amplification of barcodes followed by microarray hybridization or high throughput DNA sequencing [19–21]. In contrast to the yeast knock-out strains, for pooled shRNAs and ORFs the growth perturbing agents themselves can serve as specific barcodes because they consist of uniquely identifiable DNA sequences. This approach has been demonstrated in several recent publications using different shRNA libraries [14–17]. For each publication, a different custom array was built and quality controlled, but information on the optimization and validation was limited, and the microarrays were not made commercially available. Consequently, researchers interested in utilizing this technology would need to perform resource intensive optimization and quality assurance on a new custom microarray for each shRNA library - an approach that is cost-inefficient, and does not promote standardized datasets amenable to high level informatics analysis. In order to facilitate loss-of-function and gain-of-function pooled screens using commercially available gene modulation libraries, we designed and validated a single microarray detection platform for deconvolution of loss- and gain-of-function pooled screens in human, mouse and yeast cells.

One can now easily procure custom microarrays designed to detect novel barcoded libraries. In order to evaluate and maximize the effectiveness of such barcoded platforms, extensive testing and optimization is required. For gene expression studies, extensive development work has already been done for a number of detection platforms and, as a result, they have become straightforward and routine to use. For new custom barcoded systems, similarly rigorous testing, optimization of protocols, and development of quality control standards must be done. This manuscript describes a detailed evaluation and analysis of the performance of the Gene Modulation Array Platform (GMAP) microarray, and carefully optimized protocols for every step of the process from sample preparation to data analysis. In essence, we provide a guide to use the GMAP chip as well as highlight the design principles and quality control measures that may be useful for other customized microarray platforms. In addition, we have also developed a software tool (using open source scripts) for use in extracting and analysing the data from the GMAP.

We designed the multipurpose, cost-effective GMAP to enable researchers to inexpensively and comprehensively collect data from genome-scale pooled gene-dosage modulation screens performed in human, mouse, and yeast cells using commercially available gene modulation libraries on a standard platform. Specifically, the GMAP enables readout of clone enrichments and depletions from pooled screens using the RNAi consortium (TRC) human and mouse libraries [11, 14, 22], human ORF expression pools [23, 24], and pooled screens using gene deletion-associated barcodes or ORFs from budding yeast[1, 25].

In order to achieve consistent results, a simple-to-follow standard operating procedure (SOP) was developed for preparing samples for hybridization (see Additional file 2 for detailed protocols and recipes). As part of the SOP optimization process, a standardized amount of shRNA probe was hybridized to the GMAP array. Hybridizations were performed with quantities of 1 μg, 2 μg, and 3 μg of purified shRNA probe (see Materials and Methods) that was generated from a plasmid pool of 78,432 (78 k) human shRNAs (Additional file 1, Figure S2a). 2 μg of probe gave the best signal-to-noise ratio and was therefore chosen as the standard quantity for all 78 k pool hybridizations. For pools of different complexity, the amount of probe applied to GMAP is altered proportionally compared to the 2 μg used for 78 k pools (for example ~1.35 μg for 54 k pools). In addition to optimizing the quantity of probe, a range of hybridization and washing temperatures were tested using 2 μg of 78 k shRNA pool probe. Hybridization was tested at 40°C and 45°C, with washing tested at 30°C and 35°C. Hybridization at 40°C, with array washing at 30°C was found to provide the highest probe signal with minimal signal from features corresponding to shRNAs not included in the pool (Additional file 1, Figure S2b).

To assess GMAP chip performance with complex shRNA populations, probe generated from the 78 k human shRNA pool or a different 78 k mouse shRNA pool was hybridized and analyzed. Signal intensity histograms and cumulative distribution plots generated from the collected data indicate that the signal intensities for probes corresponding to shRNAs present in the 78 k pools were well resolved from the background signal of GMAP features corresponding to shRNAs not present in the pools (Additional file 1, Figure S3a,b). The small overlap between the signals for features matched to either of the human or mouse shRNA pools versus the features that have no matching shRNA in the corresponding pools indicates low rates of library composition errors and cross-hybridization. A surface plot of intensities for all shRNA features on the GMAP array for the human 78 k pool versus the mouse 78 k pool indicates that large pools of equal complexity can be easily resolved from each other (Spearman correlation R < 0.01) and from shRNA features not corresponding to either experimental pool (Figure 1a).

One major application of pooled shRNA screens is to perform negative selection screens, so called "dropout screens", in cancer cell lines to identify cancer-essential genes that represent potential therapeutic targets. To assess the performance of the GMAP in deconvolving genome-scale shRNA drop-out screens, we first calculated a minimum signal in the reference data for inclusion in analysis. This was accomplished by adding 1.96 standard deviations (the 95% confidence limit) to the mean background log signal measured on human 78 k pool hybridizations, yielding a threshold of log₂ = 7.89 (Additional file 1, Figure S3c). Consequently, 89.9% of the shRNA signals in the reference data exceeded this threshold and were retained for analysis. We simulated shRNA dropout screens by altering the concentration of sub-fractions of the human 78 k pool in two different shRNA dilution experiments. The first approach utilized four different 78 k human pools, each containing the same 78,432 shRNAs. Approximately 70,100 shRNAs were present in equal concentration in all four pools, while an identical sub-fraction of ~8,300 shRNAs were either undiluted (the reference "Even pool"), or diluted 4-fold, 16-fold, or 64-fold (the "4x", "16x" and "64x" pools" respectively). Distribution plots of shRNA log₂ signal intensity difference between the diluted pools and the Even pool demonstrate that the diluted subpools can be distinctly resolved from each other (Figure 1b). In a second approach, we constructed a single 78 k shRNA pool (the "2x-20x" pool) in which, relative to the bulk population, four distinct subsets of ~8400 shRNAs each were diluted 2-fold, 5-fold, 10-fold and 20-fold, respectively. Similar to the previous experiment, distribution plots of shRNA signal difference between diluted pools and the reference population in the Even pool demonstrate that the diluted subpools are resolvable (Figure 1c). We did observe that the signal reduction in diluted subpools on the GMAP chip is not directly proportional to the dilution factor, indicating that changes in microarray signal are not linearly proportional to shRNA concentration across the full observed signal range. In particular, the subpool dilution signals are spaced more closely, or compressed, than would be expected for a linear concentration versus signal relationship.

To assess the performance of GMAP relative to the lower density and previous generation TRCBC array, we amplified shRNA probes from genomic DNA extracted from human BT-474 cells infected with a ~54,000 shRNA pool of lentiviruses corresponding to all available human shRNA features on the TRCBC chip, divided the probe and hybridized samples to both TRCBC and GMAP chips. Comparison of signal intensity between the two different array formats revealed a Spearman correlation coefficient of R = 0.96 (Additional file 1, Figure S4). This result confirms that moving from 11 μm feature sizes to 5 μm feature sizes does not cause any significant reduction in signal strength or specificity, agreeing with reported observations on yeast barcode arrays [26].

A major concern in using large shRNA pools with microarray detection strategies is the extent to which cross-hybridization occurs. To address this issue, we measured the frequency of significant signals from GMAP shRNA features that did not correspond to constituents of a pool of shRNA probes. Specifically, we examined features corresponding to the 78 k mouse pool for signal when the 78 k human pool was hybridized. Features on GMAP that had 100% identity to shRNAs in both the mouse and human pools were removed from the analysis. After doing so, we found that only 2637 of 77,690 (3.39%) mouse 78 k pool features had significant signal (log₂≥7.89). From this finding, we infer that amongst the human shRNA features, the cross-hybridization rate from human 78 k pooled probe would be similar.

Frequency distribution plots of sequencing read counts for the Even, 4x, 16x, 64x and 2-20x pools (Figure 2b,c) show similar characteristics to the distributions generated for the same pools by microarray detection on the GMAP (Figure 1b,c) with some exceptions. First, the distributions tended to be slightly wider for sequence data. Second, the distributions for sequence data exhibit a linear relationship that more accurately reflects the actual experimental dilution of shRNAs in the dilution pools. In other words, the distribution curves for sequencing data tend to center over the correct fold-dilution. These observations demonstrate that sequencing barcode pools results in linear, quantifiable signals whereas microarray detection displays nonlinear signal, a behaviour previously observed by Pierce et al.[27]. In addition, the dynamic range of signal obtained from GMAP is compressed relative to deep sequencing (Additional file 1, Figure S5). However, these differences aside, overall, GMAP detection of shRNA sequences was similar to sequencing-based deconvolution in its sensitivity and quantitative reproducibility over a range of shRNA concentrations.

Microarray experiments have historically suffered from subtle to substantial differences in data produced from the same or similar templates when performed on different dates, by different users, or in different locations [28]. A candidate method to enable detection and/or correction of these differences is to include a standard set of synthetic oligonucleotide probes that are "spiked" into array hybridization mixtures in defined amounts. These hairpin spike-in (HPSI) probes, designed to have identical length and similar sequence characteristics to shRNAs, may be used as yardsticks for artifact detection and data normalization methods. We replicated 25 clusters of 200 HSPI features across the GMAP (Table 1). Spatially localized signal intensity fluctuations between HSPI clusters may indicate potentially poor hybridization or washing performance, contamination, or physical damage to the array surface. As a trial, 12 HSPI probes were tested in replicate array hybridizations to examine their hybridization characteristics and it was found that they behaved in a dose-dependent manner similar to labelled shRNA probes (Additional file 1, Figure S6).

To examine GMAP performance in a simulated cell screening experiment, two large pools of 90,408 shRNAs, each targeting ~18,000 human genes were constructed. For the "90 k Reference" pool, all of the shRNA plasmids were combined at approximately equal concentrations, while a dilution series pool, the "90 k Dilution", contained four sub-sets of ~3,500 shRNAs each that were diluted 2-fold, 4-fold, 10-fold or 20-fold with respect to their counterparts in the 90 k Reference pool. Distribution plots of shRNA log₂ signal difference between the 90 k Dilution and the 90 k Reference pools demonstrate that the diluted sub fractions are clearly resolvable (Figure 2d). We generated lentivirus from the 90 k Reference and 90 k Dilution pools that was subsequently used to infect A549 cells. Genomic DNA prepared from these cells containing integrated shRNA-expressing cassettes was used as template for probe generation and GMAP hybridization. The resulting distribution plots of log₂ signal difference between the 90 k Dilution and 90 k Reference pools post-infection (Figure 2e) were very similar to those achieved with probe generated from plasmid template for the 90 k pools and 78 k pools. A scatter plot comparing data for the 90 k Reference pool derived from plasmid and genomic DNA template reveals excellent correlation (R > 0.97, see Additional file 1, Figure S7) further demonstrating that reliable results can be obtained from pooled cell-based screening experiments with shRNA pools spanning essentially the entire human genome with coverage of 4-5 shRNAs per gene.

To enable pooled ORF over-expression screening using the GMAP array to detect and quantify ORF sequences, we designed features against 22,449 distinct human ORF sequences in the Mammalian Genome Collection (MGC) (Table 1 and [29, 30]). Between 1 and 8 probes were designed against each ORF, with a median of 7 probes per sequence. For comparative purposes, we also included up to three additional features found on the human expression profiling Affymetrix Gene 1.0 ST array for 18,088 of these sequences. To assess the GMAP performance with human ORF hybridization, we developed 41 plasmid pools of entry clones (15,347 open reading frames) from the human ORFeome v5.1 collection that were combined to generate a single master pool. Subsequently, 15,347 ORFs were amplified in pooled format with common flanking primers, labelled and hybridized to both the Human Gene 1.0 ST array and GMAP arrays in duplicate (see Materials and Methods). Signal for features shared between the two arrays was highly correlated (Pearson's correlation coefficient R = 0.953, Figure 3a), with similar distribution of signal across the features for each array (Figure 3b) and similar signal-to-noise ratios (data not shown). These results demonstrate that the GMAP has robust reporting of ORF data, and suggests that the GMAP can be used for a number of human gene assays including human ORF/cDNA overexpression screens.

To compare the dynamic range of signal for huORF and huGene features on the GMAP, five different quantities of probe generated from ORFeome plasmid pools were hybridized to GMAP in duplicate. The resulting data indicated that two-fold changes in probe input produced highly correlated signals across a 16-fold dynamic range (Figure 3c), thus we combined huORF and huGene features into one feature set. Our methodology to detect ORF sequences on microarrays depends on common flanking primers that amplify entire open reading frames. This provides an opportunity to examine the utility of the three sets of ORF features on the GMAP microarray, including the 22-mer hairpin features designed to detect shRNA barcodes. The signal concordance between sequence identity-conserved hORF features and shRNA features on GMAP was interchangeable across the entire spectrum of probe concentrations (Figure 3d). This concordance can be further exploited to expand the feature set for each gene to obtain more accurate measurements of ORFs in a given hybridization mix, which is particularly important for discriminating isoforms in a polyclonal library of open reading frames.

GMAP also contains triplicate copies of features for the collection of ~16,000 20 nucleotide Yeast Knockout Collection barcodes[2], and single copies of ~12,000 25 nucleotide yeast ORF-specific features (Table 1). The yeast ORFs represented on this array are identical to those found on the Tag4 barcode array with two distinct probes designed against each of 5718 yeast ORF sequences. These features can serve as additional controls as they are non-homologous to the shRNA or human ORFeome probes on the GMAP. The yeast features on the GMAP display comparable performance to the TAG4 microarray (Affymetrix) from which they are derived (data not shown).

A major hurdle to using a new microarray platform is the informatics associated with data extraction and analysis. The GMAP chip contains several subsets of features, some of which can be organized as traditional Affymetrix feature sets as well as features for interrogation of short input DNA sequences (shRNA barcodes). Extraction of signal from these arrays can be done with the Affymetrix Power Tool (APT) collection. To help aid adoption of the GMAP, we have developed a Java application to allow data reorganization, graphical summaries and downstream analysis of the datasets extracted from these arrays (files and applications available online at http://chemogenomics.med.utoronto.ca/supplemental/gmap/). This application makes use of a variety of R and Bioconductor libraries (see Materials and Methods). The R functions that we have developed for this analysis will also be available as scripts/libraries to allow end users to work directly in the R environment. The current functionality of the application includes methods to extract signals into a single tab-delimited file for subsets of features and GC-matched background signal, normalization procedures, routines to generate signal ratios against one or more reference chips with options to fine-tune analyses and standard or user-defined annotation files for merging analyzed signal.

Our goal was to develop a comprehensive and standardized microarray platform for quantification of pooled screening results using commercially available gene expression modulation libraries for human, mouse and budding yeast cells. The advantages of a standardized microarray are many-fold. First, a highly validated microarray is invaluable to research labs that want to adopt compatible screening approaches but do not have extensive resources to build, test and optimize a custom barcode microarray themselves. The GMAP presents an accessible solution because the obstacle of designing, constructing, and testing of the array as well as development of rigorous SOPs has already been overcome. Second, screening costs are reduced since the GMAP supports multiple commercially available functional screening approaches in a single solution. As such, individual labs or facilities do not have to re-optimize hybridization conditions for different microarray platforms, and a single framework can be used for data extraction and quality control analysis. Third, a standard microarray permits use of uniform methods across the many laboratories using these gene modulation libraries. Lastly, a multipurpose array can serve as a useful validation tool as new technologies for detecting barcodes utilizing deep sequencing mature and become more widely used. To enable facile adoption of the GMAP array for large-scale screen deconvolution, we have also provided a series of supplemental protocols and software tools to help extract data from functional screens using commercially available gene modulation libraries.

A thorough comparison of barcode detection strategies for large, complex shRNA pools has been lacking in the literature. We undertook such a comparison with pools comprised of ~78,000 and ~90,000 distinct lentiviral shRNAs, the largest such shRNA screening pools described to date. We directly compared microarray detection with deconvolution by Illumina deep sequencing and conclude that, at least in the current context, data quality is comparable using either method. For many or most users, the array-based read-out is currently more cost-effective on a per-sample basis when costs per sample are tabulated. Equally important as cost for many researchers is the turn-around time required for sample preparation, sample run time and data acquisition/processing, which is currently accomplished more rapidly using microarrays. Sample run time on the GMAP chip spans ~18 hours between probe application, hybridization, washing and scanning. Typical microarray platform equipped labs or core facilities should be able to easily accommodate more than 30 experimental samples in a typical work-week.

In addition to examining a plasmid pool containing >78,000 shRNA clones, we also simulated a cell-based dropout screen with a larger pool of >90,000 shRNA clones. Genomic DNA was isolated from each of the transduced populations and shRNA barcodes were detected using the GMAP. Importantly, the resolution of barcode detection was nearly equivalent whether the starting template was from plasmid DNA or genomic DNA. This experiment demonstrates that with the protocols developed here, the GMAP offers an effective and efficient means of screening very large pools in a cellular context. Our trials with complex shRNA and human ORF pools demonstrate that the GMAP array is capable of efficiently quantifying human and mouse genome-scale gene modulation screens. As well, the GMAP chip detection performance is equally suited for quantifying shRNA abundance in lower complexity pools. The GMAP array is not limited to use with the pooled reagents described here - in fact, any screen which uses the yeast deletion barcodes or TRC shRNA sequences as molecular barcodes can be analyzed using the GMAP array. Further generations of GMAP, utilizing even higher feature-count microarrays, could be expanded to include other shRNA and/or ORF libraries. However, benefits of such an expansion would have to be weighed against the increased cost of such expanded arrays. Generating specific arrays corresponding to different shRNA libraries might be more efficient and cost effective. In summary, this cost-effective platform provides any academic laboratory with access to a standardized array and a suite of methods and analytical tools to perform systematic genome-scale genetics on mammalian cells.